Intial Commit

.gitignore

@@ -0,0 +1,27 @@
+__pycache__/
+*.py[cod]
+/.vscode/
+/.idea/
+*.sw[pon]
+
+/results/
+*.jpg
+*.png
+*.jpeg
+*.gif
+*.avi
+*.mp4
+
+*.npy
+*.json
+*.log
+*.html
+*.tar
+*.zip
+events.*
+
+*.pth
+*.pt
+*.pkl
+*.h5
+*.dat

README.md

@@ -5,6 +5,7 @@ colorFrom: blue
 colorTo: indigo
 sdk: gradio
 sdk_version: 3.24.1
+python_version: 3.7
 app_file: app.py
 pinned: false
 ---

app.py

@@ -0,0 +1,72 @@
+import gradio as gr
+from inference import initialize_styleres
+from utils import AppUtils
+from datasets.process_image import ImageProcessor
+from argparse import ArgumentParser
+
+EXAMPLES = [
+    ["test_sample/1.jpg", "InterfaceGAN",  "Smile", "2.0",  False],
+]
+
+parser = ArgumentParser()
+parser.add_argument('--device', type=str, default='cpu', help='Which device to use')
+args = parser.parse_args()
+
+utils = AppUtils()
+methods = utils.get_methods()
+styleres = initialize_styleres('checkpoints/styleres_ffhq.pth', args.device)
+image_processor = ImageProcessor('checkpoints/shape_predictor_68_face_landmarks.dat')
+
+def process_image(image, method,  edit, factor, is_align_checked):
+    cfg = utils.args_to_cfg(method, edit, factor)
+    if is_align_checked:
+        image = image_processor.align_face(image)
+    image = image_processor.preprocess_image(image, is_batch=False)
+    image = styleres.edit_images(image, cfg)
+    image = image_processor.postprocess_image(image.detach().cpu().numpy(), is_batch=False)
+    return image
+
+def update_edit_dropdown(method):
+    choices = utils.get_edits(method)
+    return gr.Dropdown.update(choices=choices, value=choices[0])
+    
+def update_slider(method):
+    minimum, maximum, step= utils.get_range(method)
+    return gr.Slider.update(minimum=minimum, maximum=maximum, value=0, step=step, label=f"Strength [{minimum}, {maximum}]")
+
+
+with gr.Blocks() as demo:
+    gr.Markdown(
+    """
+    # StyleRes: Transforming the Residuals for Real Image Editing with StyleGAN (CVPR2023)
+    """)
+    with gr.Row():
+        image_input = gr.Image(type="pil", shape=(256,256), label='Input Image', value="test_sample/116.jpg")
+        image_output = gr.Image(type="pil", shape=(256,256), label='Output Image')
+
+    with gr.Row():
+        with gr.Column(scale=0.25, min_width=50):
+            methods_drowdown = gr.Dropdown(methods, label="Choose Method", value=methods[0])
+        with gr.Column(scale=0.25, min_width=50):
+            edits_dropdown = gr.Dropdown(utils.get_edits(methods[0]), label="Choose Edit", value=utils.get_edits(methods[0])[0])
+ 
+
+    with gr.Row(): 
+        with gr.Column(scale=0.1, min_width=50):
+            is_align_checked = gr.Checkbox(label="Crop + Align")
+        with gr.Column(scale=0.4, min_width=50):
+            factor_slider = gr.Slider(-5, 5, value=0, label="Strength [-5, 5]")
+
+    gr.Examples(
+        examples=EXAMPLES,
+        inputs=[image_input, methods_drowdown,  edits_dropdown, factor_slider, is_align_checked],
+        outputs=image_output,
+        fn=process_image,
+        cache_examples=True,
+    )
+    methods_drowdown.change(update_edit_dropdown, inputs=methods_drowdown, outputs=edits_dropdown )
+    methods_drowdown.change(update_slider, inputs=methods_drowdown, outputs=factor_slider)
+    factor_slider.release(process_image, inputs=[image_input, methods_drowdown,  edits_dropdown, factor_slider, is_align_checked], outputs=image_output)
+
+demo.launch(debug=True)
+

datasets/demo_dataset.py

@@ -0,0 +1,11 @@
+import numpy as np
+import torch
+
+def preprocess_demo_image(image):
+    image = image.resize( (256, 256))
+    image = np.asarray(image).transpose(2, 0, 1).astype(np.float32) # C,H,W -> H,W,C
+    image = torch.FloatTensor(image.copy())
+    image = (image - 127.5) / 127.5     # Normalize
+    return image.unsqueeze(0)
+
+    

datasets/inference_dataset.py

@@ -0,0 +1,49 @@
+# python3.7
+"""Contains the class of dataset."""
+
+import os
+from PIL import Image
+from .process_image import ImageProcessor
+from torch.utils.data import Dataset
+
+class InferenceDataset(Dataset):
+    
+    def __init__(self,
+                 root_dir,
+                 resolution=256,
+                 aligner_path=None
+                 ):
+        """Initializes the dataset.
+
+        Args:
+            root_dir: Root directory containing the dataset.
+            resolution: The resolution of the returned image.
+            transform: The transform function for pre-processing.
+                (default: `datasets.transforms.normalize_image()`)
+        """
+
+        self.root_dir = root_dir
+        self.resolution = resolution
+        self.image_paths = sorted(os.listdir(self.root_dir))
+        self.num_samples = len(self.image_paths)
+        self.processor = ImageProcessor(aligner_path)
+
+    def __len__(self):
+        return self.num_samples
+
+    def __getitem__(self, idx):
+        data = dict()
+
+        image_path = self.image_paths[idx]
+        image = Image.open(os.path.join(self.root_dir, image_path))
+        image = self.processor.align_face(image)
+        image = self.processor.preprocess_image(image)
+        # image = image.resize( (self.resolution, self.resolution))
+        # image = np.asarray(image).transpose(2, 0, 1).astype(np.float32) # C,H,W -> H,W,C
+        # image = torch.FloatTensor(image.copy())
+        # image = (image - 127.5) / 127.5     # Normalize
+
+        data.update({'image': image})
+        data.update({'name': image_path})
+        return data
+

datasets/process_image.py

@@ -0,0 +1,166 @@
+
+import numpy as np
+import torch
+from PIL import Image
+import dlib
+import numpy as np
+import PIL
+import PIL.Image
+import scipy
+import scipy.ndimage
+
+class ImageProcessor():
+    def __init__(self, predictor_path=None) -> None:
+        self.predictor = None
+        if predictor_path:
+            self.predictor =  dlib.shape_predictor(predictor_path)
+    
+    @staticmethod
+    def preprocess_image(image, is_batch=True):
+        image = image.resize( (256, 256))
+        image = np.asarray(image).transpose(2, 0, 1).astype(np.float32) # C,H,W -> H,W,C
+        image = torch.FloatTensor(image.copy())
+        image = (image - 127.5) / 127.5     # Normalize
+        if not is_batch:
+            image = image.unsqueeze(0)
+        return image
+
+    """
+        Input: A numpy image with shape NxCxHxW.
+        Output: Output image with NxHxWxC with values between 0-255
+    """
+    @staticmethod
+    def postprocess_image(image, min_val=-1.0, max_val=1.0, is_batch=True):
+        image = image.astype(np.float64)
+        image = (image - min_val) * 255 / (max_val - min_val)
+        image = np.clip(image + 0.5, 0, 255).astype(np.uint8)
+        image = image.transpose(0, 2, 3, 1)
+        if not is_batch:
+           image = Image.fromarray(image[0]).resize((256,256))
+        return image
+
+    """
+    brief: face alignment with FFHQ method (https://github.com/NVlabs/ffhq-dataset)
+    author: lzhbrian (https://lzhbrian.me)
+    date: 2020.1.5
+    note: code is heavily borrowed from
+        https://github.com/NVlabs/ffhq-dataset
+        http://dlib.net/face_landmark_detection.py.html
+    requirements:
+        apt install cmake
+        conda install Pillow numpy scipy
+        pip install dlib
+        # download face landmark model from:
+        # http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2
+    """
+
+    def get_landmark(self, image):
+        """get landmark with dlib
+        :return: np.array shape=(68, 2)
+        """
+        detector = dlib.get_frontal_face_detector()
+
+        # img = dlib.load_rgb_image(filepath)
+        img = np.asarray(image)
+        dets = detector(img, 1)
+
+        for k, d in enumerate(dets):
+            shape = self.predictor(img, d)
+
+        t = list(shape.parts())
+        a = []
+        for tt in t:
+            a.append([tt.x, tt.y])
+        lm = np.array(a)
+        return lm
+
+    def align_face(self, img):
+        """
+        :param image: PIL image
+        :return: PIL Image
+        """
+        if self.predictor is None:
+            return img
+
+        lm = self.get_landmark(img)
+
+        lm_chin = lm[0: 17]  # left-right
+        lm_eyebrow_left = lm[17: 22]  # left-right
+        lm_eyebrow_right = lm[22: 27]  # left-right
+        lm_nose = lm[27: 31]  # top-down
+        lm_nostrils = lm[31: 36]  # top-down
+        lm_eye_left = lm[36: 42]  # left-clockwise
+        lm_eye_right = lm[42: 48]  # left-clockwise
+        lm_mouth_outer = lm[48: 60]  # left-clockwise
+        lm_mouth_inner = lm[60: 68]  # left-clockwise
+
+        # Calculate auxiliary vectors.
+        eye_left = np.mean(lm_eye_left, axis=0)
+        eye_right = np.mean(lm_eye_right, axis=0)
+        eye_avg = (eye_left + eye_right) * 0.5
+        eye_to_eye = eye_right - eye_left
+        mouth_left = lm_mouth_outer[0]
+        mouth_right = lm_mouth_outer[6]
+        mouth_avg = (mouth_left + mouth_right) * 0.5
+        eye_to_mouth = mouth_avg - eye_avg
+
+        # Choose oriented crop rectangle.
+        x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1]
+        x /= np.hypot(*x)
+        x *= max(np.hypot(*eye_to_eye) * 2.0, np.hypot(*eye_to_mouth) * 1.8)
+        y = np.flipud(x) * [-1, 1]
+        c = eye_avg + eye_to_mouth * 0.1
+        quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])
+        qsize = np.hypot(*x) * 2
+
+        # read image
+        # img = PIL.Image.open(filepath)
+
+        output_size = 512
+        transform_size = 1024
+        enable_padding = True
+
+        # Shrink.
+        shrink = int(np.floor(qsize / output_size * 0.5))
+        if shrink > 1:
+            rsize = (int(np.rint(float(img.size[0]) / shrink)), int(np.rint(float(img.size[1]) / shrink)))
+            img = img.resize(rsize, PIL.Image.ANTIALIAS)
+            quad /= shrink
+            qsize /= shrink
+
+        # Crop.
+        border = max(int(np.rint(qsize * 0.1)), 3)
+        crop = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
+                int(np.ceil(max(quad[:, 1]))))
+        crop = (max(crop[0] - border, 0), max(crop[1] - border, 0), min(crop[2] + border, img.size[0]),
+                min(crop[3] + border, img.size[1]))
+        if crop[2] - crop[0] < img.size[0] or crop[3] - crop[1] < img.size[1]:
+            img = img.crop(crop)
+            quad -= crop[0:2]
+
+        # Pad.
+        pad = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
+            int(np.ceil(max(quad[:, 1]))))
+        pad = (max(-pad[0] + border, 0), max(-pad[1] + border, 0), max(pad[2] - img.size[0] + border, 0),
+            max(pad[3] - img.size[1] + border, 0))
+        if enable_padding and max(pad) > border - 4:
+            pad = np.maximum(pad, int(np.rint(qsize * 0.3)))
+            img = np.pad(np.float32(img), ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'reflect')
+            h, w, _ = img.shape
+            y, x, _ = np.ogrid[:h, :w, :1]
+            mask = np.maximum(1.0 - np.minimum(np.float32(x) / pad[0], np.float32(w - 1 - x) / pad[2]),
+                            1.0 - np.minimum(np.float32(y) / pad[1], np.float32(h - 1 - y) / pad[3]))
+            blur = qsize * 0.02
+            img += (scipy.ndimage.gaussian_filter(img, [blur, blur, 0]) - img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0)
+            img += (np.median(img, axis=(0, 1)) - img) * np.clip(mask, 0.0, 1.0)
+            img = PIL.Image.fromarray(np.uint8(np.clip(np.rint(img), 0, 255)), 'RGB')
+            quad += pad[:2]
+
+        # Transform.
+        img = img.transform((transform_size, transform_size), PIL.Image.QUAD, (quad + 0.5).flatten(), PIL.Image.BILINEAR)
+        if output_size < transform_size:
+            img = img.resize((output_size, output_size), PIL.Image.ANTIALIAS)
+
+        # Save aligned image.
+        img.save('aligned.png')
+        return img

editings/base_encoder_runner.py

@@ -0,0 +1,861 @@
+# python3.7
+"""Contains the base class for Encoder (GAN Inversion) runner."""
+
+import os
+import shutil
+
+import torch
+import torch.distributed as dist
+import torchvision.transforms as T
+
+from utils.visualizer import HtmlPageVisualizer
+from utils.visualizer import get_grid_shape
+from utils.visualizer import postprocess_image
+from utils.visualizer import save_image
+from utils.visualizer import load_image
+from utils.visualizer import postprocess_tensor
+
+from metrics.inception import build_inception_model
+from metrics.fid import extract_feature
+from metrics.fid import compute_fid
+from metrics.MSSIM import MSSSIM
+from metrics.LPIPS import LPIPS
+import numpy as np
+from .base_runner import BaseRunner
+from datasets import BaseDataset
+from torch.utils.data import DataLoader
+from PIL import Image
+from runners.controllers.summary_writer import log_image
+import torchvision
+from editings.latent_editor import LatentEditor
+from editings.styleclip.edit_hfgi import styleclip_edit, load_stylegan_generator,load_direction_calculator
+from editings.GradCtrl.manipulate import main as gradctrl
+import torch.nn.functional as F
+import time
+
+
+__all__ = ['BaseEncoderRunner']
+
+
+class BaseEncoderRunner(BaseRunner):
+    """Defines the base class for Encoder runner."""
+
+    def __init__(self, config, logger):
+        super().__init__(config, logger)
+        self.inception_model = None
+
+    def build_models(self):
+        super().build_models()
+        assert 'encoder' in self.models
+        assert 'generator_smooth' in self.models
+        assert 'discriminator' in self.models
+
+        self.resolution = self.models['generator_smooth'].resolution
+        self.G_kwargs_train = self.config.modules['generator_smooth'].get(
+            'kwargs_train', dict())
+        self.G_kwargs_val = self.config.modules['generator_smooth'].get(
+            'kwargs_val', dict())
+        self.D_kwargs_train = self.config.modules['discriminator'].get(
+            'kwargs_train', dict())
+        self.D_kwargs_val = self.config.modules['discriminator'].get(
+            'kwargs_val', dict())
+        if self.config.use_disc2:
+            self.D2_kwargs_train = self.config.modules['discriminator2'].get(
+                'kwargs_train', dict())
+            self.D2_kwargs_val = self.config.modules['discriminator2'].get(
+                'kwargs_val', dict())
+        if self.config.mapping_method != 'pretrained':
+            self.M_kwargs_train = self.config.modules['mapping'].get(
+                'kwargs_train', dict())
+            self.M_kwargs_val = self.config.modules['mapping'].get(
+                'kwargs_val', dict())
+        if self.config.create_mixing_network:
+            self.MIX_kwargs_train = self.config.modules['mixer'].get(
+                'kwargs_train', dict())
+            self.MIX_kwargs_val = self.config.modules['mixer'].get(
+                'kwargs_val', dict())
+
+        
+
+    def train_step(self, data, **train_kwargs):
+        raise NotImplementedError('Should be implemented in derived class.')
+
+    def mse(self, mse_num):
+
+        if mse_num == 0:
+            return -1
+        self.set_mode('val')
+
+        if self.val_loader is None:
+            self.build_dataset('val')
+        if mse_num == "auto":
+            mse_num = len(self.val_loader.dataset)
+
+        indices = list(range(self.rank, mse_num, self.world_size))
+        self.logger.init_pbar()
+        task1 = self.logger.add_pbar_task('MSE-LPIPS-SSIM', total=mse_num)
+
+        lpips = LPIPS()
+        ssim = MSSSIM(size_average=False)
+        n_evals = 3
+        gather_list = [torch.zeros( (self.val_batch_size, n_evals), device=torch.cuda.current_device()) for i in range(self.world_size)]
+        all_errors = np.zeros( (mse_num, n_evals), dtype=np.float64)
+        shared_tensor = torch.zeros((self.val_batch_size, n_evals), device=torch.cuda.current_device())
+        gather_idx = 0
+        for batch_idx in range(0, len(indices), self.val_batch_size):
+            sub_indices = indices[batch_idx:batch_idx + self.val_batch_size]
+            batch_size = len(sub_indices)
+            data = next(self.val_loader)
+            for key in data:
+                if key != 'name':
+                    data[key] = data[key][:batch_size].cuda(
+                        torch.cuda.current_device(), non_blocking=True)
+
+            with torch.no_grad():
+                real_images = data['image']
+                return_dict = self.forward_pass(data, return_vals='fakes, wp_mixed')
+                fakes = return_dict['fakes']
+                shared_tensor[:, 0] = torch.mean((fakes - real_images)**2, dim=(1,2,3))   #MSE Error
+                shared_tensor[:, 1]= lpips(real_images, fakes)
+                shared_tensor[:, 2]= ssim(real_images, fakes)  #ssim (real_images[0].unsqueeze(0), fakes[0].unsqueeze(0) )
+
+                dist.all_gather(gather_list, shared_tensor)
+                if self.rank == 0:
+                    for t in gather_list:
+                        all_errors[gather_idx:gather_idx+batch_size, 0] = t[:,0].cpu().numpy()
+                        all_errors[gather_idx:gather_idx+batch_size, 1] = t[:,1].cpu().numpy()
+                        all_errors[gather_idx:gather_idx+batch_size, 2] = t[:,2].cpu().numpy()
+                        gather_idx = gather_idx+batch_size
+                self.logger.update_pbar(task1, batch_size * self.world_size)
+        self.logger.close_pbar()
+        mean_lst, std_lst = np.mean(all_errors, axis=0), np.std(all_errors, axis=0)
+        mse_mean, lpips_mean, ssim_mean = mean_lst[0].item(), mean_lst[1].item(), mean_lst[2].item()
+        mse_std, lpips_std, ssim_std = std_lst[0].item(), std_lst[1].item(), std_lst[2].item()
+        return_vals = {'mse': (mse_mean,mse_std), 'lpips': (lpips_mean, lpips_std), 'ssim':(ssim_mean, ssim_std)}
+        return return_vals
+
+    def fid_attribute(self,
+            fid_num,
+            z=None,
+            ignore_cache=False,
+            align_tf=True,
+            attribute='smile', factor=1, direction=None):
+        """Computes the FID metric."""
+        self.set_mode('val')
+        direction = torch.load(f'editings/interfacegan_directions/{attribute}.pt').cuda()
+        if factor < 0:
+            self.config.data['smile']['root_dir'] = f'/media/hdd2/adundar/hamza/genforce/data/temp/smile_with_original'
+        elif factor > 0:
+            self.config.data['smile']['root_dir'] = f"/media/hdd2/adundar/hamza/genforce/data/temp/smile_without_original"
+        
+        fake_loader = self.build_dataset(f"smile")
+        
+        #fid_num = min(fid_num, len(self.val_loader.dataset))
+        fid_num = len(fake_loader.dataset)
+        if self.inception_model is None:
+            if align_tf:
+                self.logger.info(f'Building inception model '
+                                 f'(aligned with TensorFlow) ...')
+            else:
+                self.logger.info(f'Building inception model '
+                                 f'(using torchvision) ...')
+            self.inception_model = build_inception_model(align_tf).cuda()
+            self.logger.info(f'Finish building inception model.')
+
+        if z is not None:
+            assert isinstance(z, np.ndarray)
+            assert z.ndim == 2 and z.shape[1] == self.z_space_dim
+            fid_num = min(fid_num, z.shape[0])
+            z = torch.from_numpy(z).type(torch.FloatTensor)
+        if not fid_num:
+            return -1
+
+        indices = list(range(self.rank, fid_num, self.world_size))
+
+        self.logger.init_pbar()
+        # Extract features from fake images.
+        fake_feature_list = []
+        task1 = self.logger.add_pbar_task(f'FID-{attribute}_fake', total=fid_num)
+        for batch_idx in range(0, len(indices), self.val_batch_size):
+            sub_indices = indices[batch_idx:batch_idx + self.val_batch_size]
+            batch_size = len(sub_indices)
+            data = next(fake_loader)
+            for key in data:
+                if key != 'name':
+                    data[key] = data[key][:batch_size].cuda(
+                        torch.cuda.current_device(), non_blocking=True)
+            with torch.no_grad():
+                real_images = data['image']
+                #valids = data['valid']
+                return_dict = self.forward_pass(data, return_vals='all', only_enc = True)
+                wp = return_dict['wp_mixed']
+                eouts = return_dict['eouts']
+                edit_wp = wp + factor * direction
+                edited_images, _ = self.runG(edit_wp, "synthesis", highres_outs=eouts)
+                fake_feature_list.append(
+                    extract_feature(self.inception_model, edited_images))
+            self.logger.update_pbar(task1, batch_size * self.world_size)
+
+        np.save(f'{self.work_dir}/fake_fid_features_{self.rank}.npy',
+                np.concatenate(fake_feature_list, axis=0))
+        self.logger.close_pbar()
+
+     
+        #Extract features from real images if needed.
+        cached_fid_file = f'{self.work_dir}/real_{attribute}_{factor}_fid.npy'
+        do_real_test = (not os.path.exists(cached_fid_file) or ignore_cache)
+        if do_real_test:
+            real_feature_list = []
+            self.logger.init_pbar()
+
+            if factor < 0:
+                self.config.data['smile']['root_dir'] = f"/media/hdd2/adundar/hamza/genforce/data/temp/smile_without_original"
+            elif factor > 0:
+                self.config.data['smile']['root_dir'] = f"/media/hdd2/adundar/hamza/genforce/data/temp/smile_with_original"
+            real_loader = self.build_dataset(f"smile")
+
+            fid_num = len(real_loader.dataset)
+            indices = list(range(self.rank, fid_num, self.world_size))
+            task2 = self.logger.add_pbar_task(f"{attribute}_real", total=fid_num)
+            for batch_idx in range(0, len(indices), self.val_batch_size):
+                sub_indices = indices[batch_idx:batch_idx + self.val_batch_size]
+                batch_size = len(sub_indices)
+                data = next(real_loader)
+                for key in data:
+                    if key != 'name':
+                        data[key] = data[key][:batch_size].cuda(
+                            torch.cuda.current_device(), non_blocking=True)
+                with torch.no_grad():
+                    real_images = data['image']
+                    real_feature_list.append(
+                        extract_feature(self.inception_model, real_images))
+                self.logger.update_pbar(task2, batch_size * self.world_size)
+            np.save(f'{self.work_dir}/real_fid_features_{self.rank}.npy',
+                    np.concatenate(real_feature_list, axis=0))
+
+        dist.barrier()
+        if self.rank != 0:
+            return -1
+        self.logger.close_pbar()
+
+        # Collect fake features.
+        fake_feature_list.clear()
+        for rank in range(self.world_size):
+            fake_feature_list.append(
+                np.load(f'{self.work_dir}/fake_fid_features_{rank}.npy'))
+            os.remove(f'{self.work_dir}/fake_fid_features_{rank}.npy')
+        fake_features = np.concatenate(fake_feature_list, axis=0)
+        # assert fake_features.ndim == 2 and fake_features.shape[0] == fid_num
+        feature_dim = fake_features.shape[1]
+        feature_num = fake_features.shape[0]
+        pad = feature_num % self.world_size #feature_dim.shape[0]
+        if pad:
+            pad = self.world_size - pad
+        fake_features = np.pad(fake_features, ((0, pad), (0, 0)))
+        fake_features = fake_features.reshape(self.world_size, -1, feature_dim)
+        fake_features = fake_features.transpose(1, 0, 2)
+        fake_features = fake_features.reshape(-1, feature_dim)[:feature_num]
+
+        # Collect (or load) real features.
+        if do_real_test:
+            real_feature_list.clear()
+            for rank in range(self.world_size):
+                real_feature_list.append(
+                    np.load(f'{self.work_dir}/real_fid_features_{rank}.npy'))
+                os.remove(f'{self.work_dir}/real_fid_features_{rank}.npy')
+            real_features = np.concatenate(real_feature_list, axis=0)
+            # assert real_features.shape == (fid_num, feature_dim)
+            feature_dim = real_features.shape[1]
+            feature_num = real_features.shape[0]
+            pad = feature_num % self.world_size
+            if pad:
+                pad = self.world_size - pad
+            real_features = np.pad(real_features, ((0, pad), (0, 0)))
+            real_features = real_features.reshape(
+                self.world_size, -1, feature_dim)
+            real_features = real_features.transpose(1, 0, 2)
+            real_features = real_features.reshape(-1, feature_dim)[:feature_num]
+            np.save(cached_fid_file, real_features)
+        else:
+            real_features = np.load(cached_fid_file)
+            # assert real_features.shape == (fid_num, feature_dim)
+
+        fid_value = compute_fid(fake_features, real_features)
+        return fid_value
+
+    def fid(self,
+            fid_num,
+            z=None,
+            ignore_cache=False,
+            align_tf=True):
+        """Computes the FID metric."""
+        self.set_mode('val')
+
+        if self.val_loader is None:
+            self.build_dataset('val')
+        fid_num = min(fid_num, len(self.val_loader.dataset))
+
+        if self.inception_model is None:
+            if align_tf:
+                self.logger.info(f'Building inception model '
+                                 f'(aligned with TensorFlow) ...')
+            else:
+                self.logger.info(f'Building inception model '
+                                 f'(using torchvision) ...')
+            self.inception_model = build_inception_model(align_tf).cuda()
+            self.logger.info(f'Finish building inception model.')
+
+        if z is not None:
+            assert isinstance(z, np.ndarray)
+            assert z.ndim == 2 and z.shape[1] == self.z_space_dim
+            fid_num = min(fid_num, z.shape[0])
+            z = torch.from_numpy(z).type(torch.FloatTensor)
+        if not fid_num:
+            return -1
+
+        indices = list(range(self.rank, fid_num, self.world_size))
+
+        self.logger.init_pbar()
+        #generator = self.run_with_optim if run_with_optim else self.run_without_optim
+        # Extract features from fake images.
+        fake_feature_list = []
+        real_feature_list = []
+        task1 = self.logger.add_pbar_task('FID', total=fid_num)
+        for batch_idx in range(0, len(indices), self.val_batch_size):
+            sub_indices = indices[batch_idx:batch_idx + self.val_batch_size]
+            batch_size = len(sub_indices)
+            data = next(self.val_loader)
+            for key in data:
+                if key != 'name':
+                    data[key] = data[key][:batch_size].cuda(
+                        torch.cuda.current_device(), non_blocking=True)
+
+            # z_rand = torch.randn((batch_size,self.config.z_count,self.config.latent_dim)
+            #                     , device=torch.cuda.current_device())
+            # data['z_rand'] = z_rand
+            with torch.no_grad():
+                real_images = data['image']
+                #valids = data['valid']
+                return_dict = self.forward_pass(data, return_vals='fakes, wp_mixed')
+                fakes = return_dict['fakes']
+            if self.config.test_time_optims != 0:
+                wp_mixed = return_dict['wp_mixed']
+                fakes = self.optimize(data, wp_mixed)
+            with torch.no_grad():
+                #final_out = real_images * valids + fakes * (1.0-valids) #Final output is the mixed one. 
+                fake_feature_list.append(
+                    extract_feature(self.inception_model, fakes))
+            # Extract features from real images if needed.
+            cached_fid_file = f'{self.work_dir}/real_fid{fid_num}.npy'
+            do_real_test = (not os.path.exists(cached_fid_file) or ignore_cache)
+            if do_real_test:
+                with torch.no_grad():
+                    real_feature_list.append(
+                            extract_feature(self.inception_model, real_images))
+            self.logger.update_pbar(task1, batch_size * self.world_size)
+
+        np.save(f'{self.work_dir}/fake_fid_features_{self.rank}.npy',
+                np.concatenate(fake_feature_list, axis=0))
+        if (do_real_test):
+            np.save(f'{self.work_dir}/real_fid_features_{self.rank}.npy',
+                        np.concatenate(real_feature_list, axis=0))
+
+        # Extract features from real images if needed.
+        # cached_fid_file = f'{self.work_dir}/real_fid{fid_num}.npy'
+        # do_real_test = (not os.path.exists(cached_fid_file) or ignore_cache)
+        # if do_real_test:
+        #     real_feature_list = []
+        #     task2 = self.logger.add_pbar_task("Real", total=fid_num)
+        #     for batch_idx in range(0, len(indices), self.val_batch_size):
+        #         sub_indices = indices[batch_idx:batch_idx + self.val_batch_size]
+        #         batch_size = len(sub_indices)
+        #         data = next(self.val_loader)
+        #         for key in data:
+        #             data[key] = data[key][:batch_size].cuda(
+        #                 torch.cuda.current_device(), non_blocking=True)
+        #         with torch.no_grad():
+        #             real_images = data['image']
+        #             real_feature_list.append(
+        #                 extract_feature(self.inception_model, real_images))
+        #         self.logger.update_pbar(task2, batch_size * self.world_size)
+        #     np.save(f'{self.work_dir}/real_fid_features_{self.rank}.npy',
+        #             np.concatenate(real_feature_list, axis=0))
+
+        dist.barrier()
+        if self.rank != 0:
+            return -1
+        self.logger.close_pbar()
+
+        # Collect fake features.
+        fake_feature_list.clear()
+        for rank in range(self.world_size):
+            fake_feature_list.append(
+                np.load(f'{self.work_dir}/fake_fid_features_{rank}.npy'))
+            os.remove(f'{self.work_dir}/fake_fid_features_{rank}.npy')
+        fake_features = np.concatenate(fake_feature_list, axis=0)
+        assert fake_features.ndim == 2 and fake_features.shape[0] == fid_num
+        feature_dim = fake_features.shape[1]
+        pad = fid_num % self.world_size
+        if pad:
+            pad = self.world_size - pad
+        fake_features = np.pad(fake_features, ((0, pad), (0, 0)))
+        fake_features = fake_features.reshape(self.world_size, -1, feature_dim)
+        fake_features = fake_features.transpose(1, 0, 2)
+        fake_features = fake_features.reshape(-1, feature_dim)[:fid_num]
+
+        # Collect (or load) real features.
+        if do_real_test:
+            real_feature_list.clear()
+            for rank in range(self.world_size):
+                real_feature_list.append(
+                    np.load(f'{self.work_dir}/real_fid_features_{rank}.npy'))
+                os.remove(f'{self.work_dir}/real_fid_features_{rank}.npy')
+            real_features = np.concatenate(real_feature_list, axis=0)
+            assert real_features.shape == (fid_num, feature_dim)
+            real_features = np.pad(real_features, ((0, pad), (0, 0)))
+            real_features = real_features.reshape(
+                self.world_size, -1, feature_dim)
+            real_features = real_features.transpose(1, 0, 2)
+            real_features = real_features.reshape(-1, feature_dim)[:fid_num]
+            np.save(cached_fid_file, real_features)
+        else:
+            real_features = np.load(cached_fid_file)
+            assert real_features.shape == (fid_num, feature_dim)
+
+        fid_value = compute_fid(fake_features, real_features)
+        return fid_value
+
+    def val(self, **val_kwargs):
+        self.synthesize(**val_kwargs)
+
+    def synthesize(self,
+                   num,
+                   html_name=None,
+                   save_raw_synthesis=False):
+        """Synthesizes images.
+
+        Args:
+            num: Number of images to synthesize.
+            z: Latent codes used for generation. If not specified, this function
+                will sample latent codes randomly. (default: None)
+            html_name: Name of the output html page for visualization. If not
+                specified, no visualization page will be saved. (default: None)
+            save_raw_synthesis: Whether to save raw synthesis on the disk.
+                (default: False)
+        """
+
+        dist.barrier()
+        if self.rank != 0:
+            return
+            
+        if not html_name and not save_raw_synthesis:
+            return
+
+        self.set_mode('val')
+
+        if self.val_loader is None:
+            self.build_dataset('val')
+
+        # temp_dir = os.path.join(self.work_dir, 'synthesize_results')
+        # os.makedirs(temp_dir, exist_ok=True)
+
+        if not num:
+            return
+        # if num % self.val_batch_size != 0:
+        #     num =  (num //self.val_batch_size +1)*self.val_batch_size
+        # TODO: Use same z during the entire training process.
+        
+        self.logger.init_pbar()
+        task = self.logger.add_pbar_task('Synthesis', total=num)
+        for i in range(num):
+            data = next(self.val_loader)
+            for key in data:
+                if key != 'name':
+                    data[key] = data[key].cuda(
+                        torch.cuda.current_device(), non_blocking=True)
+
+            with torch.no_grad():
+                real_images = data['image']
+                return_dict = self.forward_pass(data, return_vals='all')
+                fakes = return_dict['fakes']
+                wp_mixed = return_dict['wp_mixed']
+                eouts = return_dict['eouts']
+
+                log_list_gpu = {"real": real_images, "fake": fakes}
+
+                # Add editings to log_list
+                editings = ['age', 'pose', 'smile']
+                for edit in editings:
+                    direction = torch.load(f'editings/interfacegan_directions/{edit}.pt').cuda()
+                    factors = [+3, -3]
+                    for factor in factors:
+                        name = f"{edit}_{factor}"
+                        edit_wp = wp_mixed + factor * direction
+                        edited_images, _ = self.runG(edit_wp, "synthesis", highres_outs=eouts)
+                        # if edit == 'smile' and factor == -3:
+                        #     res = gouts['gates'].shape[-1]
+                        #     log_list_gpu[f'smile_-3_gate'] = ( torch.mean((gouts_edits['gates']) , dim=1, keepdim=True), 0)
+                        #edited_images = F.adaptive_avg_pool2d(edited_images, 256)
+                        log_list_gpu[name] = edited_images
+                        #log_list_gpu[f'{name}_gate'] = ( torch.mean((temp['gates']) , dim=1, keepdim=True), 0)
+
+                #Add gate to log_list
+                # res = gouts['gates'].shape[-1]
+                # log_list_gpu[f'gate{res}x{res}'] = ( torch.mean((gouts['gates']) , dim=1, keepdim=True), 0)
+            
+                #Log images
+                for log_name, log_val in log_list_gpu.items():
+                    log_im = log_val[0] if type(log_val) is tuple else log_val
+                    min_val = log_val[1] if type(log_val) is tuple else -1
+                    cpu_img = postprocess_tensor(log_im.detach().cpu(), min_val=min_val)
+                    grid = torchvision.utils.make_grid(cpu_img, nrow=5)
+                    log_image( name = f"image/{log_name}", grid=grid, iter=self.iter)
+            self.logger.update_pbar(task, 1)
+        self.logger.close_pbar()
+
+    def save_edited_images(self, opts):
+        dist.barrier()
+        if self.rank != 0:
+            return
+        self.set_mode('val')
+
+        if opts.method  == 'inversion':
+            pass
+        elif opts.method == 'interfacegan':
+            direction = torch.load(f'editings/interfacegan_directions/{opts.edit}.pt').cuda()
+        elif opts.method == 'ganspace':
+            ganspace_pca = torch.load('editings/ganspace_pca/ffhq_pca.pt')
+            direction = {
+                'eye_openness':            (54,  7,  8,  20),
+                'smile':                   (46,  4,  5, -20),
+                'beard':                   (58,  7,  9,  -20),
+                'white_hair':              (57,  7, 10, -24),
+                'lipstick':                (34, 10, 11,  20),
+                'overexposed':             (27,  8, 18,  15),
+                'screaming':               (35,  3,  7, -10),
+                'head_angle_up':           (11,  1,  4,  10),
+            }
+            editor = LatentEditor()
+        elif opts.method == 'styleclip':
+            #model_path = '/media/hdd2/adundar/hamza/hyperstyle/pretrained_models/stylegan2-ffhq-config-f.pt'
+            # calculator_args = {
+		    # 'delta_i_c': 'editings/styleclip/global_directions/ffhq/fs3.npy',
+		    # 's_statistics': 'editings/styleclip/global_directions/ffhq/S_mean_std',
+		    # 'text_prompt_templates': 'editings/styleclip/global_directions/templates.txt'
+	        #  }
+            stylegan_model = load_stylegan_generator(opts.model_path)
+            global_direction_calculator = load_direction_calculator(opts.calculator_args)
+            #Eyeglasses 5, bangs 2, bobcut 5
+            # edit_args = {'alpha_min': 2, 'alpha_max': 2, 'num_alphas':1, 'beta_min':0.11, 'beta_max':0.11, 'num_betas': 1,
+            #     'neutral_text':'face', 'target_text': 'face with bangs'}
+    
+        
+        self.config.data['val']['root_dir'] = opts.dataset
+
+        dataset = BaseDataset(**self.config.data['val'])
+        val_loader = DataLoader(dataset, batch_size=1, shuffle=False, num_workers=1)
+
+        temp_dir = os.path.join(self.work_dir, opts.output)
+        os.makedirs(temp_dir, exist_ok=True)
+ 
+
+        self.logger.init_pbar()
+        task = self.logger.add_pbar_task('Synthesis', total=len(val_loader))
+        global_i = 0
+        all_latents = {}
+        for idx, data in enumerate(val_loader):
+            for key in data:
+                if key != 'name':
+                    data[key] = data[key].cuda(
+                        torch.cuda.current_device(), non_blocking=True)
+            with torch.no_grad():
+                return_dict = self.forward_pass(data, return_vals='all', only_enc=True)
+                wp_mixed = return_dict['wp_mixed']
+                eouts = return_dict['eouts']
+                #fakes = return_dict['fakes']
+                factors = np.linspace(0, 3, 100)
+                global_i = 0
+
+                # for factor in factors:
+                if opts.method == 'interfacegan':
+                    wp_mixed = wp_mixed + opts.factor * direction
+                if opts.method == 'ganspace':
+                    #interpolate_dir = direction[opts.edit][0:3] + (factor,) 
+                    wp_mixed = editor.apply_ganspace_(wp_mixed, ganspace_pca, [direction[opts.edit]])
+                    #wp_edit = editor.apply_ganspace_(wp_mixed, ganspace_pca, [interpolate_dir])
+
+
+                # z = torch.randn((1,self.config.latent_dim), device=torch.cuda.current_device())
+                # z = self.runM(z)
+                # diff = z - wp_mixed
+                # edit = (diff * 3.5) / 10
+                # wp_mixed = wp_mixed + edit
+                edited_images, gouts_edits = self.runG(wp_mixed, "synthesis", highres_outs=eouts, resize=False)
+                if opts.method == 'styleclip':
+                    # opts.edit_args['alpha_min'] = factor
+                    edited_images = styleclip_edit(wp_mixed, gouts_edits['additions'], stylegan_model, global_direction_calculator, opts.edit_args)
+                edited_images = T.Resize((256,256))(edited_images)
+                edited_images = postprocess_image(edited_images.detach().cpu().numpy())
+                for j in range(edited_images.shape[0]):
+                    # dir_name = data['name'][j][:-4]
+                    # os.makedirs(os.path.join(temp_dir, dir_name), exist_ok=True)
+                    # save_name = f'{global_i:03d}_' + data['name'][j]
+                    save_name = data['name'][j]
+                    pil_img = Image.fromarray(edited_images[j]) #.resize((256,256))
+                    #pil_img.save(os.path.join(temp_dir, dir_name,  save_name ))
+                    pil_img.save(os.path.join(temp_dir,  save_name ))
+                    global_i += 1
+                # if global_i >= 1000:
+                #     break
+                # if global_i % 100 == 0:
+                #     print(f"{global_i}/1000")
+
+            self.logger.update_pbar(task, 1)
+        self.logger.close_pbar()
+
+    # def forward_pass(self, data, return_vals='all', only_enc=False):
+    #     encoder_type = self.config.encoder_type
+    #     forward_func = getattr(self, f'{encoder_type}_forward')
+    #     return_dict = forward_func(data,only_enc)
+    #     return return_dict
+    #     # if return_vals == 'all':
+    #     #     return return_dict
+    #     # requested = return_vals.split(',')
+    #     # modified_dict = {}
+    #     # for request in requested:
+    #     #     stripped_request = request.strip()
+    #     #     modified_dict[stripped_request] = return_dict[stripped_request]
+    #     # return modified_dict
+    
+    # def base_forward(self, data):
+    #     reals = data['image']
+    #     valids = data['valid']
+    #     z_rand = data['z_rand']
+
+    #     wp_rand = self.runM(z_rand)
+    #     wp_enc, blender = self.runE(reals, valids)
+    #     wp_mixed = self.mix(wp_enc, wp_rand, blender)
+    #     fakes = self.runG(wp_mixed, 'synthesis')
+    #     return_dict = {'fakes': fakes, 'wp_enc': wp_enc, 'blender': blender, 'wp_mixed':wp_mixed}
+    #     return return_dict
+
+    # def pSp_forward(self, data, only_enc):
+    #     return self.e4e_forward(data, only_enc)
+
+    # def train_forward(self, data, iscycle=False):
+    #     reals = data['image']
+    #     direction = data['direction']
+    #     edit_name = data['edit_name']
+    #     factor = data['factor']
+    #     E = self.models['encoder']
+    #     with torch.no_grad():
+    #         wp, eouts = E(reals)
+    #     #wp = wp + self.meanw.repeat(reals.shape[0], 1, 1)
+    #     edit = torch.zeros_like(wp)
+    #     for i in range (edit.shape[0]):
+    #         if edit_name[i] is None:
+    #             edit[i] = 0
+    #         elif edit_name[i] == 'randw':
+    #             diff = direction[i] - wp[i]
+    #             # one_hot = [1] * 8 + [0] * 10
+    #             # one_hot = torch.tensor(one_hot, device=diff.device).unsqueeze(1)
+    #             # diff = diff * one_hot   
+    #             #norm = torch.linalg.norm(diff, dim=1, keepdim=True)
+    #             edit[i] = (diff * factor[i]) / 10
+    #         elif edit_name[i] == 'interface':
+    #             edit[i] = (factor[i] * direction[i])
+
+    #     # # Debug
+    #     # with torch.no_grad():
+    #     #     fakes,_ =self.runG(wp, 'synthesis', highres_outs=None)
+    #     #     fakes = postprocess_image(fakes.detach().cpu().numpy())
+    #     #     for i in range(fakes.shape[0]):
+    #     #         pil_img = Image.fromarray(fakes[i]).resize((256,256))
+    #     #         pil_img.save(f'{self.iter}_orig.png')
+
+    #     #     fakes,_ =self.runG(wp+edit, 'synthesis', highres_outs=None)
+    #     #     fakes = postprocess_image(fakes.detach().cpu().numpy())
+    #     #     for i in range(fakes.shape[0]):
+    #     #         pil_img = Image.fromarray(fakes[i]).resize((256,256))
+    #     #         pil_img.save(f'{self.iter}_edit.png')
+
+    #     #     fakes,_ =self.runG(direction.unsqueeze(1).repeat(1,18,1), 'synthesis', highres_outs=None)
+    #     #     fakes = postprocess_image(fakes.detach().cpu().numpy())
+    #     #     for i in range(fakes.shape[0]):
+    #     #         pil_img = Image.fromarray(fakes[i]).resize((256,256))
+    #     #         pil_img.save(f'{self.iter}_rand.png')
+    #     with torch.no_grad():
+    #         eouts['inversion'] = self.runG(wp, 'synthesis', highres_outs=None, return_f=True)
+    #     wp = wp + edit
+    #     fakes, gouts = self.runG(wp, 'synthesis', highres_outs=eouts)
+    #     #fakes = F.adaptive_avg_pool2d(fakes, (256,256))
+    #     fakes_cycle = None
+    #     if iscycle:
+    #         # wp_cycle = wp_cycle + self.meanw.repeat(reals.shape[0], 1, 1)
+    #         with torch.no_grad():
+    #             wp_cycle, eout_cycle = E(fakes)
+    #             eout_cycle['inversion'] = self.runG(wp_cycle, 'synthesis', highres_outs=None, return_f=True)
+
+    #         #wp_cycle = wp_cycle - edit
+    #         wp_cycle = wp_cycle - edit
+    #         #wp_cycle = wp_cycle - (data['factor'] * data['direction']).unsqueeze(1)
+    #         fakes_cycle, _ = self.runG(wp_cycle, 'synthesis', highres_outs=eout_cycle)
+    #         #fakes_cycle = F.adaptive_avg_pool2d(fakes, (256,256))
+    #         #cycle = F.mse_loss(fakes_cycle, reals, reduction='mean')
+    #     return_dict = {'fakes': fakes, 'wp_mixed':wp, 'gouts':gouts, 'eouts': eouts, 'cycle': fakes_cycle}
+    #     return return_dict
+
+    # def e4e_forward(self, data, only_enc=False):
+    #     #return self.base_forward(data)
+    #     reals = data['image']
+    #     #valids = data['valid']
+    #     E = self.models['encoder']
+    #     wp_mixed, eouts = E(reals)
+    #     #wp_mixed = wp_mixed + self.meanw.repeat(reals.shape[0], 1, 1)
+    #     eouts['inversion'] = self.runG(wp_mixed, 'synthesis', highres_outs=None, return_f=True)
+    #     if only_enc:
+    #         return_dict = {'wp_mixed':wp_mixed,'eouts': eouts}
+    #         return return_dict
+    #     fakes, gouts = self.runG(wp_mixed, 'synthesis', highres_outs=eouts)
+    #     #fakes = self.runG(wp_mixed, 'synthesis', highres_outs=None)
+    #     #fakes = F.adaptive_avg_pool2d(fakes, (256,256))
+    #     return_dict = {'fakes': fakes, 'wp_mixed':wp_mixed, 'gouts':gouts, 'eouts': eouts}
+    #     return return_dict
+
+
+    # def hyperstyle_forward(self, data):
+    #     return_dict = self.base_forward(data)
+        
+    #     E = self.models['encoder']
+    #     reals = data['image']
+    #     valids = data['valid']
+
+    #     #HyperNetwork 
+    #     weight_deltas = E(reals, valids, mode='hyper', gouts=return_dict['fakes'])
+    #     fakes = self.runG(return_dict['wp_mixed'], 'synthesis', weight_deltas=weight_deltas)
+    #     return_dict['fakes'] = fakes
+
+    #     return return_dict
+
+    def interface_generate(self, num, edit, factor):
+        direction = torch.load(f'editings/interfacegan_directions/{edit}.pt').cuda()
+        indices = list(range(self.rank, num, self.world_size))
+        gt_path = os.path.join(self.work_dir, f'interfacegan_gt')
+        smile_add_path = os.path.join(self.work_dir, f'interfacegan_{edit}_{factor}')
+        smile_rm_path = os.path.join(self.work_dir, f'interfacegan_{edit}_-{factor}')
+        if self.rank == 0:
+            os.makedirs(gt_path, exist_ok=True)
+            os.makedirs(smile_add_path, exist_ok=True)
+            os.makedirs(smile_rm_path, exist_ok=True)
+        dist.barrier()
+
+        self.logger.init_pbar()
+        task = self.logger.add_pbar_task('Interfacegan', total=num)
+
+
+        for batch_idx in range(0, len(indices), self.val_batch_size):
+            sub_indices = indices[batch_idx:batch_idx + self.val_batch_size]
+            batch_size = len(sub_indices)
+
+            z = torch.randn((batch_size,512), device=torch.cuda.current_device())
+            w_r = self.runM(z, repeat_w=True)
+            gt_imgs,_ = self.runG(w_r, resize=False)
+            gt_imgs = postprocess_image(gt_imgs.detach().cpu().numpy())
+            for i in range(gt_imgs.shape[0]):
+                save_name = str(sub_indices[i]) + ".png"
+                pil_img = Image.fromarray(gt_imgs[i]).resize((256,256))
+                pil_img.save(os.path.join(gt_path, save_name ))
+            
+            smile_added, _ = self.runG(w_r + factor*direction, resize=False)
+            smile_added = postprocess_image(smile_added.detach().cpu().numpy())
+            for i in range(gt_imgs.shape[0]):
+                save_name = str(sub_indices[i]) + ".png"
+                pil_img = Image.fromarray(smile_added[i]).resize((256,256))
+                pil_img.save(os.path.join(smile_add_path, save_name ))
+
+            smile_removed, _= self.runG(w_r - factor*direction, resize=False)
+            smile_removed = postprocess_image(smile_removed.detach().cpu().numpy())
+            for i in range(gt_imgs.shape[0]):
+                save_name = str(sub_indices[i]) + ".png"
+                pil_img = Image.fromarray(smile_removed[i]).resize((256,256))
+                pil_img.save(os.path.join(smile_rm_path, save_name ))
+
+            self.logger.update_pbar(task, batch_size * self.world_size)
+        self.logger.close_pbar()
+
+    def grad_edit(self, edit, factor, dataset=None):
+        dist.barrier()
+        if self.rank != 0:
+            return
+        self.set_mode('val')
+        edit_name = edit
+        edit = 'val'
+        self.config.data[edit]['root_dir'] = dataset
+
+        dataset = BaseDataset(**self.config.data[edit])
+        val_loader = DataLoader(dataset, batch_size=1, shuffle=False, num_workers=0)
+
+        temp_dir = os.path.join(self.work_dir, f'fakes_{edit_name}_{factor}')
+        os.makedirs(temp_dir, exist_ok=True)
+
+        self.logger.init_pbar()
+        task = self.logger.add_pbar_task('Synthesis', total=len(val_loader))
+        global_i = 0
+        args = {'model': 'ffhq', 'model_dir': '/media/hdd2/adundar/hamza/genforce/editings/GradCtrl/model_ffhq',
+                'attribute': edit_name, 'exclude': 'default', 'top_channels': 'default', 'layerwise': 'default' }
+        for idx, data in enumerate(val_loader):
+            for key in data:
+                if key != 'name':
+                    data[key] = data[key].cuda(
+                        torch.cuda.current_device(), non_blocking=True)
+            with torch.no_grad():
+                return_dict = self.forward_pass(data, return_vals='all', only_enc=True)
+                wp_mixed = return_dict['wp_mixed']
+                eouts = return_dict['eouts']
+
+            #fakes = return_dict['fakes']
+            edit_wp = gradctrl(args, wp_mixed, factor)
+            edited_images, gouts_edits = self.runG(edit_wp, "synthesis", highres_outs=eouts, resize=False)
+            #edited_images, gouts_edits = self.runG(wp_mixed, "synthesis", highres_outs=eouts, resize=False)
+            edited_images = postprocess_image(edited_images.detach().cpu().numpy())
+            for j in range(edited_images.shape[0]):
+                save_name = data['name'][j]
+                pil_img = Image.fromarray(edited_images[j]).resize((256,256))
+                pil_img.save(os.path.join(temp_dir, save_name ))
+                global_i += 1
+
+            self.logger.update_pbar(task, 1)
+        self.logger.close_pbar()
+
+    def measure_time(self, edit, factor, dataset=None, save_latents=False):
+            dist.barrier()
+            if self.rank != 0:
+                return
+            self.set_mode('val')
+            edit_name = edit
+            if dataset is not None:
+                edit = 'val'
+                self.config.data[edit]['root_dir'] = dataset
+
+            dataset = BaseDataset(**self.config.data[edit])
+            val_loader = DataLoader(dataset, batch_size=1, shuffle=False, num_workers=0)
+
+            global_i = 0
+            time_list = []
+            for idx, data in enumerate(val_loader):
+                for key in data:
+                    if key != 'name':
+                        data[key] = data[key].cuda(
+                            torch.cuda.current_device(), non_blocking=True)
+                with torch.no_grad():
+                    start = time.time()
+                    return_dict = self.forward_pass(data, return_vals='all', only_enc=True)
+                    wp_mixed = return_dict['wp_mixed']
+                    eouts = return_dict['eouts']
+                    edited_images, gouts_edits = self.runG(wp_mixed, "synthesis", highres_outs=eouts, resize=False)
+                    end = time.time()
+                    time_list.append(end-start)
+                    print(np.mean(time_list))
+            print(np.mean(time_list[1:]))
+
+            

editings/editor.py

@@ -0,0 +1,23 @@
+from .interfacegan import InterFaceGAN
+from .ganspace import GanSpace
+from .styleclip import StyleClip
+from options import Settings
+
+"""
+Entry class for all the edits. 
+"""
+class Editor():
+    def __init__(self) -> None:
+        self.interfacegan_editor = InterFaceGAN()
+        self.ganspace_editor = GanSpace()
+        self.styleclip_editor = StyleClip()
+
+    def edit(self, latent, cfg):
+        # Finds the corresponding function using method name
+        if cfg.method == 'inversion':
+            return latent
+
+        editor = getattr(self, f'{cfg.method}_editor')
+        return editor.edit(latent, cfg)
+
+        

editings/ganspace.py

@@ -0,0 +1,49 @@
+import torch
+import csv
+from options import Settings
+import os
+
+class GanSpace():
+    def __init__(self) -> None:
+
+        self.gan_space_configs = {}
+        
+        with open(os.path.join(Settings.ganspace_directions, 'ganspace_configs.csv'), "r") as f:
+            reader = csv.reader(f, delimiter="\t")
+            for row in reader:
+                key = row.pop(0)
+                self.gan_space_configs[key] = list(map(int, row))
+
+    def edit(self, latent, cfg):
+        with torch.no_grad():
+            self.load_ganspace_pca()
+            gan_space_config = self.gan_space_configs[cfg.edit]
+            gan_space_config[-1] = cfg.strength
+            return self.edit_ganspace(latent, gan_space_config)
+
+    def load_ganspace_pca(self):
+        try:   # Check if loaded
+            getattr(self, f"pca")
+        except:
+            pca = torch.load(os.path.join(Settings.ganspace_directions, 'ffhq_pca.pt'))
+            setattr(self, f"pca", pca)
+        
+
+    def edit_ganspace(self, latents, config):
+        edit_latents = []
+        pca_idx, start, end, strength = config
+        for latent in latents:
+            delta = self.get_delta( latent, pca_idx, strength)
+            delta_padded = torch.zeros(latent.shape).to(Settings.device)
+            delta_padded[start:end] += delta.repeat(end - start, 1)
+            edit_latents.append(latent + delta_padded)
+        return torch.stack(edit_latents)
+
+    def get_delta(self, latent, idx, strength):
+        # pca: ganspace checkpoint. latent: (16, 512) w+
+        w_centered = latent - self.pca['mean'].to(Settings.device)
+        lat_comp = self.pca['comp'].to(Settings.device)
+        lat_std = self.pca['std'].to(Settings.device)
+        w_coord = torch.sum(w_centered[0].reshape(-1)*lat_comp[idx].reshape(-1)) / lat_std[idx]
+        delta = (strength - w_coord)*lat_comp[idx]*lat_std[idx]
+        return delta

editings/ganspace_pca/ganspace_configs.csv

@@ -0,0 +1,42 @@
+frizzy_hair	31	2	6	20
+background_blur	49	6	9	20
+bald	21	2	5	20
+big_smile	19	4	5	20
+caricature_smile	26	3	8	13
+scary_eyes	33	6	8	20
+curly_hair	47	3	6	20
+dark_bg_shiny_hair	13	8	9	20
+dark_hair_and_light_pos	14	8	9	20
+dark_hair	16	8	9	20
+disgusted	43	6	8	-30
+displeased	36	4	7	20
+eye_openness	54	7	8	20
+eye_wrinkles	28	6	8	20
+eyebrow_thickness	37	8	9	20
+face_roundness	37	0	5	20
+fearful_eyes	54	4	10	20
+hairline	21	4	5	-20
+happy_frizzy_hair	30	0	8	20
+happy_elderly_lady	27	4	7	20
+head_angle_up	11	1	4	20
+huge_grin	28	4	6	20
+in_awe	23	3	6	-15
+wide_smile	23	3	6	20
+large_jaw	22	3	6	20
+light_lr	15	8	9	10
+lipstick_and_age	34	6	11	20
+lipstick	34	10	11	20
+mascara_vs_beard	41	6	9	20
+nose_length	51	4	5	-20
+elderly_woman	34	6	7	20
+overexposed	27	8	18	15
+screaming	35	3	7	-15
+short_face	32	2	6	-20
+show_front_teeth	59	4	5	40
+smile	46	4	5	-20
+straight_bowl_cut	20	4	5	-20
+sunlight_in_face	10	8	9	10
+trimmed_beard	58	7	9	20
+white_hair	57	7	10	-24
+wrinkles	20	6	7	-18
+boyishness	8	2	5	20

editings/interfacegan.py

@@ -0,0 +1,27 @@
+import torch
+from options import Settings
+import os
+
+class InterFaceGAN():
+    def __init__(self) -> None:
+        pass
+
+    def edit(self, latent, cfg):
+        with torch.no_grad():
+            return latent + cfg.strength * self.get_direction(cfg.edit)
+
+    def get_direction(self, editname):
+        try:
+            direction = getattr(self, f"{editname}_direction")
+        except:
+            direction = self.load_direction(editname)
+            if Settings.device != 'cpu':
+                direction = direction.to(Settings.device)
+            setattr(self, f"{editname}_direction", direction.clone())
+        return direction
+
+    def load_direction(self, editname):
+        direction = torch.load(os.path.join( Settings.interfacegan_directions, f'{editname}.pt'))
+        if Settings.device != 'cpu':
+            direction = direction.cuda()
+        return direction

editings/styleclip.py

@@ -0,0 +1,49 @@
+
+from editings.styleclip_directions.styleclip_mapper_network import LevelsMapper
+import torch
+import csv
+from options import Settings
+import os
+
+class Options():
+    def __init__(self, no_coarse_mapper, no_medium_mapper, no_fine_mapper) -> None:
+        self.no_coarse_mapper = no_coarse_mapper
+        self.no_medium_mapper = no_medium_mapper
+        self.no_fine_mapper = no_fine_mapper
+
+class StyleClip():
+    def __init__(self) -> None:
+        self.styleclip_mapping_configs = {}
+        
+        with open(os.path.join(Settings.styleclip_settings, 'styleclip_mapping_configs.csv'), "r") as f:
+            reader = csv.reader(f)
+            for row in reader:
+                key = row.pop(0)
+                self.styleclip_mapping_configs[key] = list(map(lambda x: True if x == "True" else False, row))
+        
+    def edit(self, latent, cfg):
+        with torch.no_grad():
+            if cfg.type == 'mapper':
+                mapper = self.build_mapper(cfg.edit)
+                return latent + cfg.strength * mapper(latent)
+            if cfg.type == 'global':
+                
+                return latent + 10 * torch.load(os.path.join(Settings.styleclip_global_directions, 'makeup.pt'))
+
+    # def load_global_direction(self, editname):
+    #     pass
+
+    def build_mapper(self, editname):
+        try:   # Check if loaded
+            mapper = getattr(self, f"{editname}_mapper")
+        except:
+            opts = Options(*self.styleclip_mapping_configs[editname])
+            mapper = LevelsMapper(opts)
+            ckpt = torch.load(os.path.join(Settings.styleclip_mapper_directions, f'{editname}.pt'))
+            mapper.load_state_dict(ckpt, strict=True)
+            mapper.to(device=Settings.device)
+            for param in mapper.parameters():
+                param.requires_grad = False
+            mapper.eval()
+            setattr(self, f"{editname}_mapper", mapper)
+        return mapper

editings/styleclip_directions/styleclip_directions/global_directions/templates.txt

@@ -0,0 +1,79 @@
+a bad photo of a {}.
+a sculpture of a {}.
+a photo of the hard to see {}.
+a low resolution photo of the {}.
+a rendering of a {}.
+graffiti of a {}.
+a bad photo of the {}.
+a cropped photo of the {}.
+a tattoo of a {}.
+the embroidered {}.
+a photo of a hard to see {}.
+a bright photo of a {}.
+a photo of a clean {}.
+a photo of a dirty {}.
+a dark photo of the {}.
+a drawing of a {}.
+a photo of my {}.
+the plastic {}.
+a photo of the cool {}.
+a close-up photo of a {}.
+a black and white photo of the {}.
+a painting of the {}.
+a painting of a {}.
+a pixelated photo of the {}.
+a sculpture of the {}.
+a bright photo of the {}.
+a cropped photo of a {}.
+a plastic {}.
+a photo of the dirty {}.
+a jpeg corrupted photo of a {}.
+a blurry photo of the {}.
+a photo of the {}.
+a good photo of the {}.
+a rendering of the {}.
+a {} in a video game.
+a photo of one {}.
+a doodle of a {}.
+a close-up photo of the {}.
+a photo of a {}.
+the origami {}.
+the {} in a video game.
+a sketch of a {}.
+a doodle of the {}.
+a origami {}.
+a low resolution photo of a {}.
+the toy {}.
+a rendition of the {}.
+a photo of the clean {}.
+a photo of a large {}.
+a rendition of a {}.
+a photo of a nice {}.
+a photo of a weird {}.
+a blurry photo of a {}.
+a cartoon {}.
+art of a {}.
+a sketch of the {}.
+a embroidered {}.
+a pixelated photo of a {}.
+itap of the {}.
+a jpeg corrupted photo of the {}.
+a good photo of a {}.
+a plushie {}.
+a photo of the nice {}.
+a photo of the small {}.
+a photo of the weird {}.
+the cartoon {}.
+art of the {}.
+a drawing of the {}.
+a photo of the large {}.
+a black and white photo of a {}.
+the plushie {}.
+a dark photo of a {}.
+itap of a {}.
+graffiti of the {}.
+a toy {}.
+itap of my {}.
+a photo of a cool {}.
+a photo of a small {}.
+a tattoo of the {}.

editings/styleclip_directions/styleclip_mapper_network.py

@@ -0,0 +1,121 @@
+import torch
+from torch import nn
+from torch.nn import Module
+from torch.nn import functional as F
+import math
+
+def fused_leaky_relu(input, bias, negative_slope=0.2, scale=2 ** 0.5):
+    rest_dim = [1] * (input.ndim - bias.ndim - 1)
+    input = input   #.cuda()
+    if input.ndim == 3:
+        return (
+            F.leaky_relu(
+                input + bias.view(1, *rest_dim, bias.shape[0]), negative_slope=negative_slope
+            )
+            * scale
+        )
+    else:
+        return (
+            F.leaky_relu(
+                input + bias.view(1, bias.shape[0], *rest_dim), negative_slope=negative_slope
+            )
+            * scale
+        )
+
+
+class PixelNorm(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, input):
+        return input * torch.rsqrt(torch.mean(input ** 2, dim=1, keepdim=True) + 1e-8)
+
+class EqualLinear(nn.Module):
+    def __init__(
+        self, in_dim, out_dim, bias=True, bias_init=0, lr_mul=1, activation=None
+    ):
+        super().__init__()
+
+        self.weight = nn.Parameter(torch.randn(out_dim, in_dim).div_(lr_mul))
+
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_dim).fill_(bias_init))
+
+        else:
+            self.bias = None
+
+        self.activation = activation
+
+        self.scale = (1 / math.sqrt(in_dim)) * lr_mul
+        self.lr_mul = lr_mul
+
+    def forward(self, input):
+        if self.activation:
+            out = F.linear(input, self.weight * self.scale)
+            out = fused_leaky_relu(out, self.bias * self.lr_mul)
+
+        else:
+            out = F.linear(
+                input, self.weight * self.scale, bias=self.bias * self.lr_mul
+            )
+
+        return out
+
+class Mapper(Module):
+
+    def __init__(self, latent_dim=512):
+        super(Mapper, self).__init__()
+
+        layers = [PixelNorm()]
+
+        for i in range(4):
+            layers.append(
+                EqualLinear(
+                    latent_dim, latent_dim, lr_mul=0.01, activation='fused_lrelu'
+                )
+            )
+
+        self.mapping = nn.Sequential(*layers)
+
+
+    def forward(self, x):
+        x = self.mapping(x)
+        return x
+
+
+class LevelsMapper(Module):
+
+    def __init__(self, opts):
+        super(LevelsMapper, self).__init__()
+
+        self.opts = opts
+
+        if not opts.no_coarse_mapper:
+            self.course_mapping = Mapper()
+        if not opts.no_medium_mapper:
+            self.medium_mapping = Mapper()
+        if not opts.no_fine_mapper:
+            self.fine_mapping = Mapper()
+
+    def forward(self, x):
+        x_coarse = x[:, :4, :]
+        x_medium = x[:, 4:8, :]
+        x_fine = x[:, 8:, :]
+
+        if not self.opts.no_coarse_mapper:
+            x_coarse = self.course_mapping(x_coarse)
+        else:
+            x_coarse = torch.zeros_like(x_coarse)
+        if not self.opts.no_medium_mapper:
+            x_medium = self.medium_mapping(x_medium)
+        else:
+            x_medium = torch.zeros_like(x_medium)
+        if not self.opts.no_fine_mapper:
+            x_fine = self.fine_mapping(x_fine)
+        else:
+            x_fine = torch.zeros_like(x_fine)
+
+
+        out = torch.cat([x_coarse, x_medium, x_fine], dim=1)
+
+        return out

editings/styleclip_directions/styleclip_mapping_configs.csv

@@ -0,0 +1,14 @@
+afro,False,False,True
+angry,False,False,True
+beyonce,False,False,False
+bobcut,False,False,True
+bowlcut,False,False,True
+curly_hair,False,False,True
+hilary_clinton,False,False,False
+depp,False,False,False
+mohawk,False,False,True
+purple_hair,False,False,False
+surprised,False,False,True
+taylor_swift,False,False,False
+trump,False,False,False
+zuckerberg,False,False,False

inference.py

@@ -0,0 +1,71 @@
+import os
+
+from PIL import Image
+import torch
+from torch.utils.data import DataLoader
+
+from datasets.inference_dataset import InferenceDataset
+from datasets.process_image import ImageProcessor
+from models.styleres import StyleRes
+from options.inference_options import InferenceOptions
+from options import Settings
+from utils import parse_config
+from tqdm import tqdm
+
+def initialize_styleres(checkpoint_path, device):
+    Settings.device = device
+    model = StyleRes()
+    model.load_ckpt(checkpoint_path)
+    model.send_to_device()
+    model.eval()
+    for param in model.parameters():
+        param.requires_grad = False
+    return model
+
+def run():
+    args = InferenceOptions().parse()
+    edit_configs = parse_config(args.edit_configs)
+    if torch.cuda.is_available():
+        device = torch.device('cuda'  if torch.cuda.is_available() else 'cpu')
+    
+    dataset = InferenceDataset(args.datadir, aligner_path=args.aligner_path)
+    print(f"Dataset is created. Number of images is {len(dataset)}")
+    dataloader = DataLoader(dataset, batch_size = args.test_batch_size,
+                            shuffle=False,
+                            num_workers=int(args.test_workers),
+                            drop_last=False)
+
+    if args.n_images == None:
+        args.n_images = len(dataset)
+
+    # Create output directories
+    output_dir = args.outdir
+    os.makedirs(output_dir, exist_ok=True)
+    for edit_config in edit_configs:
+        cfg_vals = edit_config.values()
+        edit_config.outdir = '_'.join( str(i) for i in cfg_vals)
+        os.makedirs( os.path.join(output_dir, edit_config.outdir), exist_ok=True)
+
+    resize_amount = (1024, 1024)
+    if args.resize_outputs:
+        resize_amount = (256,256)
+    
+    # Setup model
+    model = initialize_styleres(args.checkpoint_path, device)
+        
+    n_images = 0
+    for data in tqdm(dataloader):
+        if n_images >= args.n_images:
+            break
+        n_images = n_images + data['image'].shape[0]
+        for edit_config in edit_configs:
+            images = model.edit_images( data['image'], edit_config)
+            images = ImageProcessor.postprocess_image(images.detach().cpu().numpy())
+            for j in range( images.shape[0]):
+                save_name = data['name'][j]
+                pil_img = Image.fromarray(images[j]).resize(resize_amount)
+                pil_img.save(os.path.join(output_dir,  edit_config.outdir, save_name))
+
+
+if __name__ == '__main__':
+    run()

models/dnnlib/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+from .util import EasyDict

models/dnnlib/util.py

@@ -0,0 +1,29 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Miscellaneous utility classes and functions."""
+
+# Util classes
+# ------------------------------------------------------------------------------------------
+
+class EasyDict(dict):
+    """Convenience class that behaves like a dict but allows access with the attribute syntax."""
+
+    def __getattr__(self, name: str):
+        try:
+            return self[name]
+        except KeyError:
+            raise AttributeError(name)
+
+    def __setattr__(self, name, value):
+        self[name] = value
+
+    def __delattr__(self, name):
+        del self[name]
+
+

models/e4e.py

@@ -0,0 +1,348 @@
+import math
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from collections import namedtuple
+
+def _upsample_add(x, y):
+    _, _, H, W = y.size()
+    return F.interpolate(x, size=(H, W), mode='bilinear', align_corners=True) + y
+
+
+
+class EqualLinear(nn.Module):
+    def __init__(
+            self, in_dim, out_dim, bias=True, bias_init=0, lr_mul=1, activation=None
+    ):
+        super().__init__()
+
+        self.weight = nn.Parameter(torch.randn(out_dim, in_dim).div_(lr_mul))
+
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_dim).fill_(bias_init))
+
+        else:
+            self.bias = None
+
+        self.activation = activation
+
+        self.scale = (1 / math.sqrt(in_dim)) * lr_mul
+        self.lr_mul = lr_mul
+
+    def forward(self, input):
+        # if self.activation:
+        #     out = F.linear(input, self.weight * self.scale)
+        #     out = fused_leaky_relu(out, self.bias * self.lr_mul)
+
+        # else:
+        out = F.linear(
+            input, self.weight * self.scale, bias=self.bias * self.lr_mul
+        )
+
+        return out
+
+    def __repr__(self):
+        return (
+            f'{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]})'
+        )
+
+class GradualStyleBlock(nn.Module):
+    def __init__(self, in_c, out_c, spatial):
+        super(GradualStyleBlock, self).__init__()
+        self.out_c = out_c
+        self.spatial = spatial
+        num_pools = int(np.log2(spatial))
+        modules = []
+        modules += [nn.Conv2d(in_c, out_c, kernel_size=3, stride=2, padding=1),
+                    nn.LeakyReLU()]
+        for i in range(num_pools - 1):
+            modules += [
+                nn.Conv2d(out_c, out_c, kernel_size=3, stride=2, padding=1),
+                nn.LeakyReLU()
+            ]
+        self.convs = nn.Sequential(*modules)
+        self.linear = EqualLinear(out_c, out_c, lr_mul=1)
+
+    def forward(self, x):
+        x = self.convs(x)
+        x = x.view(-1, self.out_c)
+        x = self.linear(x)
+        return x
+
+class Bottleneck(namedtuple('Block', ['in_channel', 'depth', 'stride'])):
+    """ A named tuple describing a ResNet block. """
+
+class bottleneck_IR(nn.Module):
+    def __init__(self, in_channel, depth, stride):
+        super(bottleneck_IR, self).__init__()
+        if in_channel == depth:
+            self.shortcut_layer = nn.MaxPool2d(1, stride)
+        else:
+            self.shortcut_layer = nn.Sequential(
+                nn.Conv2d(in_channel, depth, (1, 1), stride, bias=False),
+                nn.BatchNorm2d(depth)
+            )
+        self.res_layer = nn.Sequential(
+            nn.BatchNorm2d(in_channel),
+            nn.Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False), nn.PReLU(depth),
+            nn.Conv2d(depth, depth, (3, 3), stride, 1, bias=False), nn.BatchNorm2d(depth)
+        )
+
+    def forward(self, x):
+        shortcut = self.shortcut_layer(x)
+        res = self.res_layer(x)
+        return res + shortcut
+
+class SEModule(nn.Module):
+    def __init__(self, channels, reduction):
+        super(SEModule, self).__init__()
+        self.avg_pool = nn.AdaptiveAvgPool2d(1)
+        self.fc1 = nn.Conv2d(channels, channels // reduction, kernel_size=1, padding=0, bias=False)
+        self.relu = nn.ReLU(inplace=True)
+        self.fc2 = nn.Conv2d(channels // reduction, channels, kernel_size=1, padding=0, bias=False)
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x):
+        module_input = x
+        x = self.avg_pool(x)
+        x = self.fc1(x)
+        x = self.relu(x)
+        x = self.fc2(x)
+        x = self.sigmoid(x)
+        return module_input * x
+
+class bottleneck_IR_SE(nn.Module):
+    def __init__(self, in_channel, depth, stride):
+        super(bottleneck_IR_SE, self).__init__()
+        if in_channel == depth:
+            self.shortcut_layer = nn.MaxPool2d(1, stride)
+        else:
+            self.shortcut_layer = nn.Sequential(
+                nn.Conv2d(in_channel, depth, (1, 1), stride, bias=False),
+                nn.BatchNorm2d(depth)
+            )
+        self.res_layer = nn.Sequential(
+            nn.BatchNorm2d(in_channel),
+            nn.Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False),
+            nn.PReLU(depth),
+            nn.Conv2d(depth, depth, (3, 3), stride, 1, bias=False),
+            nn.BatchNorm2d(depth),
+            SEModule(depth, 16)
+        )
+
+    def forward(self, x):
+        shortcut = self.shortcut_layer(x)
+        res = self.res_layer(x)
+        return res + shortcut
+
+
+def get_block(in_channel, depth, num_units, stride=2):
+    return [Bottleneck(in_channel, depth, stride)] + [Bottleneck(depth, depth, 1) for i in range(num_units - 1)]
+
+def get_blocks(num_layers):
+    if num_layers == 50:
+        blocks = [
+            get_block(in_channel=64, depth=64, num_units=3),
+            get_block(in_channel=64, depth=128, num_units=4),
+            get_block(in_channel=128, depth=256, num_units=14),
+            get_block(in_channel=256, depth=512, num_units=3)
+        ]
+    elif num_layers == 100:
+        blocks = [
+            get_block(in_channel=64, depth=64, num_units=3),
+            get_block(in_channel=64, depth=128, num_units=13),
+            get_block(in_channel=128, depth=256, num_units=30),
+            get_block(in_channel=256, depth=512, num_units=3)
+        ]
+    elif num_layers == 152:
+        blocks = [
+            get_block(in_channel=64, depth=64, num_units=3),
+            get_block(in_channel=64, depth=128, num_units=8),
+            get_block(in_channel=128, depth=256, num_units=36),
+            get_block(in_channel=256, depth=512, num_units=3)
+        ]
+    else:
+        raise ValueError("Invalid number of layers: {}. Must be one of [50, 100, 152]".format(num_layers))
+    return blocks
+
+class Encoder4Editing(nn.Module):
+    def __init__(self, num_layers, mode='ir', stylegan_size=1024, out_res=64):
+        super(Encoder4Editing, self).__init__()
+        assert num_layers in [50, 100, 152], 'num_layers should be 50,100, or 152'
+        assert mode in ['ir', 'ir_se'], 'mode should be ir or ir_se'
+        blocks = get_blocks(num_layers)
+        if mode == 'ir':
+            unit_module = bottleneck_IR
+        elif mode == 'ir_se':
+            unit_module = bottleneck_IR_SE
+        self.out_res = out_res
+        self.input_layer = nn.Sequential(nn.Conv2d(3, 64, (3, 3), 1, 1, bias=False),
+                                      nn.BatchNorm2d(64),
+                                      nn.PReLU(64))
+        modules = []
+        for block in blocks:
+            for bottleneck in block:
+                modules.append(unit_module(bottleneck.in_channel,
+                                           bottleneck.depth,
+                                           bottleneck.stride))
+        self.body = nn.Sequential(*modules)
+
+        self.styles = nn.ModuleList()
+        log_size = int(math.log(stylegan_size, 2))
+        self.style_count = 2 * log_size - 2
+        self.coarse_ind = 3
+        self.middle_ind = 7
+
+        for i in range(self.style_count):
+            if i < self.coarse_ind:
+                style = GradualStyleBlock(512, 512, 16)
+            elif i < self.middle_ind:
+                style = GradualStyleBlock(512, 512, 32)
+            else:
+                style = GradualStyleBlock(512, 512, 64)
+            self.styles.append(style)
+
+        self.latlayer1 = nn.Conv2d(256, 512, kernel_size=1, stride=1, padding=0)
+        self.latlayer2 = nn.Conv2d(128, 512, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x):
+        x = self.input_layer(x)
+
+        modulelist = list(self.body._modules.values())
+        for i, l in enumerate(modulelist):
+            x = l(x)
+            if i == 2:
+                c0 = x
+            if i == 6:
+                c1 = x
+            elif i == 20:
+                c2 = x
+            elif i == 23:
+                c3 = x
+
+        # Infer main W and duplicate it
+        w0 = self.styles[0](c3)
+        w = w0.repeat(self.style_count, 1, 1).permute(1, 0, 2)
+
+        features = c3
+        for i in range(1, self.style_count):  # Infer additional deltas
+            if i == self.coarse_ind:
+                p2 = _upsample_add(c3, self.latlayer1(c2))  # FPN's middle features
+                features = p2
+            elif i == self.middle_ind:
+                p1 = _upsample_add(p2, self.latlayer2(c1))  # FPN's fine features
+                features = p1
+            delta_i = self.styles[i](features)
+            w[:, i] += delta_i
+
+        c = {   128: c0,
+                64: c1,
+                32: c2,
+                16: c3
+             }.get(self.out_res)
+        return w, c
+
+class EqualConv2d(nn.Module):
+    def __init__(
+            self, in_channel, out_channel, kernel_size, stride=1, padding=0, bias=True
+    ):
+        super().__init__()
+
+        self.weight = nn.Parameter(
+            torch.randn(out_channel, in_channel, kernel_size, kernel_size)
+        )
+        self.scale = 1 / math.sqrt(in_channel * kernel_size ** 2)
+
+        self.stride = stride
+        self.padding = padding
+
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_channel))
+
+        else:
+            self.bias = None
+
+    def forward(self, input):
+        out = F.conv2d(
+            input,
+            self.weight * self.scale,
+            bias=self.bias,
+            stride=self.stride,
+            padding=self.padding,
+        )
+
+        return out
+
+    def __repr__(self):
+        return (
+            f'{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]},'
+            f' {self.weight.shape[2]}, stride={self.stride}, padding={self.padding})'
+        )
+
+class ScaledLeakyReLU(nn.Module):
+    def __init__(self, negative_slope=0.2):
+        super().__init__()
+
+        self.negative_slope = negative_slope
+
+    def forward(self, input):
+        out = F.leaky_relu(input, negative_slope=self.negative_slope)
+
+        return out * math.sqrt(2)
+
+class HighResFeat(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        super(HighResFeat, self).__init__()
+
+        self.shared = EqualConv2d(in_channels, out_channels, kernel_size=3, padding=1, bias=True)
+
+        self.conv1 =  EqualConv2d(out_channels, 1, kernel_size=3, padding=1, bias=True)
+        self.conv2 =  EqualConv2d(out_channels, out_channels, kernel_size=3, padding=1, bias=True)
+        self.activation = ScaledLeakyReLU(0.2)
+
+        self.sigmoid = nn.Sigmoid()
+
+        self.skip = None
+        if in_channels != out_channels:
+            self.skip = EqualConv2d(in_channels, out_channels, kernel_size=1, padding=0, bias=False)
+    
+    def forward(self, x):
+
+        shared_feats = self.shared(x)
+        shared_feats = self.activation(shared_feats)
+
+        gate = self.conv1(shared_feats)
+        gate = self.sigmoid(gate)
+
+        addition = self.conv2(shared_feats)
+        addition = self.activation(addition)
+
+        if self.skip is not None:
+            x = self.skip(x)
+        return gate, addition+x
+
+class E4E_Inversion(nn.Module):
+    def __init__(self, resolution, num_layers = 50, mode='ir_se', out_res=64):
+        super(E4E_Inversion, self).__init__()
+        self.out_res = out_res
+        resolution = 1024
+        self.basic_encoder = Encoder4Editing(num_layers, mode, resolution, self.out_res)
+        self.latent_avg = None
+        # ckpt = torch.load(e4e_path, map_location='cpu')
+        # self.latent_avg = ckpt['latent_avg'].cuda()
+        # ckpt = {k[k.find(".")+1:]: v for k, v in ckpt['state_dict'].items() if "decoder" not in k}
+        # self.basic_encoder.load_state_dict(ckpt, strict=True)
+
+    def freeze_basic_encoder(self):
+        self.basic_encoder.eval()   #Basic Encoder always in eval mode.
+        #No backprop to basic Encoder
+        for param in self.basic_encoder.parameters():
+            param.requires_grad = False 
+
+    def forward(self, reals):
+        self.freeze_basic_encoder()
+        w, c = self.basic_encoder(reals)
+        w = w + self.latent_avg
+        highres_outs = {f"{self.out_res}x{self.out_res}": c} #{"gates": gates, "additions": additions}
+        return w, highres_outs

models/stylegan2.py

@@ -0,0 +1,965 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Network architectures from the paper
+"Analyzing and Improving the Image Quality of StyleGAN".
+Matches the original implementation of configs E-F by Karras et al. at
+https://github.com/NVlabs/stylegan2/blob/master/training/networks_stylegan2.py"""
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from .torch_utils import misc
+from .torch_utils.ops import conv2d_resample
+from .torch_utils.ops import upfirdn2d
+from .torch_utils.ops import bias_act
+from .torch_utils.ops import fma
+
+#----------------------------------------------------------------------------
+
+
+def normalize_2nd_moment(x, dim=1, eps=1e-8):
+    return x * (x.square().mean(dim=dim, keepdim=True) + eps).rsqrt()
+
+#----------------------------------------------------------------------------
+
+
+def modulated_conv2d(
+    x,                          # Input tensor of shape [batch_size, in_channels, in_height, in_width].
+    weight,                     # Weight tensor of shape [out_channels, in_channels, kernel_height, kernel_width].
+    styles,                     # Modulation coefficients of shape [batch_size, in_channels].
+    noise           = None,     # Optional noise tensor to add to the output activations.
+    up              = 1,        # Integer upsampling factor.
+    down            = 1,        # Integer downsampling factor.
+    padding         = 0,        # Padding with respect to the upsampled image.
+    resample_filter = None,     # Low-pass filter to apply when resampling activations. Must be prepared beforehand by calling upfirdn2d.setup_filter().
+    demodulate      = True,     # Apply weight demodulation?
+    flip_weight     = True,     # False = convolution, True = correlation (matches torch.nn.functional.conv2d).
+    fused_modconv   = True,     # Perform modulation, convolution, and demodulation as a single fused operation?
+    weigth_deltas   = None
+):
+    batch_size = x.shape[0]
+    out_channels, in_channels, kh, kw = weight.shape
+    misc.assert_shape(weight, [out_channels, in_channels, kh, kw]) # [OIkk]
+    misc.assert_shape(x, [batch_size, in_channels, None, None]) # [NIHW]
+    misc.assert_shape(styles, [batch_size, in_channels]) # [NI]
+
+    # Pre-normalize inputs to avoid FP16 overflow.
+    if x.dtype == torch.float16 and demodulate:
+        weight = weight * (1 / np.sqrt(in_channels * kh * kw) / weight.norm(float('inf'), dim=[1,2,3], keepdim=True)) # max_Ikk
+        styles = styles / styles.norm(float('inf'), dim=1, keepdim=True) # max_I
+
+    # Calculate per-sample weights and demodulation coefficients.
+    w = None
+    dcoefs = None
+    if demodulate or fused_modconv:
+        w = weight.unsqueeze(0) # [NOIkk]
+        #HyperStyle Addition for the Generator
+        if weigth_deltas is None:
+            w = w * styles.reshape(batch_size, 1, -1, 1, 1) # [NOIkk]
+        else:
+            w = w * (1 + weigth_deltas) * styles.reshape(batch_size, 1, -1, 1, 1)
+    if demodulate:
+        dcoefs = (w.square().sum(dim=[2,3,4]) + 1e-8).rsqrt() # [NO]
+    if demodulate and fused_modconv:
+        w = w * dcoefs.reshape(batch_size, -1, 1, 1, 1) # [NOIkk]
+
+    # Execute by scaling the activations before and after the convolution.
+    if not fused_modconv:
+        x = x * styles.to(x.dtype).reshape(batch_size, -1, 1, 1)
+        x = conv2d_resample.conv2d_resample(x=x, w=weight.to(x.dtype), f=resample_filter, up=up, down=down, padding=padding, flip_weight=flip_weight)
+        if demodulate and noise is not None:
+            x = fma.fma(x, dcoefs.to(x.dtype).reshape(batch_size, -1, 1, 1), noise.to(x.dtype))
+        elif demodulate:
+            x = x * dcoefs.to(x.dtype).reshape(batch_size, -1, 1, 1)
+        elif noise is not None:
+            x = x.add_(noise.to(x.dtype))
+        return x
+
+    # Execute as one fused op using grouped convolution.
+    with misc.suppress_tracer_warnings(): # this value will be treated as a constant
+        batch_size = int(batch_size)
+    misc.assert_shape(x, [batch_size, in_channels, None, None])
+    x = x.reshape(1, -1, *x.shape[2:])
+    w = w.reshape(-1, in_channels, kh, kw)
+    x = conv2d_resample.conv2d_resample(x=x, w=w.to(x.dtype), f=resample_filter, up=up, down=down, padding=padding, groups=batch_size, flip_weight=flip_weight)
+    x = x.reshape(batch_size, -1, *x.shape[2:])
+    if noise is not None:
+        x = x.add_(noise)
+    return x
+
+#----------------------------------------------------------------------------
+
+
+class FullyConnectedLayer(torch.nn.Module):
+    def __init__(self,
+        in_features,                # Number of input features.
+        out_features,               # Number of output features.
+        bias            = True,     # Apply additive bias before the activation function?
+        activation      = 'linear', # Activation function: 'relu', 'lrelu', etc.
+        lr_multiplier   = 1,        # Learning rate multiplier.
+        bias_init       = 0,        # Initial value for the additive bias.
+    ):
+        super().__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.activation = activation
+        self.weight = torch.nn.Parameter(torch.randn([out_features, in_features]) / lr_multiplier)
+        self.bias = torch.nn.Parameter(torch.full([out_features], np.float32(bias_init))) if bias else None
+        self.weight_gain = lr_multiplier / np.sqrt(in_features)
+        self.bias_gain = lr_multiplier
+
+    def forward(self, x):
+        w = self.weight.to(x.dtype) * self.weight_gain
+        b = self.bias
+        if b is not None:
+            b = b.to(x.dtype)
+            if self.bias_gain != 1:
+                b = b * self.bias_gain
+
+        if self.activation == 'linear' and b is not None:
+            x = torch.addmm(b.unsqueeze(0), x, w.t())
+        else:
+            x = x.matmul(w.t())
+            x = bias_act.bias_act(x, b, act=self.activation)
+        return x
+
+    def extra_repr(self):
+        return f'in_features={self.in_features:d}, out_features={self.out_features:d}, activation={self.activation:s}'
+
+#----------------------------------------------------------------------------
+
+
+class Conv2dLayer(torch.nn.Module):
+    def __init__(self,
+        in_channels,                    # Number of input channels.
+        out_channels,                   # Number of output channels.
+        kernel_size,                    # Width and height of the convolution kernel.
+        bias            = True,         # Apply additive bias before the activation function?
+        activation      = 'linear',     # Activation function: 'relu', 'lrelu', etc.
+        up              = 1,            # Integer upsampling factor.
+        down            = 1,            # Integer downsampling factor.
+        resample_filter = [1,3,3,1],    # Low-pass filter to apply when resampling activations.
+        conv_clamp      = None,         # Clamp the output to +-X, None = disable clamping.
+        channels_last   = False,        # Expect the input to have memory_format=channels_last?
+        trainable       = True,         # Update the weights of this layer during training?
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.activation = activation
+        self.up = up
+        self.down = down
+        self.conv_clamp = conv_clamp
+        self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
+        self.padding = kernel_size // 2
+        self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size ** 2))
+        self.act_gain = bias_act.activation_funcs[activation].def_gain
+
+        memory_format = torch.channels_last if channels_last else torch.contiguous_format
+        weight = torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(memory_format=memory_format)
+        bias = torch.zeros([out_channels]) if bias else None
+        if trainable:
+            self.weight = torch.nn.Parameter(weight)
+            self.bias = torch.nn.Parameter(bias) if bias is not None else None
+        else:
+            self.register_buffer('weight', weight)
+            if bias is not None:
+                self.register_buffer('bias', bias)
+            else:
+                self.bias = None
+
+    def forward(self, x, gain=1):
+        w = self.weight * self.weight_gain
+        b = self.bias.to(x.dtype) if self.bias is not None else None
+        flip_weight = (self.up == 1) # slightly faster
+        x = conv2d_resample.conv2d_resample(x=x, w=w.to(x.dtype), f=self.resample_filter, up=self.up, down=self.down, padding=self.padding, flip_weight=flip_weight)
+
+        act_gain = self.act_gain * gain
+        act_clamp = self.conv_clamp * gain if self.conv_clamp is not None else None
+        x = bias_act.bias_act(x, b, act=self.activation, gain=act_gain, clamp=act_clamp)
+        return x
+
+    def extra_repr(self):
+        return ' '.join([
+            f'in_channels={self.in_channels:d}, out_channels={self.out_channels:d}, activation={self.activation:s},',
+            f'up={self.up}, down={self.down}'])
+
+#----------------------------------------------------------------------------
+
+
+class MappingNetwork(torch.nn.Module):
+    def __init__(self,
+        z_dim,                      # Input latent (Z) dimensionality, 0 = no latent.
+        c_dim,                      # Conditioning label (C) dimensionality, 0 = no label.
+        w_dim,                      # Intermediate latent (W) dimensionality.
+        num_ws,                     # Number of intermediate latents to output, None = do not broadcast.
+        num_layers      = 8,        # Number of mapping layers.
+        embed_features  = None,     # Label embedding dimensionality, None = same as w_dim.
+        layer_features  = None,     # Number of intermediate features in the mapping layers, None = same as w_dim.
+        activation      = 'lrelu',  # Activation function: 'relu', 'lrelu', etc.
+        lr_multiplier   = 0.01,     # Learning rate multiplier for the mapping layers.
+        w_avg_beta      = 0.998,    # Decay for tracking the moving average of W during training, None = do not track.
+    ):
+        super().__init__()
+        self.z_dim = z_dim
+        self.c_dim = c_dim
+        self.w_dim = w_dim
+        self.num_ws = num_ws
+        self.num_layers = num_layers
+        self.w_avg_beta = w_avg_beta
+
+        if embed_features is None:
+            embed_features = w_dim
+        if c_dim == 0:
+            embed_features = 0
+        if layer_features is None:
+            layer_features = w_dim
+        features_list = [z_dim + embed_features] + [layer_features] * (num_layers - 1) + [w_dim]
+
+        if c_dim > 0:
+            self.embed = FullyConnectedLayer(c_dim, embed_features)
+        for idx in range(num_layers):
+            in_features = features_list[idx]
+            out_features = features_list[idx + 1]
+            layer = FullyConnectedLayer(in_features, out_features, activation=activation, lr_multiplier=lr_multiplier)
+            setattr(self, f'fc{idx}', layer)
+
+        if num_ws is not None and w_avg_beta is not None:
+            self.register_buffer('w_avg', torch.zeros([w_dim]))
+
+    def forward(self, z, c, truncation_psi=1, truncation_cutoff=None, update_emas=False, repeat_w = False):
+        # Embed, normalize, and concat inputs.
+        x = None
+        with torch.autograd.profiler.record_function('input'):
+            if self.z_dim > 0:
+                misc.assert_shape(z, [None, self.z_dim])
+                x = normalize_2nd_moment(z.to(torch.float32))
+            if self.c_dim > 0:
+                misc.assert_shape(c, [None, self.c_dim])
+                y = normalize_2nd_moment(self.embed(c.to(torch.float32)))
+                x = torch.cat([x, y], dim=1) if x is not None else y
+
+        # Main layers.
+        for idx in range(self.num_layers):
+            layer = getattr(self, f'fc{idx}')
+            x = layer(x)
+
+        # Update moving average of W.
+        if update_emas and self.w_avg_beta is not None:
+            with torch.autograd.profiler.record_function('update_w_avg'):
+                self.w_avg.copy_(x.detach().mean(dim=0).lerp(self.w_avg, self.w_avg_beta))
+
+        # Broadcast.
+        #if self.num_ws is not None:
+        if repeat_w:
+            with torch.autograd.profiler.record_function('broadcast'):
+                x = x.unsqueeze(1).repeat([1, self.num_ws, 1])
+
+        # Apply truncation.
+        if truncation_psi != 1:
+            with torch.autograd.profiler.record_function('truncate'):
+                assert self.w_avg_beta is not None
+                if self.num_ws is None or truncation_cutoff is None:
+                    x = self.w_avg.lerp(x, truncation_psi)
+                else:
+                    x[:, :truncation_cutoff] = self.w_avg.lerp(x[:, :truncation_cutoff], truncation_psi)
+        return x
+
+    def extra_repr(self):
+        return f'z_dim={self.z_dim:d}, c_dim={self.c_dim:d}, w_dim={self.w_dim:d}, num_ws={self.num_ws:d}'
+
+#----------------------------------------------------------------------------
+
+
+class SynthesisLayer(torch.nn.Module):
+    def __init__(self,
+        in_channels,                    # Number of input channels.
+        out_channels,                   # Number of output channels.
+        w_dim,                          # Intermediate latent (W) dimensionality.
+        resolution,                     # Resolution of this layer.
+        kernel_size     = 3,            # Convolution kernel size.
+        up              = 1,            # Integer upsampling factor.
+        use_noise       = True,         # Enable noise input?
+        activation      = 'lrelu',      # Activation function: 'relu', 'lrelu', etc.
+        resample_filter = [1,3,3,1],    # Low-pass filter to apply when resampling activations.
+        conv_clamp      = None,         # Clamp the output of convolution layers to +-X, None = disable clamping.
+        channels_last   = False,        # Use channels_last format for the weights?
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.w_dim = w_dim
+        self.resolution = resolution
+        self.up = up
+        self.use_noise = use_noise
+        self.activation = activation
+        self.conv_clamp = conv_clamp
+        self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
+        self.padding = kernel_size // 2
+        self.act_gain = bias_act.activation_funcs[activation].def_gain
+
+        self.affine = FullyConnectedLayer(w_dim, in_channels, bias_init=1)
+        memory_format = torch.channels_last if channels_last else torch.contiguous_format
+        self.weight = torch.nn.Parameter(torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(memory_format=memory_format))
+        if use_noise:
+            self.register_buffer('noise_const', torch.randn([resolution, resolution]))
+            self.noise_strength = torch.nn.Parameter(torch.zeros([]))
+        self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
+
+    def forward(self, x, w, noise_mode='random', n = None, weight_deltas = None,fused_modconv=True, gain=1):
+        assert noise_mode in ['random', 'const', 'none']
+        in_resolution = self.resolution // self.up
+        misc.assert_shape(x, [None, self.in_channels, in_resolution, in_resolution])
+        styles = self.affine(w)
+
+        noise = None
+        if self.use_noise and noise_mode == 'random':
+            noise = torch.randn([x.shape[0], 1, self.resolution, self.resolution], device=x.device) * self.noise_strength
+        if self.use_noise and noise_mode == 'const':
+            if n is not None:
+                noise = n * self.noise_strength
+            else:
+                noise = self.noise_const * self.noise_strength
+
+        flip_weight = (self.up == 1) # slightly faster
+        x = modulated_conv2d(x=x, weight=self.weight, styles=styles, noise=noise, up=self.up,
+            padding=self.padding, resample_filter=self.resample_filter, flip_weight=flip_weight, fused_modconv=fused_modconv, weigth_deltas=weight_deltas)
+
+        act_gain = self.act_gain * gain
+        act_clamp = self.conv_clamp * gain if self.conv_clamp is not None else None
+        x = bias_act.bias_act(x, self.bias.to(x.dtype), act=self.activation, gain=act_gain, clamp=act_clamp)
+        return x
+
+    def extra_repr(self):
+        return ' '.join([
+            f'in_channels={self.in_channels:d}, out_channels={self.out_channels:d}, w_dim={self.w_dim:d},',
+            f'resolution={self.resolution:d}, up={self.up}, activation={self.activation:s}'])
+
+#----------------------------------------------------------------------------
+
+
+class ToRGBLayer(torch.nn.Module):
+    def __init__(self, in_channels, out_channels, w_dim, kernel_size=1, conv_clamp=None, channels_last=False):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.w_dim = w_dim
+        self.conv_clamp = conv_clamp
+        self.affine = FullyConnectedLayer(w_dim, in_channels, bias_init=1)
+        memory_format = torch.channels_last if channels_last else torch.contiguous_format
+        self.weight = torch.nn.Parameter(torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(memory_format=memory_format))
+        self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
+        self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size ** 2))
+
+    def forward(self, x, w, fused_modconv=True):
+        styles = self.affine(w) * self.weight_gain
+        x = modulated_conv2d(x=x, weight=self.weight, styles=styles, demodulate=False, fused_modconv=fused_modconv)
+        x = bias_act.bias_act(x, self.bias.to(x.dtype), clamp=self.conv_clamp)
+        return x
+
+    def extra_repr(self):
+        return f'in_channels={self.in_channels:d}, out_channels={self.out_channels:d}, w_dim={self.w_dim:d}'
+
+#----------------------------------------------------------------------------
+class ResLayers(nn.Module):
+    def __init__(self, in_channels, out_channels, stride=1):
+        super().__init__()
+        
+        if (in_channels == out_channels) and stride==1:
+            self.shortcut_layer = nn.Identity()
+        else:
+            self.shortcut_layer = nn.Sequential(
+                nn.Conv2d(in_channels, out_channels, (1, 1), stride, bias=False))
+        
+        self.res_layer = nn.Sequential(
+            nn.Conv2d(in_channels, out_channels, (3, 3), (1, 1), 1, bias=True), nn.LeakyReLU(0.2),
+            nn.Conv2d(out_channels, out_channels, (3, 3), stride, 1, bias=True) )
+
+    def forward(self, x):
+        shortcut = self.shortcut_layer(x)
+        res = self.res_layer(x)
+        return res + shortcut
+
+class FeatureEdit(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        super().__init__()
+        self.convs = nn.ModuleList()
+        iter_num = in_channels // out_channels
+        i_c = in_channels
+        for i in range(iter_num-1):
+            out_c = i_c - out_channels
+            self.convs.append( ResLayers(i_c,out_c,1) )
+            i_c = out_c
+    def forward(self, diff):
+        for block in self.convs:
+            diff = block(diff)
+        return diff
+
+class FeatureAlignment(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        super().__init__()
+        t_channel = 512
+        self.first_layer = nn.Conv2d(in_channels, t_channel, kernel_size=1, padding=0, bias=True)
+
+        self.conv1 =  nn.Sequential(*[ResLayers(t_channel,t_channel,1)])
+        self.conv2 =  nn.Sequential(*[ResLayers(t_channel,t_channel,2), ResLayers(t_channel,t_channel,1)])
+        self.conv3 =  nn.Sequential(*[ResLayers(t_channel,t_channel,2), ResLayers(t_channel,t_channel,1)])
+
+        self.dconv1 = nn.Sequential(*[ResLayers(t_channel,t_channel,1), ResLayers(t_channel,t_channel,1)])
+        self.dconv2 = nn.Sequential(*[ResLayers(t_channel,t_channel,1), ResLayers(t_channel,t_channel,1)])
+        self.dconv3 = nn.Sequential(*[ResLayers(t_channel,t_channel,1), ResLayers(t_channel,t_channel,1)])
+
+        self.out_layer = nn.Conv2d(t_channel, out_channels, kernel_size=1, padding=0, bias=True)
+
+    def forward(self, encoder_feats, generator_feats):
+
+        x = torch.cat((encoder_feats,generator_feats), dim=1)
+        x = self.first_layer(x)
+
+        f1 = self.conv1(x)
+        f2 = self.conv2(f1)
+        f3 = self.conv3(f2)
+        shape = f3.shape[-1]
+        df1 = F.interpolate(f3, size=(shape*2,shape*2) , mode='bilinear', align_corners=True)
+        df2 = self.dconv1(df1 + f2)
+        df2 =  F.interpolate(df2, size=(shape*4,shape*4) , mode='bilinear', align_corners=True)
+        df3 = self.dconv2(df2 + f1)
+
+        aligned_feats = self.out_layer(df3)
+
+        return aligned_feats
+
+class FeatureExtraction(nn.Module):
+    def __init__(self,in_channels, out_channels ):
+        super().__init__() 
+        t_channel = 512
+        self.first_layer = nn.Conv2d(in_channels, t_channel, kernel_size=1, padding=0, bias=True)
+        self.convs = nn.Sequential(*[ResLayers(t_channel,t_channel,1), ResLayers(t_channel,out_channels,1), ResLayers(out_channels,out_channels,1) ])
+        #self.out_layer = nn.Conv2d(t_channel, out_channels, kernel_size=1, padding=0, bias=False)
+
+    def forward(self, aligned_feats):
+        #x = aligned_feats - generator_feats
+        y = self.first_layer(aligned_feats)
+        y = self.convs(y)
+        #deltaF = self.out_layer(x)
+        return y
+
+class GateNetwork(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        super().__init__()
+        t_channel = 256
+        self.down1 = nn.Conv2d(in_channels, t_channel, kernel_size=3, padding=1, bias=True)
+        self.down2 = nn.Conv2d(in_channels, t_channel, kernel_size=3, padding=1, bias=True)
+        self.sigmoid = nn.Sigmoid()
+        self.convs = nn.Sequential(*[ResLayers(in_channels,in_channels,1), ResLayers(in_channels,out_channels,1), ResLayers(out_channels,out_channels,1) ])
+        self.convs2 = nn.Sequential(*[ResLayers(in_channels,in_channels,1), ResLayers(in_channels,out_channels,1), ResLayers(out_channels,1,1) ])
+
+
+    def forward(self, generator_feats, y):
+        generator_feats = self.down1(generator_feats)
+        y = self.down2(y)
+        x = torch.cat((generator_feats, y), dim=1)
+        deltaF = self.convs(x)
+        gate = self.convs2(x)
+        gate = self.sigmoid(gate)
+        return deltaF, gate
+
+g_e_concat_shape={64: 640, 32:768}
+e_shape = {64: 128, 32:256}
+
+class SynthesisBlock(torch.nn.Module):
+    def __init__(self,
+        in_channels,                            # Number of input channels, 0 = first block.
+        out_channels,                           # Number of output channels.
+        w_dim,                                  # Intermediate latent (W) dimensionality.
+        resolution,                             # Resolution of this block.
+        img_channels,                           # Number of output color channels.
+        is_last,                                # Is this the last block?
+        architecture            = 'skip',       # Architecture: 'orig', 'skip', 'resnet'.
+        resample_filter         = [1,3,3,1],    # Low-pass filter to apply when resampling activations.
+        conv_clamp              = 256,          # Clamp the output of convolution layers to +-X, None = disable clamping.
+        use_fp16                = False,        # Use FP16 for this block?
+        fp16_channels_last      = False,        # Use channels-last memory format with FP16?
+        fused_modconv_default   = True,         # Default value of fused_modconv. 'inference_only' = True for inference, False for training.
+        embed_res               = 64,           # Which resolution we embed the images
+        **layer_kwargs,                         # Arguments for SynthesisLayer.
+    ):
+        assert architecture in ['orig', 'skip', 'resnet']
+        super().__init__()
+        self.in_channels = in_channels
+        self.w_dim = w_dim
+        self.resolution = resolution
+        self.img_channels = img_channels
+        self.is_last = is_last
+        self.architecture = architecture
+        self.use_fp16 = use_fp16
+        self.channels_last = (use_fp16 and fp16_channels_last)
+        self.fused_modconv_default = fused_modconv_default
+        self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
+        self.num_conv = 0
+        self.num_torgb = 0
+
+        if in_channels == 0:
+            self.const = torch.nn.Parameter(torch.randn([out_channels, resolution, resolution]))
+
+        if in_channels != 0:
+            self.conv0 = SynthesisLayer(in_channels, out_channels, w_dim=w_dim, resolution=resolution, up=2,
+                resample_filter=resample_filter, conv_clamp=conv_clamp, channels_last=self.channels_last, **layer_kwargs)
+            self.num_conv += 1
+
+        self.conv1 = SynthesisLayer(out_channels, out_channels, w_dim=w_dim, resolution=resolution,
+            conv_clamp=conv_clamp, channels_last=self.channels_last, **layer_kwargs)
+        self.num_conv += 1
+
+        if is_last or architecture == 'skip':
+            self.torgb = ToRGBLayer(out_channels, img_channels, w_dim=w_dim,
+                conv_clamp=conv_clamp, channels_last=self.channels_last)
+            self.num_torgb += 1
+
+        if in_channels != 0 and architecture == 'resnet':
+            self.skip = Conv2dLayer(in_channels, out_channels, kernel_size=1, bias=False, up=2,
+                resample_filter=resample_filter, channels_last=self.channels_last)
+        if resolution == embed_res:
+            in_c = g_e_concat_shape.get(embed_res)
+            #self.modify_feature_edit = FeatureEdit(in_channels=512, out_channels=e_shape.get(embed_res))
+            self.modify_feature_alignment = FeatureAlignment(in_channels=in_c, out_channels=512)
+            self.modify_feature_extraction = FeatureExtraction(in_channels=512, out_channels=512)
+            self.modify_feature_gates = GateNetwork(in_channels=512, out_channels=512)
+        self.embed_res = embed_res
+
+    def forward(self, x, img, ws, conditions=None, noise=None, weight_deltas = None, highres_outs=None, return_f = False,
+            force_fp32=False, fused_modconv=None, update_emas=False, **layer_kwargs):
+        _ = update_emas # unused
+        misc.assert_shape(ws, [None, self.num_conv + self.num_torgb, self.w_dim])
+        w_iter = iter(ws.unbind(dim=1))
+        if ws.device.type != 'cuda':
+            force_fp32 = True
+        dtype = torch.float16 if self.use_fp16 and not force_fp32 else torch.float32
+        memory_format = torch.channels_last if self.channels_last and not force_fp32 else torch.contiguous_format
+        if fused_modconv is None:
+            fused_modconv = self.fused_modconv_default
+        if fused_modconv == 'inference_only':
+            fused_modconv = (not self.training)
+
+        # Input.
+        if self.in_channels == 0:
+            x = self.const.to(dtype=dtype, memory_format=memory_format)
+            x = x.unsqueeze(0).repeat([ws.shape[0], 1, 1, 1])
+        else:
+            misc.assert_shape(x, [None, self.in_channels, self.resolution // 2, self.resolution // 2])
+            x = x.to(dtype=dtype, memory_format=memory_format)
+        gouts = {}
+        # Main layers.
+        if self.in_channels == 0:
+            x = self.conv1(x, next(w_iter), fused_modconv=fused_modconv, n=noise[0], weight_deltas=weight_deltas[0], **layer_kwargs)
+        elif self.architecture == 'resnet':
+            y = self.skip(x, gain=np.sqrt(0.5))
+            x = self.conv0(x, next(w_iter), fused_modconv=fused_modconv, **layer_kwargs)
+            x = self.conv1(x, next(w_iter), fused_modconv=fused_modconv, gain=np.sqrt(0.5), **layer_kwargs)
+            x = y.add_(x)
+        else:
+            x = self.conv0(x, next(w_iter), fused_modconv=fused_modconv, n=noise[0], weight_deltas=weight_deltas[0], **layer_kwargs)
+            #HFGI Generator Modification
+            if x.shape[-1] == 64 and conditions is not None:
+                x = x*(1+conditions[0]) + conditions[1]
+            if x.shape[-1] == self.embed_res and return_f:
+                    return x, None, None
+            #HighResFeat Generator Modification
+            if x.shape[-1] == self.embed_res and highres_outs is not None: 
+                #feature_edit = self.modify_feature_edit(x - highres_outs['inversion'])
+                #high_res = highres_outs[f'{self.embed_res}x{self.embed_res}'] + feature_edit
+                aligned_feats = self.modify_feature_alignment(highres_outs[f'{self.embed_res}x{self.embed_res}'], highres_outs['inversion'])
+                aligned_feats = self.modify_feature_extraction(aligned_feats)
+                #x = self.modify_feature_gates(x, deltaF)
+                deltaF, gate = self.modify_feature_gates(x, aligned_feats)
+                x = (x * (1-gate) ) + ( (x + deltaF) * gate )
+                gouts['gates'] = gate
+                gouts['additions'] = deltaF
+                gouts['aligned_feats'] = aligned_feats
+
+                #gouts['aligned_loss'] =F.mse_loss(aligned_feats, x, reduction='mean')
+            x = self.conv1(x, next(w_iter), fused_modconv=fused_modconv,n=noise[1], weight_deltas=weight_deltas[1], **layer_kwargs)
+
+        # ToRGB.
+        if img is not None:
+            misc.assert_shape(img, [None, self.img_channels, self.resolution // 2, self.resolution // 2])
+            img = upfirdn2d.upsample2d(img, self.resample_filter)
+        if self.is_last or self.architecture == 'skip':
+            y = self.torgb(x, next(w_iter), fused_modconv=fused_modconv)
+            y = y.to(dtype=torch.float32, memory_format=torch.contiguous_format)
+            img = img.add_(y) if img is not None else y
+
+        assert x.dtype == dtype
+        assert img is None or img.dtype == torch.float32
+        return x, img, gouts
+
+    def extra_repr(self):
+        return f'resolution={self.resolution:d}, architecture={self.architecture:s}'
+
+#----------------------------------------------------------------------------
+
+
+class SynthesisNetwork(torch.nn.Module):
+    def __init__(self,
+        w_dim,                      # Intermediate latent (W) dimensionality.
+        img_resolution,             # Output image resolution.
+        img_channels,               # Number of color channels.
+        channel_base    = 32768,    # Overall multiplier for the number of channels.
+        channel_max     = 512,      # Maximum number of channels in any layer.
+        num_fp16_res    = 4,        # Use FP16 for the N highest resolutions.
+        **block_kwargs,             # Arguments for SynthesisBlock.
+    ):
+        assert img_resolution >= 4 and img_resolution & (img_resolution - 1) == 0
+        super().__init__()
+        self.w_dim = w_dim
+        self.img_resolution = img_resolution
+        self.img_resolution_log2 = int(np.log2(img_resolution))
+        self.img_channels = img_channels
+        self.num_fp16_res = num_fp16_res
+        self.block_resolutions = [2 ** i for i in range(2, self.img_resolution_log2 + 1)]
+        channels_dict = {res: min(channel_base // res, channel_max) for res in self.block_resolutions}
+        fp16_resolution = max(2 ** (self.img_resolution_log2 + 1 - num_fp16_res), 8)
+
+        self.num_ws = 0
+        for res in self.block_resolutions:
+            in_channels = channels_dict[res // 2] if res > 4 else 0
+            out_channels = channels_dict[res]
+            use_fp16 = (res >= fp16_resolution)
+            is_last = (res == self.img_resolution)
+            block = SynthesisBlock(in_channels, out_channels, w_dim=w_dim, resolution=res,
+                img_channels=img_channels, is_last=is_last, use_fp16=use_fp16, **block_kwargs)
+            self.num_ws += block.num_conv
+            if is_last:
+                self.num_ws += block.num_torgb
+            setattr(self, f'b{res}', block)
+
+    def forward(self, ws, conditions=None, noise=None, weight_deltas=None, highres_outs=None, return_f = False, **block_kwargs):
+        block_ws = []
+        with torch.autograd.profiler.record_function('split_ws'):
+            misc.assert_shape(ws, [None, self.num_ws, self.w_dim])
+            ws = ws.to(torch.float32)
+            w_idx = 0
+            for res in self.block_resolutions:
+                block = getattr(self, f'b{res}')
+                block_ws.append(ws.narrow(1, w_idx, block.num_conv + block.num_torgb))
+                w_idx += block.num_conv
+
+        x = img = None
+        conv_idx = 0
+        gouts = {}
+        for res, cur_ws in zip(self.block_resolutions, block_ws):
+            block = getattr(self, f'b{res}')
+            if noise is not None:
+                noise_input = noise[conv_idx: conv_idx + block.num_conv]
+            else:
+                noise_input = [None] * block.num_conv
+            if weight_deltas is not None:
+                delta_input = weight_deltas[conv_idx: conv_idx + block.num_conv]
+            else:
+                delta_input = [None] * block.num_conv
+            x, img, gouts_per_res = block(x, img, cur_ws, conditions, noise_input, delta_input, highres_outs, return_f, **block_kwargs)
+            if return_f and img is None:
+                return x, None
+            if gouts_per_res: 
+                gouts.update(gouts_per_res)
+                
+            conv_idx += block.num_conv
+        return img, gouts
+
+    def extra_repr(self):
+        return ' '.join([
+            f'w_dim={self.w_dim:d}, num_ws={self.num_ws:d},',
+            f'img_resolution={self.img_resolution:d}, img_channels={self.img_channels:d},',
+            f'num_fp16_res={self.num_fp16_res:d}'])
+
+#----------------------------------------------------------------------------
+
+
+class Generator(torch.nn.Module):
+    def __init__(self,
+        z_dim,                      # Input latent (Z) dimensionality.
+        c_dim,                      # Conditioning label (C) dimensionality.
+        w_dim,                      # Intermediate latent (W) dimensionality.
+        resolution,                 # Output resolution.
+        img_channels,               # Number of output color channels.
+        mapping_kwargs      = {},   # Arguments for MappingNetwork.
+        **synthesis_kwargs,         # Arguments for SynthesisNetwork.
+    ):
+        super().__init__()
+        self.z_dim = z_dim
+        self.c_dim = c_dim
+        self.w_dim = w_dim
+        self.resolution = resolution
+        self.img_channels = img_channels
+        self.synthesis = SynthesisNetwork(w_dim=w_dim, img_resolution=resolution, img_channels=img_channels, **synthesis_kwargs)
+        self.num_ws = self.synthesis.num_ws
+        self.mapping = MappingNetwork(z_dim=z_dim, c_dim=c_dim, w_dim=w_dim, num_ws=self.num_ws, **mapping_kwargs)
+
+        # self.freeze_non_trainable_layers()
+   
+    def forward(self, lat, c, truncation_psi=1, truncation_cutoff=None, update_emas=False, mode='synthesis', return_f = False, **synthesis_kwargs):
+        # self.freeze_non_trainable_layers()
+        if mode == 'mapping':
+            ws = self.mapping(lat, c, truncation_psi=truncation_psi, truncation_cutoff=truncation_cutoff, update_emas=update_emas)
+            return ws
+        if mode == 'synthesis':
+            img = self.synthesis(lat, highres_outs = c, return_f=return_f, update_emas=False, **synthesis_kwargs)
+            return img
+
+    # def freeze_non_trainable_layers(self):
+    #     for param in self.mapping.parameters():
+    #         param.requires_grad = False
+    #     for name, param in self.synthesis.named_parameters():
+    #         if 'modify' not in name:
+    #             param.requires_grad = False
+
+#----------------------------------------------------------------------------
+
+
+class DiscriminatorBlock(torch.nn.Module):
+    def __init__(self,
+        in_channels,                        # Number of input channels, 0 = first block.
+        tmp_channels,                       # Number of intermediate channels.
+        out_channels,                       # Number of output channels.
+        resolution,                         # Resolution of this block.
+        img_channels,                       # Number of input color channels.
+        first_layer_idx,                    # Index of the first layer.
+        architecture        = 'resnet',     # Architecture: 'orig', 'skip', 'resnet'.
+        activation          = 'lrelu',      # Activation function: 'relu', 'lrelu', etc.
+        resample_filter     = [1,3,3,1],    # Low-pass filter to apply when resampling activations.
+        conv_clamp          = None,         # Clamp the output of convolution layers to +-X, None = disable clamping.
+        use_fp16            = False,        # Use FP16 for this block?
+        fp16_channels_last  = False,        # Use channels-last memory format with FP16?
+        freeze_layers       = 0,            # Freeze-D: Number of layers to freeze.
+    ):
+        assert in_channels in [0, tmp_channels]
+        assert architecture in ['orig', 'skip', 'resnet']
+        super().__init__()
+        self.in_channels = in_channels
+        self.resolution = resolution
+        self.img_channels = img_channels
+        self.first_layer_idx = first_layer_idx
+        self.architecture = architecture
+        self.use_fp16 = use_fp16
+        self.channels_last = (use_fp16 and fp16_channels_last)
+        self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
+
+        self.num_layers = 0
+        def trainable_gen():
+            while True:
+                layer_idx = self.first_layer_idx + self.num_layers
+                trainable = (layer_idx >= freeze_layers)
+                self.num_layers += 1
+                yield trainable
+        trainable_iter = trainable_gen()
+
+        if in_channels == 0 or architecture == 'skip':
+            self.fromrgb = Conv2dLayer(img_channels, tmp_channels, kernel_size=1, activation=activation,
+                trainable=next(trainable_iter), conv_clamp=conv_clamp, channels_last=self.channels_last)
+
+        self.conv0 = Conv2dLayer(tmp_channels, tmp_channels, kernel_size=3, activation=activation,
+            trainable=next(trainable_iter), conv_clamp=conv_clamp, channels_last=self.channels_last)
+
+        self.conv1 = Conv2dLayer(tmp_channels, out_channels, kernel_size=3, activation=activation, down=2,
+            trainable=next(trainable_iter), resample_filter=resample_filter, conv_clamp=conv_clamp, channels_last=self.channels_last)
+
+        if architecture == 'resnet':
+            self.skip = Conv2dLayer(tmp_channels, out_channels, kernel_size=1, bias=False, down=2,
+                trainable=next(trainable_iter), resample_filter=resample_filter, channels_last=self.channels_last)
+
+    def forward(self, x, img, force_fp32=False):
+        if (x if x is not None else img).device.type != 'cuda':
+            force_fp32 = True
+        dtype = torch.float16 if self.use_fp16 and not force_fp32 else torch.float32
+        memory_format = torch.channels_last if self.channels_last and not force_fp32 else torch.contiguous_format
+
+        # Input.
+        if x is not None:
+            misc.assert_shape(x, [None, self.in_channels, self.resolution, self.resolution])
+            x = x.to(dtype=dtype, memory_format=memory_format)
+
+        # FromRGB.
+        if self.in_channels == 0 or self.architecture == 'skip':
+            misc.assert_shape(img, [None, self.img_channels, self.resolution, self.resolution])
+            img = img.to(dtype=dtype, memory_format=memory_format)
+            y = self.fromrgb(img)
+            x = x + y if x is not None else y
+            img = upfirdn2d.downsample2d(img, self.resample_filter) if self.architecture == 'skip' else None
+
+        # Main layers.
+        if self.architecture == 'resnet':
+            y = self.skip(x, gain=np.sqrt(0.5))
+            x = self.conv0(x)
+            x = self.conv1(x, gain=np.sqrt(0.5))
+            x = y.add_(x)
+        else:
+            x = self.conv0(x)
+            x = self.conv1(x)
+
+        assert x.dtype == dtype
+        return x, img
+
+    def extra_repr(self):
+        return f'resolution={self.resolution:d}, architecture={self.architecture:s}'
+
+#----------------------------------------------------------------------------
+
+
+class MinibatchStdLayer(torch.nn.Module):
+    def __init__(self, group_size, num_channels=1):
+        super().__init__()
+        self.group_size = group_size
+        self.num_channels = num_channels
+
+    def forward(self, x):
+        N, C, H, W = x.shape
+        with misc.suppress_tracer_warnings(): # as_tensor results are registered as constants
+            G = torch.min(torch.as_tensor(self.group_size), torch.as_tensor(N)) if self.group_size is not None else N
+        F = self.num_channels
+        c = C // F
+
+        y = x.reshape(G, -1, F, c, H, W)    # [GnFcHW] Split minibatch N into n groups of size G, and channels C into F groups of size c.
+        y = y - y.mean(dim=0)               # [GnFcHW] Subtract mean over group.
+        y = y.square().mean(dim=0)          # [nFcHW]  Calc variance over group.
+        y = (y + 1e-8).sqrt()               # [nFcHW]  Calc stddev over group.
+        y = y.mean(dim=[2,3,4])             # [nF]     Take average over channels and pixels.
+        y = y.reshape(-1, F, 1, 1)          # [nF11]   Add missing dimensions.
+        y = y.repeat(G, 1, H, W)            # [NFHW]   Replicate over group and pixels.
+        x = torch.cat([x, y], dim=1)        # [NCHW]   Append to input as new channels.
+        return x
+
+    def extra_repr(self):
+        return f'group_size={self.group_size}, num_channels={self.num_channels:d}'
+
+#----------------------------------------------------------------------------
+
+
+class DiscriminatorEpilogue(torch.nn.Module):
+    def __init__(self,
+        in_channels,                    # Number of input channels.
+        cmap_dim,                       # Dimensionality of mapped conditioning label, 0 = no label.
+        resolution,                     # Resolution of this block.
+        img_channels,                   # Number of input color channels.
+        architecture        = 'resnet', # Architecture: 'orig', 'skip', 'resnet'.
+        mbstd_group_size    = 4,        # Group size for the minibatch standard deviation layer, None = entire minibatch.
+        mbstd_num_channels  = 1,        # Number of features for the minibatch standard deviation layer, 0 = disable.
+        activation          = 'lrelu',  # Activation function: 'relu', 'lrelu', etc.
+        conv_clamp          = None,     # Clamp the output of convolution layers to +-X, None = disable clamping.
+    ):
+        assert architecture in ['orig', 'skip', 'resnet']
+        super().__init__()
+        self.in_channels = in_channels
+        self.cmap_dim = cmap_dim
+        self.resolution = resolution
+        self.img_channels = img_channels
+        self.architecture = architecture
+
+        if architecture == 'skip':
+            self.fromrgb = Conv2dLayer(img_channels, in_channels, kernel_size=1, activation=activation)
+        self.mbstd = MinibatchStdLayer(group_size=mbstd_group_size, num_channels=mbstd_num_channels) if mbstd_num_channels > 0 else None
+        self.conv = Conv2dLayer(in_channels + mbstd_num_channels, in_channels, kernel_size=3, activation=activation, conv_clamp=conv_clamp)
+        self.fc = FullyConnectedLayer(in_channels * (resolution ** 2), in_channels, activation=activation)
+        self.out = FullyConnectedLayer(in_channels, 1 if cmap_dim == 0 else cmap_dim)
+
+    def forward(self, x, img, cmap, force_fp32=False):
+        misc.assert_shape(x, [None, self.in_channels, self.resolution, self.resolution]) # [NCHW]
+        _ = force_fp32 # unused
+        dtype = torch.float32
+        memory_format = torch.contiguous_format
+
+        # FromRGB.
+        x = x.to(dtype=dtype, memory_format=memory_format)
+        if self.architecture == 'skip':
+            misc.assert_shape(img, [None, self.img_channels, self.resolution, self.resolution])
+            img = img.to(dtype=dtype, memory_format=memory_format)
+            x = x + self.fromrgb(img)
+
+        # Main layers.
+        if self.mbstd is not None:
+            x = self.mbstd(x)
+        x = self.conv(x)
+        x = self.fc(x.flatten(1))
+        x = self.out(x)
+
+        # Conditioning.
+        if self.cmap_dim > 0:
+            misc.assert_shape(cmap, [None, self.cmap_dim])
+            x = (x * cmap).sum(dim=1, keepdim=True) * (1 / np.sqrt(self.cmap_dim))
+
+        assert x.dtype == dtype
+        return x
+
+    def extra_repr(self):
+        return f'resolution={self.resolution:d}, architecture={self.architecture:s}'
+
+#----------------------------------------------------------------------------
+
+
+class Discriminator(torch.nn.Module):
+    def __init__(self,
+        c_dim,                          # Conditioning label (C) dimensionality.
+        img_resolution,                 # Input resolution.
+        img_channels,                   # Number of input color channels.
+        architecture        = 'resnet', # Architecture: 'orig', 'skip', 'resnet'.
+        channel_base        = 32768,    # Overall multiplier for the number of channels.
+        channel_max         = 512,      # Maximum number of channels in any layer.
+        num_fp16_res        = 4,        # Use FP16 for the N highest resolutions.
+        conv_clamp          = 256,      # Clamp the output of convolution layers to +-X, None = disable clamping.
+        cmap_dim            = None,     # Dimensionality of mapped conditioning label, None = default.
+        block_kwargs        = {},       # Arguments for DiscriminatorBlock.
+        mapping_kwargs      = {},       # Arguments for MappingNetwork.
+        epilogue_kwargs     = {},       # Arguments for DiscriminatorEpilogue.
+    ):
+        super().__init__()
+        self.c_dim = c_dim
+        self.img_resolution = img_resolution
+        self.img_resolution_log2 = int(np.log2(img_resolution))
+        self.img_channels = img_channels
+        self.block_resolutions = [2 ** i for i in range(self.img_resolution_log2, 2, -1)]
+        channels_dict = {res: min(channel_base // res, channel_max) for res in self.block_resolutions + [4]}
+        fp16_resolution = max(2 ** (self.img_resolution_log2 + 1 - num_fp16_res), 8)
+
+        if cmap_dim is None:
+            cmap_dim = channels_dict[4]
+        if c_dim == 0:
+            cmap_dim = 0
+
+        common_kwargs = dict(img_channels=img_channels, architecture=architecture, conv_clamp=conv_clamp)
+        cur_layer_idx = 0
+        for res in self.block_resolutions:
+            in_channels = channels_dict[res] if res < img_resolution else 0
+            tmp_channels = channels_dict[res]
+            out_channels = channels_dict[res // 2]
+            use_fp16 = (res >= fp16_resolution)
+            block = DiscriminatorBlock(in_channels, tmp_channels, out_channels, resolution=res,
+                first_layer_idx=cur_layer_idx, use_fp16=use_fp16, **block_kwargs, **common_kwargs)
+            setattr(self, f'b{res}', block)
+            cur_layer_idx += block.num_layers
+        if c_dim > 0:
+            self.mapping = MappingNetwork(z_dim=0, c_dim=c_dim, w_dim=cmap_dim, num_ws=None, w_avg_beta=None, **mapping_kwargs)
+        self.b4 = DiscriminatorEpilogue(channels_dict[4], cmap_dim=cmap_dim, resolution=4, **epilogue_kwargs, **common_kwargs)
+
+    def forward(self, img, c, update_emas=False, **block_kwargs):
+        _ = update_emas # unused
+        x = None
+        for res in self.block_resolutions:
+            block = getattr(self, f'b{res}')
+            x, img = block(x, img, **block_kwargs)
+
+        cmap = None
+        if self.c_dim > 0:
+            cmap = self.mapping(None, c)
+        x = self.b4(x, img, cmap)
+        return x
+
+    def extra_repr(self):
+        return f'c_dim={self.c_dim:d}, img_resolution={self.img_resolution:d}, img_channels={self.img_channels:d}'
+
+#----------------------------------------------------------------------------

models/styleres.py

@@ -0,0 +1,75 @@
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from models.e4e import E4E_Inversion
+from models.stylegan2 import Generator
+from editings.editor import Editor
+from options import Settings
+
+class StyleRes(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.encoder = E4E_Inversion(resolution=256, num_layers = 50, mode='ir_se', out_res=64)
+        self.generator = Generator(z_dim=512, w_dim=512, c_dim=0, resolution=1024, img_channels=3, 
+                        fused_modconv_default='inference_only', embed_res=64)
+        # Set Generator arguments for eval mode
+        self.G_kwargs_val = {'noise_mode':'const', 'force_fp32':True}
+        self.device = Settings.device
+        self.editor = Editor()
+
+    def load_ckpt(self, ckpt_path):
+        ckpt = torch.load(ckpt_path, map_location='cpu')
+        self.encoder.basic_encoder.load_state_dict(ckpt['e4e'], strict=True)
+        self.encoder.latent_avg = ckpt['latent_avg']
+        self.generator.load_state_dict(ckpt['generator_smooth'], strict=True)
+        print("Model succesfully loaded")
+
+    def send_to_device(self):
+        self.encoder.to(self.device)
+        self.generator.to(self.device)
+        if self.device != 'cpu':
+            self.encoder.latent_avg = self.encoder.latent_avg.cuda()
+
+    """
+        Inputs: Input images and edit configs
+        Returns: Edited images together with the randomly generated image when the edit is interpolation.
+    """
+    def edit_images(self, image, cfg):
+        image = image.to(self.device)
+        with torch.no_grad():
+            latents, skips = self.encoder(image)
+
+        # GradCtrl requires gradients, others do not
+        latents_edited = self.editor.edit(latents, cfg) 
+
+        with torch.no_grad():
+            # Get F space features F_orig, for the original image
+            skips['inversion'], _ = self.generator(latents, skips, return_f = True, **self.G_kwargs_val)
+            # Transform F_orig to incoming image
+            images, _ = self.generator(latents_edited, skips,  **self.G_kwargs_val)
+            
+        return images
+
+    # def edit_demo_image(self, image, edit, factor):
+    #     from utils import AttrDict
+    #     cfg = AttrDict()
+    #     edit = edit.lower()
+    #     if edit in ['pose', 'age', 'smile']:
+    #         cfg.method = 'interfacegan'
+    #         cfg.edit = edit
+    #         cfg.strength = factor
+    #     image = image.to(self.device)
+    #     with torch.no_grad():
+    #         latents, skips = self.encoder(image)
+    #         latents_edited = self.editor.edit(latents, cfg) 
+    #     with torch.no_grad():
+    #         # Get F space features F_orig, for the original image
+    #         skips['inversion'], _ = self.generator(latents, skips, return_f = True, **self.G_kwargs_val)
+    #         # Transform F_orig to incoming image
+    #         images, _ = self.generator(latents_edited, skips,  **self.G_kwargs_val)
+            
+    #     return images
+
+
+

models/torch_utils/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+# empty

models/torch_utils/custom_ops.py

@@ -0,0 +1,157 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import glob
+import hashlib
+import importlib
+import os
+import re
+import shutil
+import uuid
+
+import torch
+import torch.utils.cpp_extension
+from torch.utils.file_baton import FileBaton
+
+#----------------------------------------------------------------------------
+# Global options.
+
+verbosity = 'brief' # Verbosity level: 'none', 'brief', 'full'
+
+#----------------------------------------------------------------------------
+# Internal helper funcs.
+
+def _find_compiler_bindir():
+    patterns = [
+        'C:/Program Files (x86)/Microsoft Visual Studio/*/Professional/VC/Tools/MSVC/*/bin/Hostx64/x64',
+        'C:/Program Files (x86)/Microsoft Visual Studio/*/BuildTools/VC/Tools/MSVC/*/bin/Hostx64/x64',
+        'C:/Program Files (x86)/Microsoft Visual Studio/*/Community/VC/Tools/MSVC/*/bin/Hostx64/x64',
+        'C:/Program Files (x86)/Microsoft Visual Studio */vc/bin',
+    ]
+    for pattern in patterns:
+        matches = sorted(glob.glob(pattern))
+        if len(matches):
+            return matches[-1]
+    return None
+
+#----------------------------------------------------------------------------
+
+def _get_mangled_gpu_name():
+    name = torch.cuda.get_device_name().lower()
+    out = []
+    for c in name:
+        if re.match('[a-z0-9_-]+', c):
+            out.append(c)
+        else:
+            out.append('-')
+    return ''.join(out)
+
+#----------------------------------------------------------------------------
+# Main entry point for compiling and loading C++/CUDA plugins.
+
+_cached_plugins = dict()
+
+def get_plugin(module_name, sources, headers=None, source_dir=None, **build_kwargs):
+    assert verbosity in ['none', 'brief', 'full']
+    if headers is None:
+        headers = []
+    if source_dir is not None:
+        sources = [os.path.join(source_dir, fname) for fname in sources]
+        headers = [os.path.join(source_dir, fname) for fname in headers]
+
+    # Already cached?
+    if module_name in _cached_plugins:
+        return _cached_plugins[module_name]
+
+    # Print status.
+    if verbosity == 'full':
+        print(f'Setting up PyTorch plugin "{module_name}"...')
+    elif verbosity == 'brief':
+        print(f'Setting up PyTorch plugin "{module_name}"... ', end='', flush=True)
+    verbose_build = (verbosity == 'full')
+
+    # Compile and load.
+    try: # pylint: disable=too-many-nested-blocks
+        # Make sure we can find the necessary compiler binaries.
+        if os.name == 'nt' and os.system("where cl.exe >nul 2>nul") != 0:
+            compiler_bindir = _find_compiler_bindir()
+            if compiler_bindir is None:
+                raise RuntimeError(f'Could not find MSVC/GCC/CLANG installation on this computer. Check _find_compiler_bindir() in "{__file__}".')
+            os.environ['PATH'] += ';' + compiler_bindir
+
+        # Some containers set TORCH_CUDA_ARCH_LIST to a list that can either
+        # break the build or unnecessarily restrict what's available to nvcc.
+        # Unset it to let nvcc decide based on what's available on the
+        # machine.
+        os.environ['TORCH_CUDA_ARCH_LIST'] = ''
+
+        # Incremental build md5sum trickery.  Copies all the input source files
+        # into a cached build directory under a combined md5 digest of the input
+        # source files.  Copying is done only if the combined digest has changed.
+        # This keeps input file timestamps and filenames the same as in previous
+        # extension builds, allowing for fast incremental rebuilds.
+        #
+        # This optimization is done only in case all the source files reside in
+        # a single directory (just for simplicity) and if the TORCH_EXTENSIONS_DIR
+        # environment variable is set (we take this as a signal that the user
+        # actually cares about this.)
+        #
+        # EDIT: We now do it regardless of TORCH_EXTENSIOS_DIR, in order to work
+        # around the *.cu dependency bug in ninja config.
+        #
+        all_source_files = sorted(sources + headers)
+        all_source_dirs = set(os.path.dirname(fname) for fname in all_source_files)
+        if len(all_source_dirs) == 1: # and ('TORCH_EXTENSIONS_DIR' in os.environ):
+
+            # Compute combined hash digest for all source files.
+            hash_md5 = hashlib.md5()
+            for src in all_source_files:
+                with open(src, 'rb') as f:
+                    hash_md5.update(f.read())
+
+            # Select cached build directory name.
+            source_digest = hash_md5.hexdigest()
+            build_top_dir = torch.utils.cpp_extension._get_build_directory(module_name, verbose=verbose_build) # pylint: disable=protected-access
+            cached_build_dir = os.path.join(build_top_dir, f'{source_digest}-{_get_mangled_gpu_name()}')
+
+            if not os.path.isdir(cached_build_dir):
+                tmpdir = f'{build_top_dir}/srctmp-{uuid.uuid4().hex}'
+                os.makedirs(tmpdir)
+                for src in all_source_files:
+                    shutil.copyfile(src, os.path.join(tmpdir, os.path.basename(src)))
+                try:
+                    os.replace(tmpdir, cached_build_dir) # atomic
+                except OSError:
+                    # source directory already exists, delete tmpdir and its contents.
+                    shutil.rmtree(tmpdir)
+                    if not os.path.isdir(cached_build_dir): raise
+
+            # Compile.
+            cached_sources = [os.path.join(cached_build_dir, os.path.basename(fname)) for fname in sources]
+            torch.utils.cpp_extension.load(name=module_name, build_directory=cached_build_dir,
+                verbose=verbose_build, sources=cached_sources, **build_kwargs)
+        else:
+            torch.utils.cpp_extension.load(name=module_name, verbose=verbose_build, sources=sources, **build_kwargs)
+
+        # Load.
+        module = importlib.import_module(module_name)
+
+    except:
+        if verbosity == 'brief':
+            print('Failed!')
+        raise
+
+    # Print status and add to cache dict.
+    if verbosity == 'full':
+        print(f'Done setting up PyTorch plugin "{module_name}".')
+    elif verbosity == 'brief':
+        print('Done.')
+    _cached_plugins[module_name] = module
+    return module
+
+#----------------------------------------------------------------------------

models/torch_utils/misc.py

@@ -0,0 +1,265 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import re
+import contextlib
+import numpy as np
+import torch
+import warnings
+
+#----------------------------------------------------------------------------
+# Cached construction of constant tensors. Avoids CPU=>GPU copy when the
+# same constant is used multiple times.
+
+_constant_cache = dict()
+
+def constant(value, shape=None, dtype=None, device=None, memory_format=None):
+    value = np.asarray(value)
+    if shape is not None:
+        shape = tuple(shape)
+    if dtype is None:
+        dtype = torch.get_default_dtype()
+    if device is None:
+        device = torch.device('cpu')
+    if memory_format is None:
+        memory_format = torch.contiguous_format
+
+    key = (value.shape, value.dtype, value.tobytes(), shape, dtype, device, memory_format)
+    tensor = _constant_cache.get(key, None)
+    if tensor is None:
+        tensor = torch.as_tensor(value.copy(), dtype=dtype, device=device)
+        if shape is not None:
+            tensor, _ = torch.broadcast_tensors(tensor, torch.empty(shape))
+        tensor = tensor.contiguous(memory_format=memory_format)
+        _constant_cache[key] = tensor
+    return tensor
+
+#----------------------------------------------------------------------------
+# Replace NaN/Inf with specified numerical values.
+
+try:
+    nan_to_num = torch.nan_to_num # 1.8.0a0
+except AttributeError:
+    def nan_to_num(input, nan=0.0, posinf=None, neginf=None, *, out=None): # pylint: disable=redefined-builtin
+        assert isinstance(input, torch.Tensor)
+        if posinf is None:
+            posinf = torch.finfo(input.dtype).max
+        if neginf is None:
+            neginf = torch.finfo(input.dtype).min
+        assert nan == 0
+        return torch.clamp(input.unsqueeze(0).nansum(0), min=neginf, max=posinf, out=out)
+
+#----------------------------------------------------------------------------
+# Symbolic assert.
+
+try:
+    symbolic_assert = torch._assert # 1.8.0a0 # pylint: disable=protected-access
+except AttributeError:
+    symbolic_assert = torch.Assert # 1.7.0
+
+#----------------------------------------------------------------------------
+# Context manager to temporarily suppress known warnings in torch.jit.trace().
+# Note: Cannot use catch_warnings because of https://bugs.python.org/issue29672
+
+@contextlib.contextmanager
+def suppress_tracer_warnings():
+    flt = ('ignore', None, torch.jit.TracerWarning, None, 0)
+    warnings.filters.insert(0, flt)
+    yield
+    warnings.filters.remove(flt)
+
+#----------------------------------------------------------------------------
+# Assert that the shape of a tensor matches the given list of integers.
+# None indicates that the size of a dimension is allowed to vary.
+# Performs symbolic assertion when used in torch.jit.trace().
+
+def assert_shape(tensor, ref_shape):
+    if tensor.ndim != len(ref_shape):
+        raise AssertionError(f'Wrong number of dimensions: got {tensor.ndim}, expected {len(ref_shape)}')
+    for idx, (size, ref_size) in enumerate(zip(tensor.shape, ref_shape)):
+        if ref_size is None:
+            pass
+        elif isinstance(ref_size, torch.Tensor):
+            with suppress_tracer_warnings(): # as_tensor results are registered as constants
+                symbolic_assert(torch.equal(torch.as_tensor(size), ref_size), f'Wrong size for dimension {idx}')
+        elif isinstance(size, torch.Tensor):
+            with suppress_tracer_warnings(): # as_tensor results are registered as constants
+                symbolic_assert(torch.equal(size, torch.as_tensor(ref_size)), f'Wrong size for dimension {idx}: expected {ref_size}')
+        elif size != ref_size:
+            raise AssertionError(f'Wrong size for dimension {idx}: got {size}, expected {ref_size}')
+
+#----------------------------------------------------------------------------
+# Function decorator that calls torch.autograd.profiler.record_function().
+
+def profiled_function(fn):
+    def decorator(*args, **kwargs):
+        with torch.autograd.profiler.record_function(fn.__name__):
+            return fn(*args, **kwargs)
+    decorator.__name__ = fn.__name__
+    return decorator
+
+#----------------------------------------------------------------------------
+# Sampler for torch.utils.data.DataLoader that loops over the dataset
+# indefinitely, shuffling items as it goes.
+
+class InfiniteSampler(torch.utils.data.Sampler):
+    def __init__(self, dataset, rank=0, num_replicas=1, shuffle=True, seed=0, window_size=0.5):
+        assert len(dataset) > 0
+        assert num_replicas > 0
+        assert 0 <= rank < num_replicas
+        assert 0 <= window_size <= 1
+        super().__init__(dataset)
+        self.dataset = dataset
+        self.rank = rank
+        self.num_replicas = num_replicas
+        self.shuffle = shuffle
+        self.seed = seed
+        self.window_size = window_size
+
+    def __iter__(self):
+        order = np.arange(len(self.dataset))
+        rnd = None
+        window = 0
+        if self.shuffle:
+            rnd = np.random.RandomState(self.seed)
+            rnd.shuffle(order)
+            window = int(np.rint(order.size * self.window_size))
+
+        idx = 0
+        while True:
+            i = idx % order.size
+            if idx % self.num_replicas == self.rank:
+                yield order[i]
+            if window >= 2:
+                j = (i - rnd.randint(window)) % order.size
+                order[i], order[j] = order[j], order[i]
+            idx += 1
+
+#----------------------------------------------------------------------------
+# Utilities for operating with torch.nn.Module parameters and buffers.
+
+def params_and_buffers(module):
+    assert isinstance(module, torch.nn.Module)
+    return list(module.parameters()) + list(module.buffers())
+
+def named_params_and_buffers(module):
+    assert isinstance(module, torch.nn.Module)
+    return list(module.named_parameters()) + list(module.named_buffers())
+
+def copy_params_and_buffers(src_module, dst_module, require_all=False):
+    assert isinstance(src_module, torch.nn.Module)
+    assert isinstance(dst_module, torch.nn.Module)
+    src_tensors = dict(named_params_and_buffers(src_module))
+    for name, tensor in named_params_and_buffers(dst_module):
+        assert (name in src_tensors) or (not require_all)
+        if name in src_tensors:
+            tensor.copy_(src_tensors[name].detach()).requires_grad_(tensor.requires_grad)
+
+#----------------------------------------------------------------------------
+# Context manager for easily enabling/disabling DistributedDataParallel
+# synchronization.
+
+@contextlib.contextmanager
+def ddp_sync(module, sync):
+    assert isinstance(module, torch.nn.Module)
+    if sync or not isinstance(module, torch.nn.parallel.DistributedDataParallel):
+        yield
+    else:
+        with module.no_sync():
+            yield
+
+#----------------------------------------------------------------------------
+# Check DistributedDataParallel consistency across processes.
+
+def check_ddp_consistency(module, ignore_regex=None):
+    assert isinstance(module, torch.nn.Module)
+    for name, tensor in named_params_and_buffers(module):
+        fullname = type(module).__name__ + '.' + name
+        if ignore_regex is not None and re.fullmatch(ignore_regex, fullname):
+            continue
+        tensor = tensor.detach()
+        if tensor.is_floating_point():
+            tensor = nan_to_num(tensor)
+        other = tensor.clone()
+        torch.distributed.broadcast(tensor=other, src=0)
+        assert (tensor == other).all(), fullname
+
+#----------------------------------------------------------------------------
+# Print summary table of module hierarchy.
+
+def print_module_summary(module, inputs, max_nesting=3, skip_redundant=True):
+    assert isinstance(module, torch.nn.Module)
+    assert not isinstance(module, torch.jit.ScriptModule)
+    assert isinstance(inputs, (tuple, list))
+
+    # Register hooks.
+    entries = []
+    nesting = [0]
+    def pre_hook(_mod, _inputs):
+        nesting[0] += 1
+    def post_hook(mod, _inputs, outputs):
+        nesting[0] -= 1
+        if nesting[0] <= max_nesting:
+            outputs = list(outputs) if isinstance(outputs, (tuple, list)) else [outputs]
+            outputs = [t for t in outputs if isinstance(t, torch.Tensor)]
+            entries.append(dnnlib.EasyDict(mod=mod, outputs=outputs))
+    hooks = [mod.register_forward_pre_hook(pre_hook) for mod in module.modules()]
+    hooks += [mod.register_forward_hook(post_hook) for mod in module.modules()]
+
+    # Run module.
+    outputs = module(*inputs)
+    for hook in hooks:
+        hook.remove()
+
+    # Identify unique outputs, parameters, and buffers.
+    tensors_seen = set()
+    for e in entries:
+        e.unique_params = [t for t in e.mod.parameters() if id(t) not in tensors_seen]
+        e.unique_buffers = [t for t in e.mod.buffers() if id(t) not in tensors_seen]
+        e.unique_outputs = [t for t in e.outputs if id(t) not in tensors_seen]
+        tensors_seen |= {id(t) for t in e.unique_params + e.unique_buffers + e.unique_outputs}
+
+    # Filter out redundant entries.
+    if skip_redundant:
+        entries = [e for e in entries if len(e.unique_params) or len(e.unique_buffers) or len(e.unique_outputs)]
+
+    # Construct table.
+    rows = [[type(module).__name__, 'Parameters', 'Buffers', 'Output shape', 'Datatype']]
+    rows += [['---'] * len(rows[0])]
+    param_total = 0
+    buffer_total = 0
+    submodule_names = {mod: name for name, mod in module.named_modules()}
+    for e in entries:
+        name = '<top-level>' if e.mod is module else submodule_names[e.mod]
+        param_size = sum(t.numel() for t in e.unique_params)
+        buffer_size = sum(t.numel() for t in e.unique_buffers)
+        output_shapes = [str(list(t.shape)) for t in e.outputs]
+        output_dtypes = [str(t.dtype).split('.')[-1] for t in e.outputs]
+        rows += [[
+            name + (':0' if len(e.outputs) >= 2 else ''),
+            str(param_size) if param_size else '-',
+            str(buffer_size) if buffer_size else '-',
+            (output_shapes + ['-'])[0],
+            (output_dtypes + ['-'])[0],
+        ]]
+        for idx in range(1, len(e.outputs)):
+            rows += [[name + f':{idx}', '-', '-', output_shapes[idx], output_dtypes[idx]]]
+        param_total += param_size
+        buffer_total += buffer_size
+    rows += [['---'] * len(rows[0])]
+    rows += [['Total', str(param_total), str(buffer_total), '-', '-']]
+
+    # Print table.
+    widths = [max(len(cell) for cell in column) for column in zip(*rows)]
+    print()
+    for row in rows:
+        print('  '.join(cell + ' ' * (width - len(cell)) for cell, width in zip(row, widths)))
+    print()
+    return outputs
+
+#----------------------------------------------------------------------------

models/torch_utils/ops/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+# empty

models/torch_utils/ops/bias_act.cpp

@@ -0,0 +1,99 @@
+// Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#include <torch/extension.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include "bias_act.h"
+
+//------------------------------------------------------------------------
+
+static bool has_same_layout(torch::Tensor x, torch::Tensor y)
+{
+    if (x.dim() != y.dim())
+        return false;
+    for (int64_t i = 0; i < x.dim(); i++)
+    {
+        if (x.size(i) != y.size(i))
+            return false;
+        if (x.size(i) >= 2 && x.stride(i) != y.stride(i))
+            return false;
+    }
+    return true;
+}
+
+//------------------------------------------------------------------------
+
+static torch::Tensor bias_act(torch::Tensor x, torch::Tensor b, torch::Tensor xref, torch::Tensor yref, torch::Tensor dy, int grad, int dim, int act, float alpha, float gain, float clamp)
+{
+    // Validate arguments.
+    TORCH_CHECK(x.is_cuda(), "x must reside on CUDA device");
+    TORCH_CHECK(b.numel() == 0 || (b.dtype() == x.dtype() && b.device() == x.device()), "b must have the same dtype and device as x");
+    TORCH_CHECK(xref.numel() == 0 || (xref.sizes() == x.sizes() && xref.dtype() == x.dtype() && xref.device() == x.device()), "xref must have the same shape, dtype, and device as x");
+    TORCH_CHECK(yref.numel() == 0 || (yref.sizes() == x.sizes() && yref.dtype() == x.dtype() && yref.device() == x.device()), "yref must have the same shape, dtype, and device as x");
+    TORCH_CHECK(dy.numel() == 0 || (dy.sizes() == x.sizes() && dy.dtype() == x.dtype() && dy.device() == x.device()), "dy must have the same dtype and device as x");
+    TORCH_CHECK(x.numel() <= INT_MAX, "x is too large");
+    TORCH_CHECK(b.dim() == 1, "b must have rank 1");
+    TORCH_CHECK(b.numel() == 0 || (dim >= 0 && dim < x.dim()), "dim is out of bounds");
+    TORCH_CHECK(b.numel() == 0 || b.numel() == x.size(dim), "b has wrong number of elements");
+    TORCH_CHECK(grad >= 0, "grad must be non-negative");
+
+    // Validate layout.
+    TORCH_CHECK(x.is_non_overlapping_and_dense(), "x must be non-overlapping and dense");
+    TORCH_CHECK(b.is_contiguous(), "b must be contiguous");
+    TORCH_CHECK(xref.numel() == 0 || has_same_layout(xref, x), "xref must have the same layout as x");
+    TORCH_CHECK(yref.numel() == 0 || has_same_layout(yref, x), "yref must have the same layout as x");
+    TORCH_CHECK(dy.numel() == 0 || has_same_layout(dy, x), "dy must have the same layout as x");
+
+    // Create output tensor.
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
+    torch::Tensor y = torch::empty_like(x);
+    TORCH_CHECK(has_same_layout(y, x), "y must have the same layout as x");
+
+    // Initialize CUDA kernel parameters.
+    bias_act_kernel_params p;
+    p.x     = x.data_ptr();
+    p.b     = (b.numel()) ? b.data_ptr() : NULL;
+    p.xref  = (xref.numel()) ? xref.data_ptr() : NULL;
+    p.yref  = (yref.numel()) ? yref.data_ptr() : NULL;
+    p.dy    = (dy.numel()) ? dy.data_ptr() : NULL;
+    p.y     = y.data_ptr();
+    p.grad  = grad;
+    p.act   = act;
+    p.alpha = alpha;
+    p.gain  = gain;
+    p.clamp = clamp;
+    p.sizeX = (int)x.numel();
+    p.sizeB = (int)b.numel();
+    p.stepB = (b.numel()) ? (int)x.stride(dim) : 1;
+
+    // Choose CUDA kernel.
+    void* kernel;
+    AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), "upfirdn2d_cuda", [&]
+    {
+        kernel = choose_bias_act_kernel<scalar_t>(p);
+    });
+    TORCH_CHECK(kernel, "no CUDA kernel found for the specified activation func");
+
+    // Launch CUDA kernel.
+    p.loopX = 4;
+    int blockSize = 4 * 32;
+    int gridSize = (p.sizeX - 1) / (p.loopX * blockSize) + 1;
+    void* args[] = {&p};
+    AT_CUDA_CHECK(cudaLaunchKernel(kernel, gridSize, blockSize, args, 0, at::cuda::getCurrentCUDAStream()));
+    return y;
+}
+
+//------------------------------------------------------------------------
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m)
+{
+    m.def("bias_act", &bias_act);
+}
+
+//------------------------------------------------------------------------

models/torch_utils/ops/bias_act.cu

@@ -0,0 +1,173 @@
+// Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#include <c10/util/Half.h>
+#include "bias_act.h"
+
+//------------------------------------------------------------------------
+// Helpers.
+
+template <class T> struct InternalType;
+template <> struct InternalType<double>     { typedef double scalar_t; };
+template <> struct InternalType<float>      { typedef float  scalar_t; };
+template <> struct InternalType<c10::Half>  { typedef float  scalar_t; };
+
+//------------------------------------------------------------------------
+// CUDA kernel.
+
+template <class T, int A>
+__global__ void bias_act_kernel(bias_act_kernel_params p)
+{
+    typedef typename InternalType<T>::scalar_t scalar_t;
+    int G                 = p.grad;
+    scalar_t alpha        = (scalar_t)p.alpha;
+    scalar_t gain         = (scalar_t)p.gain;
+    scalar_t clamp        = (scalar_t)p.clamp;
+    scalar_t one          = (scalar_t)1;
+    scalar_t two          = (scalar_t)2;
+    scalar_t expRange     = (scalar_t)80;
+    scalar_t halfExpRange = (scalar_t)40;
+    scalar_t seluScale    = (scalar_t)1.0507009873554804934193349852946;
+    scalar_t seluAlpha    = (scalar_t)1.6732632423543772848170429916717;
+
+    // Loop over elements.
+    int xi = blockIdx.x * p.loopX * blockDim.x + threadIdx.x;
+    for (int loopIdx = 0; loopIdx < p.loopX && xi < p.sizeX; loopIdx++, xi += blockDim.x)
+    {
+        // Load.
+        scalar_t x = (scalar_t)((const T*)p.x)[xi];
+        scalar_t b = (p.b) ? (scalar_t)((const T*)p.b)[(xi / p.stepB) % p.sizeB] : 0;
+        scalar_t xref = (p.xref) ? (scalar_t)((const T*)p.xref)[xi] : 0;
+        scalar_t yref = (p.yref) ? (scalar_t)((const T*)p.yref)[xi] : 0;
+        scalar_t dy = (p.dy) ? (scalar_t)((const T*)p.dy)[xi] : one;
+        scalar_t yy = (gain != 0) ? yref / gain : 0;
+        scalar_t y = 0;
+
+        // Apply bias.
+        ((G == 0) ? x : xref) += b;
+
+        // linear
+        if (A == 1)
+        {
+            if (G == 0) y = x;
+            if (G == 1) y = x;
+        }
+
+        // relu
+        if (A == 2)
+        {
+            if (G == 0) y = (x > 0) ? x : 0;
+            if (G == 1) y = (yy > 0) ? x : 0;
+        }
+
+        // lrelu
+        if (A == 3)
+        {
+            if (G == 0) y = (x > 0) ? x : x * alpha;
+            if (G == 1) y = (yy > 0) ? x : x * alpha;
+        }
+
+        // tanh
+        if (A == 4)
+        {
+            if (G == 0) { scalar_t c = exp(x); scalar_t d = one / c; y = (x < -expRange) ? -one : (x > expRange) ? one : (c - d) / (c + d); }
+            if (G == 1) y = x * (one - yy * yy);
+            if (G == 2) y = x * (one - yy * yy) * (-two * yy);
+        }
+
+        // sigmoid
+        if (A == 5)
+        {
+            if (G == 0) y = (x < -expRange) ? 0 : one / (exp(-x) + one);
+            if (G == 1) y = x * yy * (one - yy);
+            if (G == 2) y = x * yy * (one - yy) * (one - two * yy);
+        }
+
+        // elu
+        if (A == 6)
+        {
+            if (G == 0) y = (x >= 0) ? x : exp(x) - one;
+            if (G == 1) y = (yy >= 0) ? x : x * (yy + one);
+            if (G == 2) y = (yy >= 0) ? 0 : x * (yy + one);
+        }
+
+        // selu
+        if (A == 7)
+        {
+            if (G == 0) y = (x >= 0) ? seluScale * x : (seluScale * seluAlpha) * (exp(x) - one);
+            if (G == 1) y = (yy >= 0) ? x * seluScale : x * (yy + seluScale * seluAlpha);
+            if (G == 2) y = (yy >= 0) ? 0 : x * (yy + seluScale * seluAlpha);
+        }
+
+        // softplus
+        if (A == 8)
+        {
+            if (G == 0) y = (x > expRange) ? x : log(exp(x) + one);
+            if (G == 1) y = x * (one - exp(-yy));
+            if (G == 2) { scalar_t c = exp(-yy); y = x * c * (one - c); }
+        }
+
+        // swish
+        if (A == 9)
+        {
+            if (G == 0)
+                y = (x < -expRange) ? 0 : x / (exp(-x) + one);
+            else
+            {
+                scalar_t c = exp(xref);
+                scalar_t d = c + one;
+                if (G == 1)
+                    y = (xref > halfExpRange) ? x : x * c * (xref + d) / (d * d);
+                else
+                    y = (xref > halfExpRange) ? 0 : x * c * (xref * (two - d) + two * d) / (d * d * d);
+                yref = (xref < -expRange) ? 0 : xref / (exp(-xref) + one) * gain;
+            }
+        }
+
+        // Apply gain.
+        y *= gain * dy;
+
+        // Clamp.
+        if (clamp >= 0)
+        {
+            if (G == 0)
+                y = (y > -clamp & y < clamp) ? y : (y >= 0) ? clamp : -clamp;
+            else
+                y = (yref > -clamp & yref < clamp) ? y : 0;
+        }
+
+        // Store.
+        ((T*)p.y)[xi] = (T)y;
+    }
+}
+
+//------------------------------------------------------------------------
+// CUDA kernel selection.
+
+template <class T> void* choose_bias_act_kernel(const bias_act_kernel_params& p)
+{
+    if (p.act == 1) return (void*)bias_act_kernel<T, 1>;
+    if (p.act == 2) return (void*)bias_act_kernel<T, 2>;
+    if (p.act == 3) return (void*)bias_act_kernel<T, 3>;
+    if (p.act == 4) return (void*)bias_act_kernel<T, 4>;
+    if (p.act == 5) return (void*)bias_act_kernel<T, 5>;
+    if (p.act == 6) return (void*)bias_act_kernel<T, 6>;
+    if (p.act == 7) return (void*)bias_act_kernel<T, 7>;
+    if (p.act == 8) return (void*)bias_act_kernel<T, 8>;
+    if (p.act == 9) return (void*)bias_act_kernel<T, 9>;
+    return NULL;
+}
+
+//------------------------------------------------------------------------
+// Template specializations.
+
+template void* choose_bias_act_kernel<double>       (const bias_act_kernel_params& p);
+template void* choose_bias_act_kernel<float>        (const bias_act_kernel_params& p);
+template void* choose_bias_act_kernel<c10::Half>    (const bias_act_kernel_params& p);
+
+//------------------------------------------------------------------------

models/torch_utils/ops/bias_act.h

@@ -0,0 +1,38 @@
+// Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+//------------------------------------------------------------------------
+// CUDA kernel parameters.
+
+struct bias_act_kernel_params
+{
+    const void* x;      // [sizeX]
+    const void* b;      // [sizeB] or NULL
+    const void* xref;   // [sizeX] or NULL
+    const void* yref;   // [sizeX] or NULL
+    const void* dy;     // [sizeX] or NULL
+    void*       y;      // [sizeX]
+
+    int         grad;
+    int         act;
+    float       alpha;
+    float       gain;
+    float       clamp;
+
+    int         sizeX;
+    int         sizeB;
+    int         stepB;
+    int         loopX;
+};
+
+//------------------------------------------------------------------------
+// CUDA kernel selection.
+
+template <class T> void* choose_bias_act_kernel(const bias_act_kernel_params& p);
+
+//------------------------------------------------------------------------

models/torch_utils/ops/bias_act.py

@@ -0,0 +1,209 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Custom PyTorch ops for efficient bias and activation."""
+
+import os
+import numpy as np
+import torch
+import models.dnnlib as dnnlib
+
+from .. import custom_ops
+from .. import misc
+
+#----------------------------------------------------------------------------
+
+activation_funcs = {
+    'linear':   dnnlib.EasyDict(func=lambda x, **_:         x,                                          def_alpha=0,    def_gain=1,             cuda_idx=1, ref='',  has_2nd_grad=False),
+    'relu':     dnnlib.EasyDict(func=lambda x, **_:         torch.nn.functional.relu(x),                def_alpha=0,    def_gain=np.sqrt(2),    cuda_idx=2, ref='y', has_2nd_grad=False),
+    'lrelu':    dnnlib.EasyDict(func=lambda x, alpha, **_:  torch.nn.functional.leaky_relu(x, alpha),   def_alpha=0.2,  def_gain=np.sqrt(2),    cuda_idx=3, ref='y', has_2nd_grad=False),
+    'tanh':     dnnlib.EasyDict(func=lambda x, **_:         torch.tanh(x),                              def_alpha=0,    def_gain=1,             cuda_idx=4, ref='y', has_2nd_grad=True),
+    'sigmoid':  dnnlib.EasyDict(func=lambda x, **_:         torch.sigmoid(x),                           def_alpha=0,    def_gain=1,             cuda_idx=5, ref='y', has_2nd_grad=True),
+    'elu':      dnnlib.EasyDict(func=lambda x, **_:         torch.nn.functional.elu(x),                 def_alpha=0,    def_gain=1,             cuda_idx=6, ref='y', has_2nd_grad=True),
+    'selu':     dnnlib.EasyDict(func=lambda x, **_:         torch.nn.functional.selu(x),                def_alpha=0,    def_gain=1,             cuda_idx=7, ref='y', has_2nd_grad=True),
+    'softplus': dnnlib.EasyDict(func=lambda x, **_:         torch.nn.functional.softplus(x),            def_alpha=0,    def_gain=1,             cuda_idx=8, ref='y', has_2nd_grad=True),
+    'swish':    dnnlib.EasyDict(func=lambda x, **_:         torch.sigmoid(x) * x,                       def_alpha=0,    def_gain=np.sqrt(2),    cuda_idx=9, ref='x', has_2nd_grad=True),
+}
+
+#----------------------------------------------------------------------------
+
+_plugin = None
+_null_tensor = torch.empty([0])
+
+def _init():
+    global _plugin
+    if _plugin is None:
+        _plugin = custom_ops.get_plugin(
+            module_name='bias_act_plugin',
+            sources=['bias_act.cpp', 'bias_act.cu'],
+            headers=['bias_act.h'],
+            source_dir=os.path.dirname(__file__),
+            extra_cuda_cflags=['--use_fast_math'],
+        )
+    return True
+
+#----------------------------------------------------------------------------
+
+def bias_act(x, b=None, dim=1, act='linear', alpha=None, gain=None, clamp=None, impl='cuda'):
+    r"""Fused bias and activation function.
+
+    Adds bias `b` to activation tensor `x`, evaluates activation function `act`,
+    and scales the result by `gain`. Each of the steps is optional. In most cases,
+    the fused op is considerably more efficient than performing the same calculation
+    using standard PyTorch ops. It supports first and second order gradients,
+    but not third order gradients.
+
+    Args:
+        x:      Input activation tensor. Can be of any shape.
+        b:      Bias vector, or `None` to disable. Must be a 1D tensor of the same type
+                as `x`. The shape must be known, and it must match the dimension of `x`
+                corresponding to `dim`.
+        dim:    The dimension in `x` corresponding to the elements of `b`.
+                The value of `dim` is ignored if `b` is not specified.
+        act:    Name of the activation function to evaluate, or `"linear"` to disable.
+                Can be e.g. `"relu"`, `"lrelu"`, `"tanh"`, `"sigmoid"`, `"swish"`, etc.
+                See `activation_funcs` for a full list. `None` is not allowed.
+        alpha:  Shape parameter for the activation function, or `None` to use the default.
+        gain:   Scaling factor for the output tensor, or `None` to use default.
+                See `activation_funcs` for the default scaling of each activation function.
+                If unsure, consider specifying 1.
+        clamp:  Clamp the output values to `[-clamp, +clamp]`, or `None` to disable
+                the clamping (default).
+        impl:   Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the same shape and datatype as `x`.
+    """
+    assert isinstance(x, torch.Tensor)
+    assert impl in ['ref', 'cuda']
+    if impl == 'cuda' and x.device.type == 'cuda' and _init():
+        return _bias_act_cuda(dim=dim, act=act, alpha=alpha, gain=gain, clamp=clamp).apply(x, b)
+    return _bias_act_ref(x=x, b=b, dim=dim, act=act, alpha=alpha, gain=gain, clamp=clamp)
+
+#----------------------------------------------------------------------------
+
+@misc.profiled_function
+def _bias_act_ref(x, b=None, dim=1, act='linear', alpha=None, gain=None, clamp=None):
+    """Slow reference implementation of `bias_act()` using standard TensorFlow ops.
+    """
+    assert isinstance(x, torch.Tensor)
+    assert clamp is None or clamp >= 0
+    spec = activation_funcs[act]
+    alpha = float(alpha if alpha is not None else spec.def_alpha)
+    gain = float(gain if gain is not None else spec.def_gain)
+    clamp = float(clamp if clamp is not None else -1)
+
+    # Add bias.
+    if b is not None:
+        assert isinstance(b, torch.Tensor) and b.ndim == 1
+        assert 0 <= dim < x.ndim
+        assert b.shape[0] == x.shape[dim]
+        x = x + b.reshape([-1 if i == dim else 1 for i in range(x.ndim)])
+
+    # Evaluate activation function.
+    alpha = float(alpha)
+    x = spec.func(x, alpha=alpha)
+
+    # Scale by gain.
+    gain = float(gain)
+    if gain != 1:
+        x = x * gain
+
+    # Clamp.
+    if clamp >= 0:
+        x = x.clamp(-clamp, clamp) # pylint: disable=invalid-unary-operand-type
+    return x
+
+#----------------------------------------------------------------------------
+
+_bias_act_cuda_cache = dict()
+
+def _bias_act_cuda(dim=1, act='linear', alpha=None, gain=None, clamp=None):
+    """Fast CUDA implementation of `bias_act()` using custom ops.
+    """
+    # Parse arguments.
+    assert clamp is None or clamp >= 0
+    spec = activation_funcs[act]
+    alpha = float(alpha if alpha is not None else spec.def_alpha)
+    gain = float(gain if gain is not None else spec.def_gain)
+    clamp = float(clamp if clamp is not None else -1)
+
+    # Lookup from cache.
+    key = (dim, act, alpha, gain, clamp)
+    if key in _bias_act_cuda_cache:
+        return _bias_act_cuda_cache[key]
+
+    # Forward op.
+    class BiasActCuda(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, x, b): # pylint: disable=arguments-differ
+            ctx.memory_format = torch.channels_last if x.ndim > 2 and x.stride(1) == 1 else torch.contiguous_format
+            x = x.contiguous(memory_format=ctx.memory_format)
+            b = b.contiguous() if b is not None else _null_tensor
+            y = x
+            if act != 'linear' or gain != 1 or clamp >= 0 or b is not _null_tensor:
+                y = _plugin.bias_act(x, b, _null_tensor, _null_tensor, _null_tensor, 0, dim, spec.cuda_idx, alpha, gain, clamp)
+            ctx.save_for_backward(
+                x if 'x' in spec.ref or spec.has_2nd_grad else _null_tensor,
+                b if 'x' in spec.ref or spec.has_2nd_grad else _null_tensor,
+                y if 'y' in spec.ref else _null_tensor)
+            return y
+
+        @staticmethod
+        def backward(ctx, dy): # pylint: disable=arguments-differ
+            dy = dy.contiguous(memory_format=ctx.memory_format)
+            x, b, y = ctx.saved_tensors
+            dx = None
+            db = None
+
+            if ctx.needs_input_grad[0] or ctx.needs_input_grad[1]:
+                dx = dy
+                if act != 'linear' or gain != 1 or clamp >= 0:
+                    dx = BiasActCudaGrad.apply(dy, x, b, y)
+
+            if ctx.needs_input_grad[1]:
+                db = dx.sum([i for i in range(dx.ndim) if i != dim])
+
+            return dx, db
+
+    # Backward op.
+    class BiasActCudaGrad(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, dy, x, b, y): # pylint: disable=arguments-differ
+            ctx.memory_format = torch.channels_last if dy.ndim > 2 and dy.stride(1) == 1 else torch.contiguous_format
+            dx = _plugin.bias_act(dy, b, x, y, _null_tensor, 1, dim, spec.cuda_idx, alpha, gain, clamp)
+            ctx.save_for_backward(
+                dy if spec.has_2nd_grad else _null_tensor,
+                x, b, y)
+            return dx
+
+        @staticmethod
+        def backward(ctx, d_dx): # pylint: disable=arguments-differ
+            d_dx = d_dx.contiguous(memory_format=ctx.memory_format)
+            dy, x, b, y = ctx.saved_tensors
+            d_dy = None
+            d_x = None
+            d_b = None
+            d_y = None
+
+            if ctx.needs_input_grad[0]:
+                d_dy = BiasActCudaGrad.apply(d_dx, x, b, y)
+
+            if spec.has_2nd_grad and (ctx.needs_input_grad[1] or ctx.needs_input_grad[2]):
+                d_x = _plugin.bias_act(d_dx, b, x, y, dy, 2, dim, spec.cuda_idx, alpha, gain, clamp)
+
+            if spec.has_2nd_grad and ctx.needs_input_grad[2]:
+                d_b = d_x.sum([i for i in range(d_x.ndim) if i != dim])
+
+            return d_dy, d_x, d_b, d_y
+
+    # Add to cache.
+    _bias_act_cuda_cache[key] = BiasActCuda
+    return BiasActCuda
+
+#----------------------------------------------------------------------------

models/torch_utils/ops/conv2d_gradfix.py

@@ -0,0 +1,198 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Custom replacement for `torch.nn.functional.conv2d` that supports
+arbitrarily high order gradients with zero performance penalty."""
+
+import contextlib
+import torch
+
+# pylint: disable=redefined-builtin
+# pylint: disable=arguments-differ
+# pylint: disable=protected-access
+
+#----------------------------------------------------------------------------
+
+enabled = False                     # Enable the custom op by setting this to true.
+weight_gradients_disabled = False   # Forcefully disable computation of gradients with respect to the weights.
+
+@contextlib.contextmanager
+def no_weight_gradients(disable=True):
+    global weight_gradients_disabled
+    old = weight_gradients_disabled
+    if disable:
+        weight_gradients_disabled = True
+    yield
+    weight_gradients_disabled = old
+
+#----------------------------------------------------------------------------
+
+def conv2d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1):
+    if _should_use_custom_op(input):
+        return _conv2d_gradfix(transpose=False, weight_shape=weight.shape, stride=stride, padding=padding, output_padding=0, dilation=dilation, groups=groups).apply(input, weight, bias)
+    return torch.nn.functional.conv2d(input=input, weight=weight, bias=bias, stride=stride, padding=padding, dilation=dilation, groups=groups)
+
+def conv_transpose2d(input, weight, bias=None, stride=1, padding=0, output_padding=0, groups=1, dilation=1):
+    if _should_use_custom_op(input):
+        return _conv2d_gradfix(transpose=True, weight_shape=weight.shape, stride=stride, padding=padding, output_padding=output_padding, groups=groups, dilation=dilation).apply(input, weight, bias)
+    return torch.nn.functional.conv_transpose2d(input=input, weight=weight, bias=bias, stride=stride, padding=padding, output_padding=output_padding, groups=groups, dilation=dilation)
+
+#----------------------------------------------------------------------------
+
+def _should_use_custom_op(input):
+    assert isinstance(input, torch.Tensor)
+    if (not enabled) or (not torch.backends.cudnn.enabled):
+        return False
+    if input.device.type != 'cuda':
+        return False
+    return True
+
+def _tuple_of_ints(xs, ndim):
+    xs = tuple(xs) if isinstance(xs, (tuple, list)) else (xs,) * ndim
+    assert len(xs) == ndim
+    assert all(isinstance(x, int) for x in xs)
+    return xs
+
+#----------------------------------------------------------------------------
+
+_conv2d_gradfix_cache = dict()
+_null_tensor = torch.empty([0])
+
+def _conv2d_gradfix(transpose, weight_shape, stride, padding, output_padding, dilation, groups):
+    # Parse arguments.
+    ndim = 2
+    weight_shape = tuple(weight_shape)
+    stride = _tuple_of_ints(stride, ndim)
+    padding = _tuple_of_ints(padding, ndim)
+    output_padding = _tuple_of_ints(output_padding, ndim)
+    dilation = _tuple_of_ints(dilation, ndim)
+
+    # Lookup from cache.
+    key = (transpose, weight_shape, stride, padding, output_padding, dilation, groups)
+    if key in _conv2d_gradfix_cache:
+        return _conv2d_gradfix_cache[key]
+
+    # Validate arguments.
+    assert groups >= 1
+    assert len(weight_shape) == ndim + 2
+    assert all(stride[i] >= 1 for i in range(ndim))
+    assert all(padding[i] >= 0 for i in range(ndim))
+    assert all(dilation[i] >= 0 for i in range(ndim))
+    if not transpose:
+        assert all(output_padding[i] == 0 for i in range(ndim))
+    else: # transpose
+        assert all(0 <= output_padding[i] < max(stride[i], dilation[i]) for i in range(ndim))
+
+    # Helpers.
+    common_kwargs = dict(stride=stride, padding=padding, dilation=dilation, groups=groups)
+    def calc_output_padding(input_shape, output_shape):
+        if transpose:
+            return [0, 0]
+        return [
+            input_shape[i + 2]
+            - (output_shape[i + 2] - 1) * stride[i]
+            - (1 - 2 * padding[i])
+            - dilation[i] * (weight_shape[i + 2] - 1)
+            for i in range(ndim)
+        ]
+
+    # Forward & backward.
+    class Conv2d(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, input, weight, bias):
+            assert weight.shape == weight_shape
+            ctx.save_for_backward(
+                input if weight.requires_grad else _null_tensor,
+                weight if input.requires_grad else _null_tensor,
+            )
+            ctx.input_shape = input.shape
+
+            # Simple 1x1 convolution => cuBLAS (only on Volta, not on Ampere).
+            if weight_shape[2:] == stride == dilation == (1, 1) and padding == (0, 0) and torch.cuda.get_device_capability(input.device) < (8, 0):
+                a = weight.reshape(groups, weight_shape[0] // groups, weight_shape[1])
+                b = input.reshape(input.shape[0], groups, input.shape[1] // groups, -1)
+                c = (a.transpose(1, 2) if transpose else a) @ b.permute(1, 2, 0, 3).flatten(2)
+                c = c.reshape(-1, input.shape[0], *input.shape[2:]).transpose(0, 1)
+                c = c if bias is None else c + bias.unsqueeze(0).unsqueeze(2).unsqueeze(3)
+                return c.contiguous(memory_format=(torch.channels_last if input.stride(1) == 1 else torch.contiguous_format))
+
+            # General case => cuDNN.
+            if transpose:
+                return torch.nn.functional.conv_transpose2d(input=input, weight=weight, bias=bias, output_padding=output_padding, **common_kwargs)
+            return torch.nn.functional.conv2d(input=input, weight=weight, bias=bias, **common_kwargs)
+
+        @staticmethod
+        def backward(ctx, grad_output):
+            input, weight = ctx.saved_tensors
+            input_shape = ctx.input_shape
+            grad_input = None
+            grad_weight = None
+            grad_bias = None
+
+            if ctx.needs_input_grad[0]:
+                p = calc_output_padding(input_shape=input_shape, output_shape=grad_output.shape)
+                op = _conv2d_gradfix(transpose=(not transpose), weight_shape=weight_shape, output_padding=p, **common_kwargs)
+                grad_input = op.apply(grad_output, weight, None)
+                assert grad_input.shape == input_shape
+
+            if ctx.needs_input_grad[1] and not weight_gradients_disabled:
+                grad_weight = Conv2dGradWeight.apply(grad_output, input)
+                assert grad_weight.shape == weight_shape
+
+            if ctx.needs_input_grad[2]:
+                grad_bias = grad_output.sum([0, 2, 3])
+
+            return grad_input, grad_weight, grad_bias
+
+    # Gradient with respect to the weights.
+    class Conv2dGradWeight(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, grad_output, input):
+            ctx.save_for_backward(
+                grad_output if input.requires_grad else _null_tensor,
+                input if grad_output.requires_grad else _null_tensor,
+            )
+            ctx.grad_output_shape = grad_output.shape
+            ctx.input_shape = input.shape
+
+            # Simple 1x1 convolution => cuBLAS (on both Volta and Ampere).
+            if weight_shape[2:] == stride == dilation == (1, 1) and padding == (0, 0):
+                a = grad_output.reshape(grad_output.shape[0], groups, grad_output.shape[1] // groups, -1).permute(1, 2, 0, 3).flatten(2)
+                b = input.reshape(input.shape[0], groups, input.shape[1] // groups, -1).permute(1, 2, 0, 3).flatten(2)
+                c = (b @ a.transpose(1, 2) if transpose else a @ b.transpose(1, 2)).reshape(weight_shape)
+                return c.contiguous(memory_format=(torch.channels_last if input.stride(1) == 1 else torch.contiguous_format))
+
+            # General case => cuDNN.
+            name = 'aten::cudnn_convolution_transpose_backward_weight' if transpose else 'aten::cudnn_convolution_backward_weight'
+            flags = [torch.backends.cudnn.benchmark, torch.backends.cudnn.deterministic, torch.backends.cudnn.allow_tf32]
+            return torch._C._jit_get_operation(name)(weight_shape, grad_output, input, padding, stride, dilation, groups, *flags)
+
+        @staticmethod
+        def backward(ctx, grad2_grad_weight):
+            grad_output, input = ctx.saved_tensors
+            grad_output_shape = ctx.grad_output_shape
+            input_shape = ctx.input_shape
+            grad2_grad_output = None
+            grad2_input = None
+
+            if ctx.needs_input_grad[0]:
+                grad2_grad_output = Conv2d.apply(input, grad2_grad_weight, None)
+                assert grad2_grad_output.shape == grad_output_shape
+
+            if ctx.needs_input_grad[1]:
+                p = calc_output_padding(input_shape=input_shape, output_shape=grad_output_shape)
+                op = _conv2d_gradfix(transpose=(not transpose), weight_shape=weight_shape, output_padding=p, **common_kwargs)
+                grad2_input = op.apply(grad_output, grad2_grad_weight, None)
+                assert grad2_input.shape == input_shape
+
+            return grad2_grad_output, grad2_input
+
+    _conv2d_gradfix_cache[key] = Conv2d
+    return Conv2d
+
+#----------------------------------------------------------------------------

models/torch_utils/ops/conv2d_resample.py

@@ -0,0 +1,143 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""2D convolution with optional up/downsampling."""
+
+import torch
+
+from .. import misc
+from . import conv2d_gradfix
+from . import upfirdn2d
+from .upfirdn2d import _parse_padding
+from .upfirdn2d import _get_filter_size
+
+#----------------------------------------------------------------------------
+
+def _get_weight_shape(w):
+    with misc.suppress_tracer_warnings(): # this value will be treated as a constant
+        shape = [int(sz) for sz in w.shape]
+    misc.assert_shape(w, shape)
+    return shape
+
+#----------------------------------------------------------------------------
+
+def _conv2d_wrapper(x, w, stride=1, padding=0, groups=1, transpose=False, flip_weight=True):
+    """Wrapper for the underlying `conv2d()` and `conv_transpose2d()` implementations.
+    """
+    _out_channels, _in_channels_per_group, kh, kw = _get_weight_shape(w)
+
+    # Flip weight if requested.
+    # Note: conv2d() actually performs correlation (flip_weight=True) not convolution (flip_weight=False).
+    if not flip_weight and (kw > 1 or kh > 1):
+        w = w.flip([2, 3])
+
+    # Execute using conv2d_gradfix.
+    op = conv2d_gradfix.conv_transpose2d if transpose else conv2d_gradfix.conv2d
+    return op(x, w, stride=stride, padding=padding, groups=groups)
+
+#----------------------------------------------------------------------------
+
+@misc.profiled_function
+def conv2d_resample(x, w, f=None, up=1, down=1, padding=0, groups=1, flip_weight=True, flip_filter=False):
+    r"""2D convolution with optional up/downsampling.
+
+    Padding is performed only once at the beginning, not between the operations.
+
+    Args:
+        x:              Input tensor of shape
+                        `[batch_size, in_channels, in_height, in_width]`.
+        w:              Weight tensor of shape
+                        `[out_channels, in_channels//groups, kernel_height, kernel_width]`.
+        f:              Low-pass filter for up/downsampling. Must be prepared beforehand by
+                        calling upfirdn2d.setup_filter(). None = identity (default).
+        up:             Integer upsampling factor (default: 1).
+        down:           Integer downsampling factor (default: 1).
+        padding:        Padding with respect to the upsampled image. Can be a single number
+                        or a list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
+                        (default: 0).
+        groups:         Split input channels into N groups (default: 1).
+        flip_weight:    False = convolution, True = correlation (default: True).
+        flip_filter:    False = convolution, True = correlation (default: False).
+
+    Returns:
+        Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
+    """
+    # Validate arguments.
+    assert isinstance(x, torch.Tensor) and (x.ndim == 4)
+    assert isinstance(w, torch.Tensor) and (w.ndim == 4) and (w.dtype == x.dtype)
+    assert f is None or (isinstance(f, torch.Tensor) and f.ndim in [1, 2] and f.dtype == torch.float32)
+    assert isinstance(up, int) and (up >= 1)
+    assert isinstance(down, int) and (down >= 1)
+    assert isinstance(groups, int) and (groups >= 1)
+    out_channels, in_channels_per_group, kh, kw = _get_weight_shape(w)
+    fw, fh = _get_filter_size(f)
+    px0, px1, py0, py1 = _parse_padding(padding)
+
+    # Adjust padding to account for up/downsampling.
+    if up > 1:
+        px0 += (fw + up - 1) // 2
+        px1 += (fw - up) // 2
+        py0 += (fh + up - 1) // 2
+        py1 += (fh - up) // 2
+    if down > 1:
+        px0 += (fw - down + 1) // 2
+        px1 += (fw - down) // 2
+        py0 += (fh - down + 1) // 2
+        py1 += (fh - down) // 2
+
+    # Fast path: 1x1 convolution with downsampling only => downsample first, then convolve.
+    if kw == 1 and kh == 1 and (down > 1 and up == 1):
+        x = upfirdn2d.upfirdn2d(x=x, f=f, down=down, padding=[px0,px1,py0,py1], flip_filter=flip_filter)
+        x = _conv2d_wrapper(x=x, w=w, groups=groups, flip_weight=flip_weight)
+        return x
+
+    # Fast path: 1x1 convolution with upsampling only => convolve first, then upsample.
+    if kw == 1 and kh == 1 and (up > 1 and down == 1):
+        x = _conv2d_wrapper(x=x, w=w, groups=groups, flip_weight=flip_weight)
+        x = upfirdn2d.upfirdn2d(x=x, f=f, up=up, padding=[px0,px1,py0,py1], gain=up**2, flip_filter=flip_filter)
+        return x
+
+    # Fast path: downsampling only => use strided convolution.
+    if down > 1 and up == 1:
+        x = upfirdn2d.upfirdn2d(x=x, f=f, padding=[px0,px1,py0,py1], flip_filter=flip_filter)
+        x = _conv2d_wrapper(x=x, w=w, stride=down, groups=groups, flip_weight=flip_weight)
+        return x
+
+    # Fast path: upsampling with optional downsampling => use transpose strided convolution.
+    if up > 1:
+        if groups == 1:
+            w = w.transpose(0, 1)
+        else:
+            w = w.reshape(groups, out_channels // groups, in_channels_per_group, kh, kw)
+            w = w.transpose(1, 2)
+            w = w.reshape(groups * in_channels_per_group, out_channels // groups, kh, kw)
+        px0 -= kw - 1
+        px1 -= kw - up
+        py0 -= kh - 1
+        py1 -= kh - up
+        pxt = max(min(-px0, -px1), 0)
+        pyt = max(min(-py0, -py1), 0)
+        x = _conv2d_wrapper(x=x, w=w, stride=up, padding=[pyt,pxt], groups=groups, transpose=True, flip_weight=(not flip_weight))
+        x = upfirdn2d.upfirdn2d(x=x, f=f, padding=[px0+pxt,px1+pxt,py0+pyt,py1+pyt], gain=up**2, flip_filter=flip_filter)
+        if down > 1:
+            x = upfirdn2d.upfirdn2d(x=x, f=f, down=down, flip_filter=flip_filter)
+        return x
+
+    # Fast path: no up/downsampling, padding supported by the underlying implementation => use plain conv2d.
+    if up == 1 and down == 1:
+        if px0 == px1 and py0 == py1 and px0 >= 0 and py0 >= 0:
+            return _conv2d_wrapper(x=x, w=w, padding=[py0,px0], groups=groups, flip_weight=flip_weight)
+
+    # Fallback: Generic reference implementation.
+    x = upfirdn2d.upfirdn2d(x=x, f=(f if up > 1 else None), up=up, padding=[px0,px1,py0,py1], gain=up**2, flip_filter=flip_filter)
+    x = _conv2d_wrapper(x=x, w=w, groups=groups, flip_weight=flip_weight)
+    if down > 1:
+        x = upfirdn2d.upfirdn2d(x=x, f=f, down=down, flip_filter=flip_filter)
+    return x
+
+#----------------------------------------------------------------------------

models/torch_utils/ops/filtered_lrelu.cpp

@@ -0,0 +1,300 @@
+// Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#include <torch/extension.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include "filtered_lrelu.h"
+
+//------------------------------------------------------------------------
+
+static std::tuple<torch::Tensor, torch::Tensor, int> filtered_lrelu(
+    torch::Tensor x, torch::Tensor fu, torch::Tensor fd, torch::Tensor b, torch::Tensor si,
+    int up, int down, int px0, int px1, int py0, int py1, int sx, int sy, float gain, float slope, float clamp, bool flip_filters, bool writeSigns)
+{
+    // Set CUDA device.
+    TORCH_CHECK(x.is_cuda(), "x must reside on CUDA device");
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
+
+    // Validate arguments.
+    TORCH_CHECK(fu.device() == x.device() && fd.device() == x.device() && b.device() == x.device(), "all input tensors must reside on the same device");
+    TORCH_CHECK(fu.dtype() == torch::kFloat && fd.dtype() == torch::kFloat, "fu and fd must be float32");
+    TORCH_CHECK(b.dtype() == x.dtype(), "x and b must have the same dtype");
+    TORCH_CHECK(x.dtype() == torch::kHalf || x.dtype() == torch::kFloat, "x and b must be float16 or float32");
+    TORCH_CHECK(x.dim() == 4, "x must be rank 4");
+    TORCH_CHECK(x.size(0) * x.size(1) <= INT_MAX && x.size(2) <= INT_MAX && x.size(3) <= INT_MAX, "x is too large");
+    TORCH_CHECK(x.numel() > 0, "x is empty");
+    TORCH_CHECK((fu.dim() == 1 || fu.dim() == 2) && (fd.dim() == 1 || fd.dim() == 2), "fu and fd must be rank 1 or 2");
+    TORCH_CHECK(fu.size(0) <= INT_MAX && fu.size(-1) <= INT_MAX, "fu is too large");
+    TORCH_CHECK(fd.size(0) <= INT_MAX && fd.size(-1) <= INT_MAX, "fd is too large");
+    TORCH_CHECK(fu.numel() > 0, "fu is empty");
+    TORCH_CHECK(fd.numel() > 0, "fd is empty");
+    TORCH_CHECK(b.dim() == 1 && b.size(0) == x.size(1), "b must be a vector with the same number of channels as x");
+    TORCH_CHECK(up >= 1 && down >= 1, "up and down must be at least 1");
+
+    // Figure out how much shared memory is available on the device.
+    int maxSharedBytes = 0;
+    AT_CUDA_CHECK(cudaDeviceGetAttribute(&maxSharedBytes, cudaDevAttrMaxSharedMemoryPerBlockOptin, x.device().index()));
+    int sharedKB = maxSharedBytes >> 10;
+
+    // Populate enough launch parameters to check if a CUDA kernel exists.
+    filtered_lrelu_kernel_params p;
+    p.up      = up;
+    p.down    = down;
+    p.fuShape = make_int2((int)fu.size(-1), fu.dim() == 2 ? (int)fu.size(0) : 0); // shape [n, 0] indicates separable filter.
+    p.fdShape = make_int2((int)fd.size(-1), fd.dim() == 2 ? (int)fd.size(0) : 0);
+    filtered_lrelu_kernel_spec test_spec = choose_filtered_lrelu_kernel<float, int32_t, false, false>(p, sharedKB);
+    if (!test_spec.exec)
+    {
+        // No kernel found - return empty tensors and indicate missing kernel with return code of -1.
+        return std::make_tuple(torch::Tensor(), torch::Tensor(), -1);
+    }
+
+    // Input/output element size.
+    int64_t sz = (x.dtype() == torch::kHalf) ? 2 : 4;
+
+    // Input sizes.
+    int64_t xw = (int)x.size(3);
+    int64_t xh = (int)x.size(2);
+    int64_t fut_w = (int)fu.size(-1) - 1;
+    int64_t fut_h = (int)fu.size(0)  - 1;
+    int64_t fdt_w = (int)fd.size(-1) - 1;
+    int64_t fdt_h = (int)fd.size(0)  - 1;
+
+    // Logical size of upsampled buffer.
+    int64_t cw = xw * up + (px0 + px1) - fut_w;
+    int64_t ch = xh * up + (py0 + py1) - fut_h;
+    TORCH_CHECK(cw > fdt_w && ch > fdt_h, "upsampled buffer must be at least the size of downsampling filter");
+    TORCH_CHECK(cw <= INT_MAX && ch <= INT_MAX, "upsampled buffer is too large");
+
+    // Compute output size and allocate.
+    int64_t yw = (cw - fdt_w + (down - 1)) / down;
+    int64_t yh = (ch - fdt_h + (down - 1)) / down;
+    TORCH_CHECK(yw > 0 && yh > 0, "output must be at least 1x1");
+    TORCH_CHECK(yw <= INT_MAX && yh <= INT_MAX, "output is too large");
+    torch::Tensor y = torch::empty({x.size(0), x.size(1), yh, yw}, x.options(), x.suggest_memory_format());
+
+    // Allocate sign tensor.
+    torch::Tensor so;
+    torch::Tensor s = si;
+    bool readSigns = !!s.numel();
+    int64_t sw_active = 0; // Active width of sign tensor.
+    if (writeSigns)
+    {
+        sw_active = yw * down - (down - 1) + fdt_w;     // Active width in elements.
+        int64_t sh = yh * down - (down - 1) + fdt_h;    // Height = active height.
+        int64_t sw = (sw_active + 15) & ~15;            // Width  = active width in elements, rounded up to multiple of 16.
+        TORCH_CHECK(sh <= INT_MAX && (sw >> 2) <= INT_MAX, "signs is too large");
+        s = so = torch::empty({x.size(0), x.size(1), sh, sw >> 2}, x.options().dtype(torch::kUInt8), at::MemoryFormat::Contiguous);
+    }
+    else if (readSigns)
+        sw_active = s.size(3) << 2;
+
+    // Validate sign tensor if in use.
+    if (readSigns || writeSigns)
+    {
+        TORCH_CHECK(s.is_contiguous(), "signs must be contiguous");
+        TORCH_CHECK(s.dtype() == torch::kUInt8, "signs must be uint8");
+        TORCH_CHECK(s.device() == x.device(), "signs must reside on the same device as x");
+        TORCH_CHECK(s.dim() == 4, "signs must be rank 4");
+        TORCH_CHECK(s.size(0) == x.size(0) && s.size(1) == x.size(1), "signs must have same batch & channels as x");
+        TORCH_CHECK(s.size(2) <= INT_MAX && s.size(3) <= INT_MAX, "signs is too large");
+    }
+
+    // Populate rest of CUDA kernel parameters.
+    p.x         = x.data_ptr();
+    p.y         = y.data_ptr();
+    p.b         = b.data_ptr();
+    p.s         = (readSigns || writeSigns) ? s.data_ptr<unsigned char>() : 0;
+    p.fu        = fu.data_ptr<float>();
+    p.fd        = fd.data_ptr<float>();
+    p.pad0      = make_int2(px0, py0);
+    p.gain      = gain;
+    p.slope     = slope;
+    p.clamp     = clamp;
+    p.flip      = (flip_filters) ? 1 : 0;
+    p.xShape    = make_int4((int)x.size(3), (int)x.size(2), (int)x.size(1), (int)x.size(0));
+    p.yShape    = make_int4((int)y.size(3), (int)y.size(2), (int)y.size(1), (int)y.size(0));
+    p.sShape    = (readSigns || writeSigns) ? make_int2((int)s.size(3), (int)s.size(2)) : make_int2(0, 0); // Width is in bytes. Contiguous.
+    p.sOfs      = make_int2(sx, sy);
+    p.swLimit   = (sw_active + 3) >> 2; // Rounded up to bytes.
+
+    // x, y, b strides are in bytes.
+    p.xStride   = make_longlong4(sz * x.stride(3), sz * x.stride(2), sz * x.stride(1), sz * x.stride(0));
+    p.yStride   = make_longlong4(sz * y.stride(3), sz * y.stride(2), sz * y.stride(1), sz * y.stride(0));
+    p.bStride   = sz * b.stride(0);
+
+    // fu, fd strides are in elements.
+    p.fuStride  = make_longlong3(fu.stride(-1), fu.dim() == 2 ? fu.stride(0) : 0, 0);
+    p.fdStride  = make_longlong3(fd.stride(-1), fd.dim() == 2 ? fd.stride(0) : 0, 0);
+
+    // Determine if indices don't fit in int32. Support negative strides although Torch currently never produces those.
+    bool index64b = false;
+    if (std::abs(p.bStride * x.size(1)) > INT_MAX) index64b = true;
+    if (std::min(x.size(0) * p.xStride.w, 0ll) + std::min(x.size(1) * p.xStride.z, 0ll) + std::min(x.size(2) * p.xStride.y, 0ll) + std::min(x.size(3) * p.xStride.x, 0ll) < -INT_MAX) index64b = true;
+    if (std::max(x.size(0) * p.xStride.w, 0ll) + std::max(x.size(1) * p.xStride.z, 0ll) + std::max(x.size(2) * p.xStride.y, 0ll) + std::max(x.size(3) * p.xStride.x, 0ll) >  INT_MAX) index64b = true;
+    if (std::min(y.size(0) * p.yStride.w, 0ll) + std::min(y.size(1) * p.yStride.z, 0ll) + std::min(y.size(2) * p.yStride.y, 0ll) + std::min(y.size(3) * p.yStride.x, 0ll) < -INT_MAX) index64b = true;
+    if (std::max(y.size(0) * p.yStride.w, 0ll) + std::max(y.size(1) * p.yStride.z, 0ll) + std::max(y.size(2) * p.yStride.y, 0ll) + std::max(y.size(3) * p.yStride.x, 0ll) >  INT_MAX) index64b = true;
+    if (s.numel() > INT_MAX) index64b = true;
+
+    // Choose CUDA kernel.
+    filtered_lrelu_kernel_spec spec = { 0 };
+    AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), "filtered_lrelu_cuda", [&]
+    {
+        if constexpr (sizeof(scalar_t) <= 4) // Exclude doubles. constexpr prevents template instantiation.
+        {
+            // Choose kernel based on index type, datatype and sign read/write modes.
+            if      (!index64b &&  writeSigns && !readSigns) spec = choose_filtered_lrelu_kernel<scalar_t, int32_t, true,  false>(p, sharedKB);
+            else if (!index64b && !writeSigns &&  readSigns) spec = choose_filtered_lrelu_kernel<scalar_t, int32_t, false, true >(p, sharedKB);
+            else if (!index64b && !writeSigns && !readSigns) spec = choose_filtered_lrelu_kernel<scalar_t, int32_t, false, false>(p, sharedKB);
+            else if ( index64b &&  writeSigns && !readSigns) spec = choose_filtered_lrelu_kernel<scalar_t, int64_t, true,  false>(p, sharedKB);
+            else if ( index64b && !writeSigns &&  readSigns) spec = choose_filtered_lrelu_kernel<scalar_t, int64_t, false, true >(p, sharedKB);
+            else if ( index64b && !writeSigns && !readSigns) spec = choose_filtered_lrelu_kernel<scalar_t, int64_t, false, false>(p, sharedKB);
+        }
+    });
+    TORCH_CHECK(spec.exec, "internal error - CUDA kernel not found") // This should not happen because we tested earlier that kernel exists.
+
+    // Launch CUDA kernel.
+    void* args[] = {&p};
+    int bx = spec.numWarps * 32;
+    int gx = (p.yShape.x - 1) / spec.tileOut.x + 1;
+    int gy = (p.yShape.y - 1) / spec.tileOut.y + 1;
+    int gz = p.yShape.z * p.yShape.w;
+
+    // Repeat multiple horizontal tiles in a CTA?
+    if (spec.xrep)
+    {
+        p.tilesXrep = spec.xrep;
+        p.tilesXdim = gx;
+
+        gx = (gx + p.tilesXrep - 1) / p.tilesXrep;
+        std::swap(gx, gy);
+    }
+    else
+    {
+        p.tilesXrep = 0;
+        p.tilesXdim = 0;
+    }
+
+    // Launch filter setup kernel.
+    AT_CUDA_CHECK(cudaLaunchKernel(spec.setup, 1, 1024, args, 0, at::cuda::getCurrentCUDAStream()));
+
+    // Copy kernels to constant memory.
+    if      ( writeSigns && !readSigns) AT_CUDA_CHECK((copy_filters<true,  false>(at::cuda::getCurrentCUDAStream())));
+    else if (!writeSigns &&  readSigns) AT_CUDA_CHECK((copy_filters<false, true >(at::cuda::getCurrentCUDAStream())));
+    else if (!writeSigns && !readSigns) AT_CUDA_CHECK((copy_filters<false, false>(at::cuda::getCurrentCUDAStream())));
+
+    // Set cache and shared memory configurations for main kernel.
+    AT_CUDA_CHECK(cudaFuncSetCacheConfig(spec.exec, cudaFuncCachePreferShared));
+    if (spec.dynamicSharedKB) // Need dynamically allocated shared memory?
+        AT_CUDA_CHECK(cudaFuncSetAttribute(spec.exec, cudaFuncAttributeMaxDynamicSharedMemorySize, spec.dynamicSharedKB << 10));
+    AT_CUDA_CHECK(cudaFuncSetSharedMemConfig(spec.exec, cudaSharedMemBankSizeFourByte));
+
+    // Launch main kernel.
+    const int maxSubGz = 65535; // CUDA maximum for block z dimension.
+    for (int zofs=0; zofs < gz; zofs += maxSubGz) // Do multiple launches if gz is too big.
+    {
+        p.blockZofs = zofs;
+        int subGz = std::min(maxSubGz, gz - zofs);
+        AT_CUDA_CHECK(cudaLaunchKernel(spec.exec, dim3(gx, gy, subGz), bx, args, spec.dynamicSharedKB << 10, at::cuda::getCurrentCUDAStream()));
+    }
+
+    // Done.
+    return std::make_tuple(y, so, 0);
+}
+
+//------------------------------------------------------------------------
+
+static torch::Tensor filtered_lrelu_act(torch::Tensor x, torch::Tensor si, int sx, int sy, float gain, float slope, float clamp, bool writeSigns)
+{
+    // Set CUDA device.
+    TORCH_CHECK(x.is_cuda(), "x must reside on CUDA device");
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
+
+    // Validate arguments.
+    TORCH_CHECK(x.dim() == 4, "x must be rank 4");
+    TORCH_CHECK(x.size(0) * x.size(1) <= INT_MAX && x.size(2) <= INT_MAX && x.size(3) <= INT_MAX, "x is too large");
+    TORCH_CHECK(x.numel() > 0, "x is empty");
+    TORCH_CHECK(x.dtype() == torch::kHalf || x.dtype() == torch::kFloat || x.dtype() == torch::kDouble, "x must be float16, float32 or float64");
+
+    // Output signs if we don't have sign input.
+    torch::Tensor so;
+    torch::Tensor s = si;
+    bool readSigns = !!s.numel();
+    if (writeSigns)
+    {
+        int64_t sw = x.size(3);
+        sw = (sw + 15) & ~15; // Round to a multiple of 16 for coalescing.
+        s = so = torch::empty({x.size(0), x.size(1), x.size(2), sw >> 2}, x.options().dtype(torch::kUInt8), at::MemoryFormat::Contiguous);
+    }
+
+    // Validate sign tensor if in use.
+    if (readSigns || writeSigns)
+    {
+        TORCH_CHECK(s.is_contiguous(), "signs must be contiguous");
+        TORCH_CHECK(s.dtype() == torch::kUInt8, "signs must be uint8");
+        TORCH_CHECK(s.device() == x.device(), "signs must reside on the same device as x");
+        TORCH_CHECK(s.dim() == 4, "signs must be rank 4");
+        TORCH_CHECK(s.size(0) == x.size(0) && s.size(1) == x.size(1), "signs must have same batch & channels as x");
+        TORCH_CHECK(s.size(2) <= INT_MAX && (s.size(3) << 2) <= INT_MAX, "signs tensor is too large");
+    }
+
+    // Initialize CUDA kernel parameters.
+    filtered_lrelu_act_kernel_params p;
+    p.x         = x.data_ptr();
+    p.s         = (readSigns || writeSigns) ? s.data_ptr<unsigned char>() : 0;
+    p.gain      = gain;
+    p.slope     = slope;
+    p.clamp     = clamp;
+    p.xShape    = make_int4((int)x.size(3), (int)x.size(2), (int)x.size(1), (int)x.size(0));
+    p.xStride   = make_longlong4(x.stride(3), x.stride(2), x.stride(1), x.stride(0));
+    p.sShape    = (readSigns || writeSigns) ? make_int2((int)s.size(3) << 2, (int)s.size(2)) : make_int2(0, 0); // Width is in elements. Contiguous.
+    p.sOfs      = make_int2(sx, sy);
+
+    // Choose CUDA kernel.
+    void* func = 0;
+    AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), "filtered_lrelu_act_cuda", [&]
+    {
+        if (writeSigns)
+            func = choose_filtered_lrelu_act_kernel<scalar_t, true, false>();
+        else if (readSigns)
+            func = choose_filtered_lrelu_act_kernel<scalar_t, false, true>();
+        else
+            func = choose_filtered_lrelu_act_kernel<scalar_t, false, false>();
+    });
+    TORCH_CHECK(func, "internal error - CUDA kernel not found");
+
+    // Launch CUDA kernel.
+    void* args[] = {&p};
+    int bx = 128; // 4 warps per block.
+
+    // Logical size of launch = writeSigns ? p.s : p.x
+    uint32_t gx = writeSigns ? p.sShape.x : p.xShape.x;
+    uint32_t gy = writeSigns ? p.sShape.y : p.xShape.y;
+    uint32_t gz = p.xShape.z * p.xShape.w; // Same as in p.sShape if signs are in use.
+    gx = (gx - 1) / bx + 1;
+
+    // Make sure grid y and z dimensions are within CUDA launch limits. Kernel loops internally to do the rest.
+    const uint32_t gmax = 65535;
+    gy = std::min(gy, gmax);
+    gz = std::min(gz, gmax);
+
+    // Launch.
+    AT_CUDA_CHECK(cudaLaunchKernel(func, dim3(gx, gy, gz), bx, args, 0, at::cuda::getCurrentCUDAStream()));
+    return so;
+}
+
+//------------------------------------------------------------------------
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m)
+{
+    m.def("filtered_lrelu",      &filtered_lrelu);      // The whole thing.
+    m.def("filtered_lrelu_act_", &filtered_lrelu_act);  // Activation and sign tensor handling only. Modifies data tensor in-place.
+}
+
+//------------------------------------------------------------------------

models/torch_utils/ops/filtered_lrelu.cu

@@ -0,0 +1,1284 @@
+// Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#include <c10/util/Half.h>
+#include "filtered_lrelu.h"
+#include <cstdint>
+
+//------------------------------------------------------------------------
+// Helpers.
+
+enum // Filter modes.
+{
+    MODE_SUSD = 0,  // Separable upsampling, separable downsampling.
+    MODE_FUSD = 1,  // Full upsampling, separable downsampling.
+    MODE_SUFD = 2,  // Separable upsampling, full downsampling.
+    MODE_FUFD = 3,  // Full upsampling, full downsampling.
+};
+
+template <class T> struct InternalType;
+template <> struct InternalType<double>
+{
+    typedef double scalar_t; typedef double2 vec2_t; typedef double4 vec4_t;
+    __device__ __forceinline__ static vec2_t zero_vec2(void) { return make_double2(0, 0); }
+    __device__ __forceinline__ static vec4_t zero_vec4(void) { return make_double4(0, 0, 0, 0); }
+    __device__ __forceinline__ static double clamp(double x, double c) { return fmin(fmax(x, -c), c); }
+};
+template <> struct InternalType<float>
+{
+    typedef float scalar_t; typedef float2 vec2_t; typedef float4 vec4_t;
+    __device__ __forceinline__ static vec2_t zero_vec2(void) { return make_float2(0, 0); }
+    __device__ __forceinline__ static vec4_t zero_vec4(void) { return make_float4(0, 0, 0, 0); }
+    __device__ __forceinline__ static float clamp(float x, float c) { return fminf(fmaxf(x, -c), c); }
+};
+template <> struct InternalType<c10::Half>
+{
+    typedef float scalar_t; typedef float2 vec2_t; typedef float4 vec4_t;
+    __device__ __forceinline__ static vec2_t zero_vec2(void) { return make_float2(0, 0); }
+    __device__ __forceinline__ static vec4_t zero_vec4(void) { return make_float4(0, 0, 0, 0); }
+    __device__ __forceinline__ static float clamp(float x, float c) { return fminf(fmaxf(x, -c), c); }
+};
+
+#define MIN(A, B)       ((A) < (B) ? (A) : (B))
+#define MAX(A, B)       ((A) > (B) ? (A) : (B))
+#define CEIL_DIV(A, B) (((B)==1) ? (A) : \
+                        ((B)==2) ? ((int)((A)+1) >> 1) : \
+                        ((B)==4) ? ((int)((A)+3) >> 2) : \
+                        (((A) + ((A) > 0 ? (B) - 1 : 0)) / (B)))
+
+// This works only up to blocks of size 256 x 256 and for all N that are powers of two.
+template <int N> __device__ __forceinline__ void fast_div_mod(int& x, int& y, unsigned int i)
+{
+    if ((N & (N-1)) && N <= 256)
+        y = (i * ((1<<24)/N + 1)) >> 24; // Assumes N <= 256, i < N*256.
+    else
+        y = i/N;
+
+    x = i - y*N;
+}
+
+// Type cast stride before reading it.
+template <class T> __device__ __forceinline__ T get_stride(const int64_t& x)
+{
+    return *reinterpret_cast<const T*>(&x);
+}
+
+//------------------------------------------------------------------------
+// Filters, setup kernel, copying function.
+
+#define MAX_FILTER_SIZE 32
+
+// Combined up/down filter buffers so that transfer can be done with one copy.
+__device__              float g_fbuf[2 * MAX_FILTER_SIZE * MAX_FILTER_SIZE]; // Filters in global memory, written by setup kernel.
+__device__ __constant__ float c_fbuf[2 * MAX_FILTER_SIZE * MAX_FILTER_SIZE]; // Filters in constant memory, read by main kernel.
+
+// Accessors to combined buffers to index up/down filters individually.
+#define c_fu (c_fbuf)
+#define c_fd (c_fbuf + MAX_FILTER_SIZE * MAX_FILTER_SIZE)
+#define g_fu (g_fbuf)
+#define g_fd (g_fbuf + MAX_FILTER_SIZE * MAX_FILTER_SIZE)
+
+// Set up filters into global memory buffer.
+static __global__ void setup_filters_kernel(filtered_lrelu_kernel_params p)
+{
+    for (int idx = threadIdx.x; idx < MAX_FILTER_SIZE * MAX_FILTER_SIZE; idx += blockDim.x)
+    {
+        int x, y;
+        fast_div_mod<MAX_FILTER_SIZE>(x, y, idx);
+
+        int fu_x = p.flip ? x : (p.fuShape.x - 1 - x);
+        int fu_y = p.flip ? y : (p.fuShape.y - 1 - y);
+        if (p.fuShape.y > 0)
+            g_fu[idx] = (x >= p.fuShape.x || y >= p.fuShape.y) ? 0.0f : p.fu[fu_x * p.fuStride.x + fu_y * p.fuStride.y];
+        else
+            g_fu[idx] = (x >= p.fuShape.x || y > 0) ? 0.0f : p.fu[fu_x * p.fuStride.x];
+
+        int fd_x = p.flip ? x : (p.fdShape.x - 1 - x);
+        int fd_y = p.flip ? y : (p.fdShape.y - 1 - y);
+        if (p.fdShape.y > 0)
+            g_fd[idx] = (x >= p.fdShape.x || y >= p.fdShape.y) ? 0.0f : p.fd[fd_x * p.fdStride.x + fd_y * p.fdStride.y];
+        else
+            g_fd[idx] = (x >= p.fdShape.x || y > 0) ? 0.0f : p.fd[fd_x * p.fdStride.x];
+    }
+}
+
+// Host function to copy filters written by setup kernel into constant buffer for main kernel.
+template <bool, bool> static cudaError_t copy_filters(cudaStream_t stream)
+{
+    void* src = 0;
+    cudaError_t err = cudaGetSymbolAddress(&src, g_fbuf);
+    if (err) return err;
+    return cudaMemcpyToSymbolAsync(c_fbuf, src, 2 * MAX_FILTER_SIZE * MAX_FILTER_SIZE * sizeof(float), 0, cudaMemcpyDeviceToDevice, stream);
+}
+
+//------------------------------------------------------------------------
+// Coordinate spaces:
+// - Relative to input tensor:      inX, inY, tileInX, tileInY
+// - Relative to input tile:        relInX, relInY, tileInW, tileInH
+// - Relative to upsampled tile:    relUpX, relUpY, tileUpW, tileUpH
+// - Relative to output tile:       relOutX, relOutY, tileOutW, tileOutH
+// - Relative to output tensor:     outX, outY, tileOutX, tileOutY
+//
+// Relationships between coordinate spaces:
+// - inX = tileInX + relInX
+// - inY = tileInY + relInY
+// - relUpX = relInX * up + phaseInX
+// - relUpY = relInY * up + phaseInY
+// - relUpX = relOutX * down
+// - relUpY = relOutY * down
+// - outX = tileOutX + relOutX
+// - outY = tileOutY + relOutY
+
+extern __shared__ char s_buf_raw[]; // When sharedKB <= 48, allocate shared memory statically inside the kernel, otherwise use the externally allocated shared memory buffer.
+
+template <class T, class index_t, int sharedKB, bool signWrite, bool signRead, int filterMode, int up, int fuSize, int down, int fdSize, int tileOutW, int tileOutH, int threadsPerBlock, bool enableXrep, bool enableWriteSkip>
+static __global__ void filtered_lrelu_kernel(filtered_lrelu_kernel_params p)
+{
+    // Check that we don't try to support non-existing filter modes.
+    static_assert(up   == 1 || up   == 2 || up   == 4, "only up=1, up=2, up=4 scales supported");
+    static_assert(down == 1 || down == 2 || down == 4, "only down=1, down=2, down=4 scales supported");
+    static_assert(fuSize >= up,   "upsampling filter size must be at least upsampling factor");
+    static_assert(fdSize >= down, "downsampling filter size must be at least downsampling factor");
+    static_assert(fuSize % up   == 0, "upsampling filter size must be divisible with upsampling factor");
+    static_assert(fdSize % down == 0, "downsampling filter size must be divisible with downsampling factor");
+    static_assert(fuSize <= MAX_FILTER_SIZE && fdSize <= MAX_FILTER_SIZE, "filter size greater than MAX_FILTER_SIZE");
+    static_assert(up   != 1 || (fuSize == 1 && (filterMode == MODE_FUFD || filterMode == MODE_FUSD)), "up=1 supported only for 1x1 full filters");
+    static_assert(down != 1 || (fdSize == 1 && (filterMode == MODE_FUFD || filterMode == MODE_SUFD)), "down=1 supported only for 1x1 full filters");
+    static_assert(!(up   == 4 && (filterMode == MODE_FUFD || filterMode == MODE_FUSD)), "full filters not supported for up=4");
+    static_assert(!(down == 4 && (filterMode == MODE_FUFD || filterMode == MODE_SUFD)), "full filters not supported for down=4");
+
+    // Static definitions.
+    typedef typename InternalType<T>::scalar_t scalar_t;
+    typedef typename InternalType<T>::vec2_t vec2_t;
+    typedef typename InternalType<T>::vec4_t vec4_t;
+    const int tileUpW    = (tileOutW * down + (fdSize - 1) - (down - 1) + 3) & ~3;  // Upsampled tile width, rounded up to multiple of 4.
+    const int tileUpH    = tileOutH * down + (fdSize - 1) - (down - 1);             // Upsampled tile height.
+    const int tileInW    = CEIL_DIV(tileUpW  + (fuSize - 1), up);                   // Input tile width.
+    const int tileInH    = CEIL_DIV(tileUpH  + (fuSize - 1), up);                   // Input tile height.
+    const int tileUpH_up = CEIL_DIV(tileUpH, up) * up;                              // Upsampled tile height rounded up to a multiple of up.
+    const int tileInH_up = CEIL_DIV(tileUpH_up + (fuSize - 1), up);                 // For allocations only, to avoid shared memory read overruns with up=2 and up=4.
+
+    // Merge 1x1 downsampling into last upsampling step for upf1 and ups2.
+    const bool downInline = (down == 1) && ((up == 1 && filterMode == MODE_FUFD) || (up == 2 && filterMode == MODE_SUFD));
+
+    // Sizes of logical buffers.
+    const int szIn    = tileInH_up * tileInW;
+    const int szUpX   = tileInH_up * tileUpW;
+    const int szUpXY  = downInline ? 0 : (tileUpH * tileUpW);
+    const int szDownX = tileUpH * tileOutW;
+
+    // Sizes for shared memory arrays.
+    const int s_buf0_size_base =
+        (filterMode == MODE_SUSD) ? MAX(szIn, szUpXY) :
+        (filterMode == MODE_FUSD) ? MAX(szIn, szDownX) :
+        (filterMode == MODE_SUFD) ? MAX(szIn, szUpXY) :
+        (filterMode == MODE_FUFD) ? szIn :
+        -1;
+    const int s_buf1_size_base =
+        (filterMode == MODE_SUSD) ? MAX(szUpX, szDownX) :
+        (filterMode == MODE_FUSD) ? szUpXY :
+        (filterMode == MODE_SUFD) ? szUpX  :
+        (filterMode == MODE_FUFD) ? szUpXY :
+        -1;
+
+    // Ensure U128 alignment.
+    const int s_buf0_size = (s_buf0_size_base + 3) & ~3;
+    const int s_buf1_size = (s_buf1_size_base + 3) & ~3;
+
+    // Check at compile time that we don't use too much shared memory.
+    static_assert((s_buf0_size + s_buf1_size) * sizeof(scalar_t) <= (sharedKB << 10), "shared memory overflow");
+
+    // Declare shared memory arrays.
+    scalar_t* s_buf0;
+    scalar_t* s_buf1;
+    if (sharedKB <= 48)
+    {
+        // Allocate shared memory arrays here.
+        __shared__ scalar_t s_buf0_st[(sharedKB > 48) ? (1<<24) : (s_buf0_size + s_buf1_size)]; // Prevent launching if this isn't optimized away when unused.
+        s_buf0 = s_buf0_st;
+        s_buf1 = s_buf0 + s_buf0_size;
+    }
+    else
+    {
+        // Use the dynamically allocated shared memory array.
+        s_buf0 = (scalar_t*)s_buf_raw;
+        s_buf1 = s_buf0 + s_buf0_size;
+    }
+
+    // Pointers to the buffers.
+    scalar_t* s_tileIn;       // Input tile:                      [relInX * tileInH + relInY]
+    scalar_t* s_tileUpX;      // After horizontal upsampling:     [relInY * tileUpW + relUpX]
+    scalar_t* s_tileUpXY;     // After upsampling:                [relUpY * tileUpW + relUpX]
+    scalar_t* s_tileDownX;    // After horizontal downsampling:   [relUpY * tileOutW + relOutX]
+    if (filterMode == MODE_SUSD)
+    {
+        s_tileIn    = s_buf0;
+        s_tileUpX   = s_buf1;
+        s_tileUpXY  = s_buf0;
+        s_tileDownX = s_buf1;
+    }
+    else if (filterMode == MODE_FUSD)
+    {
+        s_tileIn    = s_buf0;
+        s_tileUpXY  = s_buf1;
+        s_tileDownX = s_buf0;
+    }
+    else if (filterMode == MODE_SUFD)
+    {
+        s_tileIn    = s_buf0;
+        s_tileUpX   = s_buf1;
+        s_tileUpXY  = s_buf0;
+    }
+    else if (filterMode == MODE_FUFD)
+    {
+        s_tileIn    = s_buf0;
+        s_tileUpXY  = s_buf1;
+    }
+
+    // Allow large grids in z direction via per-launch offset.
+    int channelIdx = blockIdx.z + p.blockZofs;
+    int batchIdx = channelIdx / p.yShape.z;
+    channelIdx -= batchIdx * p.yShape.z;
+
+    // Offset to output feature map. In bytes.
+    index_t mapOfsOut = channelIdx * get_stride<index_t>(p.yStride.z) + batchIdx * get_stride<index_t>(p.yStride.w);
+
+    // Sign shift amount.
+    uint32_t signXo = ((threadIdx.x + p.sOfs.x) << 1) & 6;
+
+    // Inner tile loop.
+    #pragma unroll 1
+    for (int tileIdx = 0; !enableXrep || (tileIdx < MIN(p.tilesXrep, p.tilesXdim - p.tilesXrep * blockIdx.y)); tileIdx++)
+    {
+        // Locate output tile.
+        int tileX = enableXrep ? blockIdx.y * p.tilesXrep + tileIdx : blockIdx.x;
+        int tileOutX = tileX * tileOutW;
+        int tileOutY = (enableXrep ? blockIdx.x : blockIdx.y) * tileOutH;
+
+        // Locate input tile.
+        int tmpX = tileOutX * down - p.pad0.x;
+        int tmpY = tileOutY * down - p.pad0.y;
+        int tileInX = CEIL_DIV(tmpX, up);
+        int tileInY = CEIL_DIV(tmpY, up);
+        const int phaseInX = tileInX * up - tmpX;
+        const int phaseInY = tileInY * up - tmpY;
+
+        // Extra sync if input and output buffers are the same and we are not on first tile.
+        if (enableXrep && tileIdx > 0 && (filterMode == MODE_FUSD || (filterMode == MODE_SUFD && !downInline) || (filterMode == MODE_FUFD && downInline)))
+            __syncthreads();
+
+        // Load input tile & apply bias. Unrolled.
+        scalar_t b = (scalar_t)*(const T*)((const char*)p.b + (channelIdx * get_stride<index_t>(p.bStride)));
+        index_t mapOfsIn = channelIdx * get_stride<index_t>(p.xStride.z) + batchIdx * get_stride<index_t>(p.xStride.w);
+        int idx = threadIdx.x;
+        const int loopCountIN = CEIL_DIV(tileInW * tileInH, threadsPerBlock);
+        #pragma unroll
+        for (int loop = 0; loop < loopCountIN; loop++)
+        {
+            int relInX, relInY;
+            fast_div_mod<tileInW>(relInX, relInY, idx);
+            int inX = tileInX + relInX;
+            int inY = tileInY + relInY;
+            scalar_t v = 0;
+
+            if ((uint32_t)inX < p.xShape.x && (uint32_t)inY < p.xShape.y)
+                v = (scalar_t)*((const T*)((const char*)p.x + (inX * get_stride<index_t>(p.xStride.x) + inY * get_stride<index_t>(p.xStride.y) + mapOfsIn))) + b;
+
+            bool skip = (loop == loopCountIN-1) && (idx >= tileInW * tileInH);
+            if (!skip)
+                s_tileIn[idx] = v;
+
+            idx += threadsPerBlock;
+        }
+
+        if (filterMode == MODE_SUSD || filterMode == MODE_SUFD) // Separable upsampling filter.
+        {
+            // Horizontal upsampling.
+            __syncthreads();
+            if (up == 4)
+            {
+                for (int idx = threadIdx.x*up; idx < tileUpW * tileInH; idx += blockDim.x*up)
+                {
+                    int relUpX0, relInY;
+                    fast_div_mod<tileUpW>(relUpX0, relInY, idx);
+                    int relInX0 = relUpX0 / up;
+                    int src0 = relInX0 + tileInW * relInY;
+                    int dst = relInY * tileUpW + relUpX0;
+                    vec4_t v = InternalType<T>::zero_vec4();
+                    scalar_t a = s_tileIn[src0];
+                    if (phaseInX == 0)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileIn[src0 + step + 1];
+                            v.y += a * (scalar_t)c_fu[step * up + 3];
+                            v.z += a * (scalar_t)c_fu[step * up + 2];
+                            v.w += a * (scalar_t)c_fu[step * up + 1];
+                        }
+                    }
+                    else if (phaseInX == 1)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 1];
+                            v.y += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileIn[src0 + step + 1];
+                            v.z += a * (scalar_t)c_fu[step * up + 3];
+                            v.w += a * (scalar_t)c_fu[step * up + 2];
+                        }
+                    }
+                    else if (phaseInX == 2)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 2];
+                            v.y += a * (scalar_t)c_fu[step * up + 1];
+                            v.z += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileIn[src0 + step + 1];
+                            v.w += a * (scalar_t)c_fu[step * up + 3];
+                        }
+                    }
+                    else // (phaseInX == 3)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 3];
+                            v.y += a * (scalar_t)c_fu[step * up + 2];
+                            v.z += a * (scalar_t)c_fu[step * up + 1];
+                            v.w += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileIn[src0 + step + 1];
+                        }
+                    }
+                    s_tileUpX[dst+0] = v.x;
+                    s_tileUpX[dst+1] = v.y;
+                    s_tileUpX[dst+2] = v.z;
+                    s_tileUpX[dst+3] = v.w;
+                }
+            }
+            else if (up == 2)
+            {
+                bool p0 = (phaseInX == 0);
+                for (int idx = threadIdx.x*up; idx < tileUpW * tileInH; idx += blockDim.x*up)
+                {
+                    int relUpX0, relInY;
+                    fast_div_mod<tileUpW>(relUpX0, relInY, idx);
+                    int relInX0 = relUpX0 / up;
+                    int src0 = relInX0 + tileInW * relInY;
+                    int dst = relInY * tileUpW + relUpX0;
+                    vec2_t v = InternalType<T>::zero_vec2();
+                    scalar_t a = s_tileIn[src0];
+                    if (p0) // (phaseInX == 0)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileIn[src0 + step + 1];
+                            v.y += a * (scalar_t)c_fu[step * up + 1];
+                        }
+                    }
+                    else // (phaseInX == 1)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 1];
+                            v.y += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileIn[src0 + step + 1];
+                        }
+                    }
+                    s_tileUpX[dst+0] = v.x;
+                    s_tileUpX[dst+1] = v.y;
+                }
+            }
+
+            // Vertical upsampling & nonlinearity.
+
+            __syncthreads();
+            int groupMask = 15 << ((threadIdx.x & 31) & ~3);
+            int minY = tileOutY ? (tileOutY - tileOutH) * down + tileUpH : 0; // Skip already written signs.
+            int sShapeMaxY = MIN(p.sShape.y, tileOutY * down + tileUpH); // Avoid out-of-tile sign writes.
+            if (up == 4)
+            {
+                minY -= 3; // Adjust according to block height.
+                for (int idx = threadIdx.x; idx < tileUpW * tileUpH_up / up; idx += blockDim.x)
+                {
+                    int relUpX, relInY0;
+                    fast_div_mod<tileUpW>(relUpX, relInY0, idx);
+                    int relUpY0 = relInY0 * up;
+                    int src0 = relInY0 * tileUpW + relUpX;
+                    int dst = relUpY0 * tileUpW + relUpX;
+                    vec4_t v = InternalType<T>::zero_vec4();
+
+                    scalar_t a = s_tileUpX[src0];
+                    if (phaseInY == 0)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileUpX[src0 + (step + 1) * tileUpW];
+                            v.y += a * (scalar_t)c_fu[step * up + 3];
+                            v.z += a * (scalar_t)c_fu[step * up + 2];
+                            v.w += a * (scalar_t)c_fu[step * up + 1];
+                        }
+                    }
+                    else if (phaseInY == 1)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 1];
+                            v.y += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileUpX[src0 + (step + 1) * tileUpW];
+                            v.z += a * (scalar_t)c_fu[step * up + 3];
+                            v.w += a * (scalar_t)c_fu[step * up + 2];
+                        }
+                    }
+                    else if (phaseInY == 2)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 2];
+                            v.y += a * (scalar_t)c_fu[step * up + 1];
+                            v.z += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileUpX[src0 + (step + 1) * tileUpW];
+                            v.w += a * (scalar_t)c_fu[step * up + 3];
+                        }
+                    }
+                    else // (phaseInY == 3)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 3];
+                            v.y += a * (scalar_t)c_fu[step * up + 2];
+                            v.z += a * (scalar_t)c_fu[step * up + 1];
+                            v.w += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileUpX[src0 + (step + 1) * tileUpW];
+                        }
+                    }
+
+                    int x = tileOutX * down + relUpX;
+                    int y = tileOutY * down + relUpY0;
+                    int signX = x + p.sOfs.x;
+                    int signY = y + p.sOfs.y;
+                    int signZ = blockIdx.z + p.blockZofs;
+                    int signXb = signX >> 2;
+                    index_t si0 = signXb + p.sShape.x * (signY + (index_t)p.sShape.y * signZ);
+                    index_t si1 = si0 + p.sShape.x;
+                    index_t si2 = si0 + p.sShape.x * 2;
+                    index_t si3 = si0 + p.sShape.x * 3;
+
+                    v.x *= (scalar_t)((float)up * (float)up * p.gain);
+                    v.y *= (scalar_t)((float)up * (float)up * p.gain);
+                    v.z *= (scalar_t)((float)up * (float)up * p.gain);
+                    v.w *= (scalar_t)((float)up * (float)up * p.gain);
+
+                    if (signWrite)
+                    {
+                        if (!enableWriteSkip)
+                        {
+                            // Determine and write signs.
+                            int sx = __float_as_uint(v.x) >> 31 <<  0;
+                            int sy = __float_as_uint(v.y) >> 31 <<  8;
+                            int sz = __float_as_uint(v.z) >> 31 << 16;
+                            int sw = __float_as_uint(v.w) >> 31 << 24;
+                            if (sx) v.x *= p.slope;
+                            if (sy) v.y *= p.slope;
+                            if (sz) v.z *= p.slope;
+                            if (sw) v.w *= p.slope;
+                            if (fabsf(v.x) > p.clamp) { sx = 2 <<  0; v.x = InternalType<T>::clamp(v.x, p.clamp); }
+                            if (fabsf(v.y) > p.clamp) { sy = 2 <<  8; v.y = InternalType<T>::clamp(v.y, p.clamp); }
+                            if (fabsf(v.z) > p.clamp) { sz = 2 << 16; v.z = InternalType<T>::clamp(v.z, p.clamp); }
+                            if (fabsf(v.w) > p.clamp) { sw = 2 << 24; v.w = InternalType<T>::clamp(v.w, p.clamp); }
+
+                            if ((uint32_t)signXb < p.swLimit && signY >= minY)
+                            {
+                                // Combine signs.
+                                uint32_t s = sx + sy + sw + sz;
+                                s <<= (signX & 3) << 1;
+                                s |= __shfl_xor_sync(groupMask, s, 1);
+                                s |= __shfl_xor_sync(groupMask, s, 2);
+
+                                // Write signs.
+                                if ((uint32_t)(signY + 0) < sShapeMaxY) { p.s[si0] = (unsigned char)(s >>  0); }
+                                if ((uint32_t)(signY + 1) < sShapeMaxY) { p.s[si1] = (unsigned char)(s >>  8); }
+                                if ((uint32_t)(signY + 2) < sShapeMaxY) { p.s[si2] = (unsigned char)(s >> 16); }
+                                if ((uint32_t)(signY + 3) < sShapeMaxY) { p.s[si3] = (unsigned char)(s >> 24); }
+                            }
+                        }
+                        else
+                        {
+                            // Determine and write signs.
+                            if ((uint32_t)signXb < p.swLimit && signY >= minY)
+                            {
+                                int sx = __float_as_uint(v.x) >> 31 <<  0;
+                                int sy = __float_as_uint(v.y) >> 31 <<  8;
+                                int sz = __float_as_uint(v.z) >> 31 << 16;
+                                int sw = __float_as_uint(v.w) >> 31 << 24;
+                                if (sx) v.x *= p.slope;
+                                if (sy) v.y *= p.slope;
+                                if (sz) v.z *= p.slope;
+                                if (sw) v.w *= p.slope;
+                                if (fabsf(v.x) > p.clamp) { sx = 2 <<  0; v.x = InternalType<T>::clamp(v.x, p.clamp); }
+                                if (fabsf(v.y) > p.clamp) { sy = 2 <<  8; v.y = InternalType<T>::clamp(v.y, p.clamp); }
+                                if (fabsf(v.z) > p.clamp) { sz = 2 << 16; v.z = InternalType<T>::clamp(v.z, p.clamp); }
+                                if (fabsf(v.w) > p.clamp) { sw = 2 << 24; v.w = InternalType<T>::clamp(v.w, p.clamp); }
+
+                                // Combine signs.
+                                uint32_t s = sx + sy + sw + sz;
+                                s <<= (signX & 3) << 1;
+                                s |= __shfl_xor_sync(groupMask, s, 1);
+                                s |= __shfl_xor_sync(groupMask, s, 2);
+
+                                // Write signs.
+                                if ((uint32_t)(signY + 0) < sShapeMaxY) { p.s[si0] = (unsigned char)(s >>  0); }
+                                if ((uint32_t)(signY + 1) < sShapeMaxY) { p.s[si1] = (unsigned char)(s >>  8); }
+                                if ((uint32_t)(signY + 2) < sShapeMaxY) { p.s[si2] = (unsigned char)(s >> 16); }
+                                if ((uint32_t)(signY + 3) < sShapeMaxY) { p.s[si3] = (unsigned char)(s >> 24); }
+                            }
+                            else
+                            {
+                                // Just compute the values.
+                                if (v.x < 0.f) v.x *= p.slope; v.x = InternalType<T>::clamp(v.x, p.clamp);
+                                if (v.y < 0.f) v.y *= p.slope; v.y = InternalType<T>::clamp(v.y, p.clamp);
+                                if (v.z < 0.f) v.z *= p.slope; v.z = InternalType<T>::clamp(v.z, p.clamp);
+                                if (v.w < 0.f) v.w *= p.slope; v.w = InternalType<T>::clamp(v.w, p.clamp);
+                            }
+                        }
+                    }
+                    else if (signRead) // Read signs and apply.
+                    {
+                        if ((uint32_t)signXb < p.swLimit)
+                        {
+                            int ss = (signX & 3) << 1;
+                            if ((uint32_t)(signY + 0) < p.sShape.y) { int s = p.s[si0] >> ss; if (s & 1) v.x *= p.slope; if (s & 2) v.x = 0.f; }
+                            if ((uint32_t)(signY + 1) < p.sShape.y) { int s = p.s[si1] >> ss; if (s & 1) v.y *= p.slope; if (s & 2) v.y = 0.f; }
+                            if ((uint32_t)(signY + 2) < p.sShape.y) { int s = p.s[si2] >> ss; if (s & 1) v.z *= p.slope; if (s & 2) v.z = 0.f; }
+                            if ((uint32_t)(signY + 3) < p.sShape.y) { int s = p.s[si3] >> ss; if (s & 1) v.w *= p.slope; if (s & 2) v.w = 0.f; }
+                        }
+                    }
+                    else // Forward pass with no sign write.
+                    {
+                        if (v.x < 0.f) v.x *= p.slope; v.x = InternalType<T>::clamp(v.x, p.clamp);
+                        if (v.y < 0.f) v.y *= p.slope; v.y = InternalType<T>::clamp(v.y, p.clamp);
+                        if (v.z < 0.f) v.z *= p.slope; v.z = InternalType<T>::clamp(v.z, p.clamp);
+                        if (v.w < 0.f) v.w *= p.slope; v.w = InternalType<T>::clamp(v.w, p.clamp);
+                    }
+
+                    s_tileUpXY[dst + 0 * tileUpW] = v.x;
+                    if (relUpY0 + 1 < tileUpH) s_tileUpXY[dst + 1 * tileUpW] = v.y;
+                    if (relUpY0 + 2 < tileUpH) s_tileUpXY[dst + 2 * tileUpW] = v.z;
+                    if (relUpY0 + 3 < tileUpH) s_tileUpXY[dst + 3 * tileUpW] = v.w;
+                }
+            }
+            else if (up == 2)
+            {
+                minY -= 1; // Adjust according to block height.
+                for (int idx = threadIdx.x; idx < tileUpW * tileUpH_up / up; idx += blockDim.x)
+                {
+                    int relUpX, relInY0;
+                    fast_div_mod<tileUpW>(relUpX, relInY0, idx);
+                    int relUpY0 = relInY0 * up;
+                    int src0 = relInY0 * tileUpW + relUpX;
+                    int dst = relUpY0 * tileUpW + relUpX;
+                    vec2_t v = InternalType<T>::zero_vec2();
+
+                    scalar_t a = s_tileUpX[src0];
+                    if (phaseInY == 0)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileUpX[src0 + (step + 1) * tileUpW];
+                            v.y += a * (scalar_t)c_fu[step * up + 1];
+                        }
+                    }
+                    else // (phaseInY == 1)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 1];
+                            v.y += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileUpX[src0 + (step + 1) * tileUpW];
+                        }
+                    }
+
+                    int x = tileOutX * down + relUpX;
+                    int y = tileOutY * down + relUpY0;
+                    int signX = x + p.sOfs.x;
+                    int signY = y + p.sOfs.y;
+                    int signZ = blockIdx.z + p.blockZofs;
+                    int signXb = signX >> 2;
+                    index_t si0 = signXb + p.sShape.x * (signY + (index_t)p.sShape.y * signZ);
+                    index_t si1 = si0 + p.sShape.x;
+
+                    v.x *= (scalar_t)((float)up * (float)up * p.gain);
+                    v.y *= (scalar_t)((float)up * (float)up * p.gain);
+
+                    if (signWrite)
+                    {
+                        if (!enableWriteSkip)
+                        {
+                            // Determine and write signs.
+                            int sx = __float_as_uint(v.x) >> 31 << 0;
+                            int sy = __float_as_uint(v.y) >> 31 << 8;
+                            if (sx) v.x *= p.slope;
+                            if (sy) v.y *= p.slope;
+                            if (fabsf(v.x) > p.clamp) { sx = 2 << 0; v.x = InternalType<T>::clamp(v.x, p.clamp); }
+                            if (fabsf(v.y) > p.clamp) { sy = 2 << 8; v.y = InternalType<T>::clamp(v.y, p.clamp); }
+
+                            if ((uint32_t)signXb < p.swLimit && signY >= minY)
+                            {
+                                // Combine signs.
+                                int s = sx + sy;
+                                s <<= signXo;
+                                s |= __shfl_xor_sync(groupMask, s, 1);
+                                s |= __shfl_xor_sync(groupMask, s, 2);
+
+                                // Write signs.
+                                if ((uint32_t)(signY + 0) < sShapeMaxY) { p.s[si0] = (unsigned char)(s >>  0); }
+                                if ((uint32_t)(signY + 1) < sShapeMaxY) { p.s[si1] = (unsigned char)(s >>  8); }
+                            }
+                        }
+                        else
+                        {
+                            // Determine and write signs.
+                            if ((uint32_t)signXb < p.swLimit && signY >= minY)
+                            {
+                                int sx = __float_as_uint(v.x) >> 31 << 0;
+                                int sy = __float_as_uint(v.y) >> 31 << 8;
+                                if (sx) v.x *= p.slope;
+                                if (sy) v.y *= p.slope;
+                                if (fabsf(v.x) > p.clamp) { sx = 2 << 0; v.x = InternalType<T>::clamp(v.x, p.clamp); }
+                                if (fabsf(v.y) > p.clamp) { sy = 2 << 8; v.y = InternalType<T>::clamp(v.y, p.clamp); }
+
+                                // Combine signs.
+                                int s = sx + sy;
+                                s <<= signXo;
+                                s |= __shfl_xor_sync(groupMask, s, 1);
+                                s |= __shfl_xor_sync(groupMask, s, 2);
+
+                                // Write signs.
+                                if ((uint32_t)(signY + 0) < sShapeMaxY) { p.s[si0] = (unsigned char)(s >>  0); }
+                                if ((uint32_t)(signY + 1) < sShapeMaxY) { p.s[si1] = (unsigned char)(s >>  8); }
+                            }
+                            else
+                            {
+                                // Just compute the values.
+                                if (v.x < 0.f) v.x *= p.slope; v.x = InternalType<T>::clamp(v.x, p.clamp);
+                                if (v.y < 0.f) v.y *= p.slope; v.y = InternalType<T>::clamp(v.y, p.clamp);
+                            }
+                        }
+                    }
+                    else if (signRead) // Read signs and apply.
+                    {
+                        if ((uint32_t)signXb < p.swLimit)
+                        {
+                            if ((uint32_t)(signY + 0) < p.sShape.y) { int s = p.s[si0] >> signXo; if (s & 1) v.x *= p.slope; if (s & 2) v.x = 0.f; }
+                            if ((uint32_t)(signY + 1) < p.sShape.y) { int s = p.s[si1] >> signXo; if (s & 1) v.y *= p.slope; if (s & 2) v.y = 0.f; }
+                        }
+                    }
+                    else // Forward pass with no sign write.
+                    {
+                        if (v.x < 0.f) v.x *= p.slope; v.x = InternalType<T>::clamp(v.x, p.clamp);
+                        if (v.y < 0.f) v.y *= p.slope; v.y = InternalType<T>::clamp(v.y, p.clamp);
+                    }
+
+                    if (!downInline)
+                    {
+                        // Write into temporary buffer.
+                        s_tileUpXY[dst] = v.x;
+                        if (relUpY0 < tileUpH - 1)
+                            s_tileUpXY[dst + tileUpW] = v.y;
+                    }
+                    else
+                    {
+                        // Write directly into output buffer.
+                        if ((uint32_t)x < p.yShape.x)
+                        {
+                            int ymax = MIN(p.yShape.y, tileUpH + tileOutY * down);
+                            index_t ofs = x * get_stride<index_t>(p.yStride.x) + y * get_stride<index_t>(p.yStride.y) + mapOfsOut;
+                            if ((uint32_t)y + 0 < p.yShape.y) *((T*)((char*)p.y + ofs)) = (T)(v.x * (scalar_t)c_fd[0]);
+                            if ((uint32_t)y + 1 < ymax) *((T*)((char*)p.y + ofs + get_stride<index_t>(p.yStride.y))) = (T)(v.y * (scalar_t)c_fd[0]);
+                        }
+                    }
+                }
+            }
+        }
+        else if (filterMode == MODE_FUSD || filterMode == MODE_FUFD)
+        {
+            // Full upsampling filter.
+
+            if (up == 2)
+            {
+                // 2 x 2-wide.
+                __syncthreads();
+                int minY = tileOutY ? (tileOutY - tileOutH) * down + tileUpH + p.sOfs.y : 0; // Skip already written signs.
+                for (int idx = threadIdx.x * 4; idx < tileUpW * tileUpH; idx += blockDim.x * 4)
+                {
+                    int relUpX0, relUpY0;
+                    fast_div_mod<tileUpW>(relUpX0, relUpY0, idx);
+                    int relInX0 = CEIL_DIV(relUpX0 - phaseInX, up);
+                    int relInY0 = CEIL_DIV(relUpY0 - phaseInY, up);
+                    int src0 = relInX0 + tileInW * relInY0;
+                    int tap0y = (relInY0 * up + phaseInY - relUpY0);
+
+                    #define X_LOOP(TAPY, PX) \
+                        for (int sx = 0; sx < fuSize / up; sx++) \
+                        { \
+                            v.x += a * (scalar_t)c_fu[(sx * up + (((PX) - 0) & (up - 1))) + (sy * up + (TAPY)) * MAX_FILTER_SIZE]; \
+                            v.z += b * (scalar_t)c_fu[(sx * up + (((PX) - 0) & (up - 1))) + (sy * up + (TAPY)) * MAX_FILTER_SIZE]; if ((PX) == 0) { a = b; b = s_tileIn[src0 + 2 + sx + sy * tileInW]; } \
+                            v.y += a * (scalar_t)c_fu[(sx * up + (((PX) - 1) & (up - 1))) + (sy * up + (TAPY)) * MAX_FILTER_SIZE]; \
+                            v.w += b * (scalar_t)c_fu[(sx * up + (((PX) - 1) & (up - 1))) + (sy * up + (TAPY)) * MAX_FILTER_SIZE]; if ((PX) == 1) { a = b; b = s_tileIn[src0 + 2 + sx + sy * tileInW]; } \
+                        }
+
+                    vec4_t v = InternalType<T>::zero_vec4();
+                    if (tap0y == 0 && phaseInX == 0)
+                        #pragma unroll
+                        for (int sy = 0; sy < fuSize / up; sy++) { scalar_t a = s_tileIn[src0 + sy * tileInW]; scalar_t b = s_tileIn[src0 + sy * tileInW + 1];
+                            #pragma unroll
+                            X_LOOP(0, 0) }
+                    if (tap0y == 0 && phaseInX == 1)
+                        #pragma unroll
+                        for (int sy = 0; sy < fuSize / up; sy++) { scalar_t a = s_tileIn[src0 + sy * tileInW]; scalar_t b = s_tileIn[src0 + sy * tileInW + 1];
+                            #pragma unroll
+                            X_LOOP(0, 1) }
+                    if (tap0y == 1 && phaseInX == 0)
+                        #pragma unroll
+                        for (int sy = 0; sy < fuSize / up; sy++) { scalar_t a = s_tileIn[src0 + sy * tileInW]; scalar_t b = s_tileIn[src0 + sy * tileInW + 1];
+                            #pragma unroll
+                            X_LOOP(1, 0) }
+                    if (tap0y == 1 && phaseInX == 1)
+                        #pragma unroll
+                        for (int sy = 0; sy < fuSize / up; sy++) { scalar_t a = s_tileIn[src0 + sy * tileInW]; scalar_t b = s_tileIn[src0 + sy * tileInW + 1];
+                            #pragma unroll
+                            X_LOOP(1, 1) }
+
+                    #undef X_LOOP
+
+                    int x = tileOutX * down + relUpX0;
+                    int y = tileOutY * down + relUpY0;
+                    int signX = x + p.sOfs.x;
+                    int signY = y + p.sOfs.y;
+                    int signZ = blockIdx.z + p.blockZofs;
+                    int signXb = signX >> 2;
+                    index_t si = signXb + p.sShape.x * (signY + (index_t)p.sShape.y * signZ);
+
+                    v.x *= (scalar_t)((float)up * (float)up * p.gain);
+                    v.y *= (scalar_t)((float)up * (float)up * p.gain);
+                    v.z *= (scalar_t)((float)up * (float)up * p.gain);
+                    v.w *= (scalar_t)((float)up * (float)up * p.gain);
+
+                    if (signWrite)
+                    {
+                        if (!enableWriteSkip)
+                        {
+                            // Determine and write signs.
+                            int sx = __float_as_uint(v.x) >> 31;
+                            int sy = __float_as_uint(v.y) >> 31;
+                            int sz = __float_as_uint(v.z) >> 31;
+                            int sw = __float_as_uint(v.w) >> 31;
+                            if (sx) v.x *= p.slope; if (fabsf(v.x) > p.clamp) { sx = 2; v.x = InternalType<T>::clamp(v.x, p.clamp); }
+                            if (sy) v.y *= p.slope; if (fabsf(v.y) > p.clamp) { sy = 2; v.y = InternalType<T>::clamp(v.y, p.clamp); }
+                            if (sz) v.z *= p.slope; if (fabsf(v.z) > p.clamp) { sz = 2; v.z = InternalType<T>::clamp(v.z, p.clamp); }
+                            if (sw) v.w *= p.slope; if (fabsf(v.w) > p.clamp) { sw = 2; v.w = InternalType<T>::clamp(v.w, p.clamp); }
+
+                            if ((uint32_t)signXb < p.swLimit && (uint32_t)signY < p.sShape.y && signY >= minY)
+                            {
+                                p.s[si] = sx + (sy << 2) + (sz << 4) + (sw << 6);
+                            }
+                        }
+                        else
+                        {
+                            // Determine and write signs.
+                            if ((uint32_t)signXb < p.swLimit && (uint32_t)signY < p.sShape.y && signY >= minY)
+                            {
+                                int sx = __float_as_uint(v.x) >> 31;
+                                int sy = __float_as_uint(v.y) >> 31;
+                                int sz = __float_as_uint(v.z) >> 31;
+                                int sw = __float_as_uint(v.w) >> 31;
+                                if (sx) v.x *= p.slope; if (fabsf(v.x) > p.clamp) { sx = 2; v.x = InternalType<T>::clamp(v.x, p.clamp); }
+                                if (sy) v.y *= p.slope; if (fabsf(v.y) > p.clamp) { sy = 2; v.y = InternalType<T>::clamp(v.y, p.clamp); }
+                                if (sz) v.z *= p.slope; if (fabsf(v.z) > p.clamp) { sz = 2; v.z = InternalType<T>::clamp(v.z, p.clamp); }
+                                if (sw) v.w *= p.slope; if (fabsf(v.w) > p.clamp) { sw = 2; v.w = InternalType<T>::clamp(v.w, p.clamp); }
+
+                                p.s[si] = sx + (sy << 2) + (sz << 4) + (sw << 6);
+                            }
+                            else
+                            {
+                                // Just compute the values.
+                                if (v.x < 0.f) v.x *= p.slope; v.x = InternalType<T>::clamp(v.x, p.clamp);
+                                if (v.y < 0.f) v.y *= p.slope; v.y = InternalType<T>::clamp(v.y, p.clamp);
+                                if (v.z < 0.f) v.z *= p.slope; v.z = InternalType<T>::clamp(v.z, p.clamp);
+                                if (v.w < 0.f) v.w *= p.slope; v.w = InternalType<T>::clamp(v.w, p.clamp);
+                            }
+                        }
+                    }
+                    else if (signRead) // Read sign and apply.
+                    {
+                        if ((uint32_t)signY < p.sShape.y)
+                        {
+                            int s = 0;
+                            if ((uint32_t)signXb     < p.swLimit) s  = p.s[si];
+                            if ((uint32_t)signXb + 1 < p.swLimit) s |= p.s[si + 1] << 8;
+                            s >>= (signX & 3) << 1;
+                            if (s & 0x01) v.x *= p.slope; if (s & 0x02) v.x = 0.f;
+                            if (s & 0x04) v.y *= p.slope; if (s & 0x08) v.y = 0.f;
+                            if (s & 0x10) v.z *= p.slope; if (s & 0x20) v.z = 0.f;
+                            if (s & 0x40) v.w *= p.slope; if (s & 0x80) v.w = 0.f;
+                        }
+                    }
+                    else // Forward pass with no sign write.
+                    {
+                        if (v.x < 0.f) v.x *= p.slope; v.x = InternalType<T>::clamp(v.x, p.clamp);
+                        if (v.y < 0.f) v.y *= p.slope; v.y = InternalType<T>::clamp(v.y, p.clamp);
+                        if (v.z < 0.f) v.z *= p.slope; v.z = InternalType<T>::clamp(v.z, p.clamp);
+                        if (v.w < 0.f) v.w *= p.slope; v.w = InternalType<T>::clamp(v.w, p.clamp);
+                    }
+
+                    s_tileUpXY[idx + 0] = v.x;
+                    s_tileUpXY[idx + 1] = v.y;
+                    s_tileUpXY[idx + 2] = v.z;
+                    s_tileUpXY[idx + 3] = v.w;
+                }
+            }
+            else if (up == 1)
+            {
+                __syncthreads();
+                uint32_t groupMask = 15 << ((threadIdx.x & 31) & ~3);
+                int minY = tileOutY ? (tileOutY - tileOutH) * down + tileUpH : 0; // Skip already written signs.
+                for (int idx = threadIdx.x; idx < tileUpW * tileUpH; idx += blockDim.x)
+                {
+                    int relUpX0, relUpY0;
+                    fast_div_mod<tileUpW>(relUpX0, relUpY0, idx);
+                    scalar_t v = s_tileIn[idx] * (scalar_t)c_fu[0]; // 1x1 filter.
+
+                    int x = tileOutX * down + relUpX0;
+                    int y = tileOutY * down + relUpY0;
+                    int signX = x + p.sOfs.x;
+                    int signY = y + p.sOfs.y;
+                    int signZ = blockIdx.z + p.blockZofs;
+                    int signXb = signX >> 2;
+                    index_t si = signXb + p.sShape.x * (signY + (index_t)p.sShape.y * signZ);
+                    v *= (scalar_t)((float)up * (float)up * p.gain);
+
+                    if (signWrite)
+                    {
+                        if (!enableWriteSkip)
+                        {
+                            // Determine and write sign.
+                            uint32_t s = 0;
+                            uint32_t signXbit = (1u << signXo);
+                            if (v < 0.f)
+                            {
+                                s = signXbit;
+                                v *= p.slope;
+                            }
+                            if (fabsf(v) > p.clamp)
+                            {
+                                s = signXbit * 2;
+                                v = InternalType<T>::clamp(v, p.clamp);
+                            }
+                            if ((uint32_t)signXb < p.swLimit && (uint32_t)signY < p.sShape.y && signY >= minY)
+                            {
+                                s += __shfl_xor_sync(groupMask, s, 1);  // Coalesce.
+                                s += __shfl_xor_sync(groupMask, s, 2);  // Coalesce.
+                                p.s[si] = s;                            // Write.
+                            }
+                        }
+                        else
+                        {
+                            // Determine and write sign.
+                            if ((uint32_t)signXb < p.swLimit && (uint32_t)signY < p.sShape.y && signY >= minY)
+                            {
+                                uint32_t s = 0;
+                                uint32_t signXbit = (1u << signXo);
+                                if (v < 0.f)
+                                {
+                                    s = signXbit;
+                                    v *= p.slope;
+                                }
+                                if (fabsf(v) > p.clamp)
+                                {
+                                    s = signXbit * 2;
+                                    v = InternalType<T>::clamp(v, p.clamp);
+                                }
+                                s += __shfl_xor_sync(groupMask, s, 1);  // Coalesce.
+                                s += __shfl_xor_sync(groupMask, s, 2);  // Coalesce.
+                                p.s[si] = s;                            // Write.
+                            }
+                            else
+                            {
+                                // Just compute the value.
+                                if (v < 0.f) v *= p.slope;
+                                v = InternalType<T>::clamp(v, p.clamp);
+                            }
+                        }
+                    }
+                    else if (signRead)
+                    {
+                        // Read sign and apply if within sign tensor bounds.
+                        if ((uint32_t)signXb < p.swLimit && (uint32_t)signY < p.sShape.y)
+                        {
+                            int s = p.s[si];
+                            s >>= signXo;
+                            if (s & 1) v *= p.slope;
+                            if (s & 2) v = 0.f;
+                        }
+                    }
+                    else // Forward pass with no sign write.
+                    {
+                        if (v < 0.f) v *= p.slope;
+                        v = InternalType<T>::clamp(v, p.clamp);
+                    }
+
+                    if (!downInline) // Write into temporary buffer.
+                        s_tileUpXY[idx] = v;
+                    else if ((uint32_t)x < p.yShape.x && (uint32_t)y < p.yShape.y) // Write directly into output buffer
+                        *((T*)((char*)p.y + (x * get_stride<index_t>(p.yStride.x) + y * get_stride<index_t>(p.yStride.y) + mapOfsOut))) = (T)(v * (scalar_t)c_fd[0]);
+                }
+            }
+        }
+
+        // Downsampling.
+        if (filterMode == MODE_SUSD || filterMode == MODE_FUSD)
+        {
+            // Horizontal downsampling.
+            __syncthreads();
+            if (down == 4 && tileOutW % 4 == 0)
+            {
+                // Calculate 4 pixels at a time.
+                for (int idx = threadIdx.x * 4; idx < tileOutW * tileUpH; idx += blockDim.x * 4)
+                {
+                    int relOutX0, relUpY;
+                    fast_div_mod<tileOutW>(relOutX0, relUpY, idx);
+                    int relUpX0 = relOutX0 * down;
+                    int src0 = relUpY * tileUpW + relUpX0;
+                    vec4_t v = InternalType<T>::zero_vec4();
+                    #pragma unroll
+                    for (int step = 0; step < fdSize; step++)
+                    {
+                        v.x += s_tileUpXY[src0 +  0 + step] * (scalar_t)c_fd[step];
+                        v.y += s_tileUpXY[src0 +  4 + step] * (scalar_t)c_fd[step];
+                        v.z += s_tileUpXY[src0 +  8 + step] * (scalar_t)c_fd[step];
+                        v.w += s_tileUpXY[src0 + 12 + step] * (scalar_t)c_fd[step];
+                    }
+                    s_tileDownX[idx+0] = v.x;
+                    s_tileDownX[idx+1] = v.y;
+                    s_tileDownX[idx+2] = v.z;
+                    s_tileDownX[idx+3] = v.w;
+                }
+            }
+            else if ((down == 2 || down == 4) && (tileOutW % 2 == 0))
+            {
+                // Calculate 2 pixels at a time.
+                for (int idx = threadIdx.x * 2; idx < tileOutW * tileUpH; idx += blockDim.x * 2)
+                {
+                    int relOutX0, relUpY;
+                    fast_div_mod<tileOutW>(relOutX0, relUpY, idx);
+                    int relUpX0 = relOutX0 * down;
+                    int src0 = relUpY * tileUpW + relUpX0;
+                    vec2_t v = InternalType<T>::zero_vec2();
+                    #pragma unroll
+                    for (int step = 0; step < fdSize; step++)
+                    {
+                        v.x += s_tileUpXY[src0 +    0 + step] * (scalar_t)c_fd[step];
+                        v.y += s_tileUpXY[src0 + down + step] * (scalar_t)c_fd[step];
+                    }
+                    s_tileDownX[idx+0] = v.x;
+                    s_tileDownX[idx+1] = v.y;
+                }
+            }
+            else
+            {
+                // Calculate 1 pixel at a time.
+                for (int idx = threadIdx.x; idx < tileOutW * tileUpH; idx += blockDim.x)
+                {
+                    int relOutX0, relUpY;
+                    fast_div_mod<tileOutW>(relOutX0, relUpY, idx);
+                    int relUpX0 = relOutX0 * down;
+                    int src = relUpY * tileUpW + relUpX0;
+                    scalar_t v = 0.f;
+                    #pragma unroll
+                    for (int step = 0; step < fdSize; step++)
+                        v += s_tileUpXY[src + step] * (scalar_t)c_fd[step];
+                    s_tileDownX[idx] = v;
+                }
+            }
+
+            // Vertical downsampling & store output tile.
+            __syncthreads();
+            for (int idx = threadIdx.x; idx < tileOutW * tileOutH; idx += blockDim.x)
+            {
+                int relOutX, relOutY0;
+                fast_div_mod<tileOutW>(relOutX, relOutY0, idx);
+                int relUpY0 = relOutY0 * down;
+                int src0 = relUpY0 * tileOutW + relOutX;
+                scalar_t v = 0;
+                #pragma unroll
+                for (int step = 0; step < fdSize; step++)
+                    v += s_tileDownX[src0 + step * tileOutW] * (scalar_t)c_fd[step];
+
+                int outX = tileOutX + relOutX;
+                int outY = tileOutY + relOutY0;
+
+                if (outX < p.yShape.x & outY < p.yShape.y)
+                    *((T*)((char*)p.y + (outX * get_stride<index_t>(p.yStride.x) + outY * get_stride<index_t>(p.yStride.y) + mapOfsOut))) = (T)v;
+            }
+        }
+        else if (filterMode == MODE_SUFD || filterMode == MODE_FUFD)
+        {
+            // Full downsampling filter.
+            if (down == 2)
+            {
+                // 2-wide.
+                __syncthreads();
+                for (int idx = threadIdx.x * 2; idx < tileOutW * tileOutH; idx += blockDim.x * 2)
+                {
+                    int relOutX0, relOutY0;
+                    fast_div_mod<tileOutW>(relOutX0, relOutY0, idx);
+                    int relUpX0 = relOutX0 * down;
+                    int relUpY0 = relOutY0 * down;
+                    int src0 = relUpY0 * tileUpW + relUpX0;
+                    vec2_t v = InternalType<T>::zero_vec2();
+                    #pragma unroll
+                    for (int sy = 0; sy < fdSize; sy++)
+                    #pragma unroll
+                    for (int sx = 0; sx < fdSize; sx++)
+                    {
+                        v.x += s_tileUpXY[src0 + 0 + sx + sy * tileUpW] * (scalar_t)c_fd[sx + sy * MAX_FILTER_SIZE];
+                        v.y += s_tileUpXY[src0 + 2 + sx + sy * tileUpW] * (scalar_t)c_fd[sx + sy * MAX_FILTER_SIZE];
+                    }
+
+                    int outX = tileOutX + relOutX0;
+                    int outY = tileOutY + relOutY0;
+                    if ((uint32_t)outY < p.yShape.y)
+                    {
+                        index_t ofs = outX * get_stride<index_t>(p.yStride.x) + outY * get_stride<index_t>(p.yStride.y) + mapOfsOut;
+                        if (outX + 0 < p.yShape.x) *((T*)((char*)p.y + ofs)) = (T)v.x;
+                        if (outX + 1 < p.yShape.x) *((T*)((char*)p.y + ofs + get_stride<index_t>(p.yStride.x))) = (T)v.y;
+                    }
+                }
+            }
+            else if (down == 1 && !downInline)
+            {
+                // Thread per pixel.
+                __syncthreads();
+                for (int idx = threadIdx.x; idx < tileOutW * tileOutH; idx += blockDim.x)
+                {
+                    int relOutX0, relOutY0;
+                    fast_div_mod<tileOutW>(relOutX0, relOutY0, idx);
+                    scalar_t v = s_tileUpXY[idx] * (scalar_t)c_fd[0]; // 1x1 filter.
+
+                    int outX = tileOutX + relOutX0;
+                    int outY = tileOutY + relOutY0;
+                    if ((uint32_t)outX < p.yShape.x && (uint32_t)outY < p.yShape.y)
+                        *((T*)((char*)p.y + (outX * get_stride<index_t>(p.yStride.x) + outY * get_stride<index_t>(p.yStride.y) + mapOfsOut))) = (T)v;
+                }
+            }
+        }
+
+        if (!enableXrep)
+            break;
+    }
+}
+
+//------------------------------------------------------------------------
+// Compute activation function and signs for upsampled data tensor, modifying data tensor in-place. Used for accelerating the generic variant.
+// Sign tensor is known to be contiguous, and p.x and p.s have the same z, w dimensions. 64-bit indexing is always used.
+
+template <class T, bool signWrite, bool signRead>
+static __global__ void filtered_lrelu_act_kernel(filtered_lrelu_act_kernel_params p)
+{
+    typedef typename InternalType<T>::scalar_t scalar_t;
+
+    // Indexing.
+    int32_t x = threadIdx.x + blockIdx.x * blockDim.x;
+    int32_t ymax = signWrite ? p.sShape.y : p.xShape.y;
+    int32_t qmax = p.xShape.z * p.xShape.w; // Combined minibatch*channel maximum index.
+
+    // Loop to accommodate oversized tensors.
+    for (int32_t q = blockIdx.z; q < qmax; q += gridDim.z)
+    for (int32_t y = blockIdx.y; y < ymax; y += gridDim.y)
+    {
+        // Extract z and w (channel, minibatch index).
+        int32_t w = q / p.xShape.z;
+        int32_t z = q - w * p.xShape.z;
+
+        // Choose behavior based on sign read/write mode.
+        if (signWrite)
+        {
+            // Process value if in p.x.
+            uint32_t s = 0;
+            if (x < p.xShape.x && y < p.xShape.y)
+            {
+                int64_t ix = x * p.xStride.x + y * p.xStride.y + z * p.xStride.z + w * p.xStride.w;
+                T* pv = ((T*)p.x) + ix;
+                scalar_t v = (scalar_t)(*pv);
+
+                // Gain, LReLU, clamp.
+                v *= p.gain;
+                if (v < 0.f)
+                {
+                    v *= p.slope;
+                    s = 1; // Sign.
+                }
+                if (fabsf(v) > p.clamp)
+                {
+                    v = InternalType<T>::clamp(v, p.clamp);
+                    s = 2; // Clamp.
+                }
+
+                *pv = (T)v; // Write value.
+            }
+
+            // Coalesce into threads 0 and 16 of warp.
+            uint32_t m = (threadIdx.x & 16) ? 0xffff0000u : 0x0000ffffu;
+            s <<= ((threadIdx.x & 15) << 1); // Shift into place.
+            s |= __shfl_xor_sync(m, s, 1); // Distribute.
+            s |= __shfl_xor_sync(m, s, 2);
+            s |= __shfl_xor_sync(m, s, 4);
+            s |= __shfl_xor_sync(m, s, 8);
+
+            // Write signs if leader and in p.s.
+            if (!(threadIdx.x & 15) && x < p.sShape.x) // y is always in.
+            {
+                uint64_t is = x + p.sShape.x * (y + (int64_t)p.sShape.y * q); // Contiguous.
+                ((uint32_t*)p.s)[is >> 4] = s;
+            }
+        }
+        else if (signRead)
+        {
+            // Process value if in p.x.
+            if (x < p.xShape.x) // y is always in.
+            {
+                int64_t ix = x * p.xStride.x + y * p.xStride.y + z * p.xStride.z + w * p.xStride.w;
+                T* pv = ((T*)p.x) + ix;
+                scalar_t v = (scalar_t)(*pv);
+                v *= p.gain;
+
+                // Apply sign buffer offset.
+                uint32_t sx = x + p.sOfs.x;
+                uint32_t sy = y + p.sOfs.y;
+
+                // Read and apply signs if we land inside valid region of sign buffer.
+                if (sx < p.sShape.x && sy < p.sShape.y)
+                {
+                    uint64_t is = (sx >> 2) + (p.sShape.x >> 2) * (sy + (uint64_t)p.sShape.y * q); // Contiguous.
+                    unsigned char s = p.s[is];
+                    s >>= (sx & 3) << 1; // Shift into place.
+                    if (s & 1) // Sign?
+                        v *= p.slope;
+                    if (s & 2) // Clamp?
+                        v = 0.f;
+                }
+
+                *pv = (T)v; // Write value.
+            }
+        }
+        else
+        {
+            // Forward pass with no sign write. Process value if in p.x.
+            if (x < p.xShape.x) // y is always in.
+            {
+                int64_t ix = x * p.xStride.x + y * p.xStride.y + z * p.xStride.z + w * p.xStride.w;
+                T* pv = ((T*)p.x) + ix;
+                scalar_t v = (scalar_t)(*pv);
+                v *= p.gain;
+                if (v < 0.f)
+                    v *= p.slope;
+                if (fabsf(v) > p.clamp)
+                    v = InternalType<T>::clamp(v, p.clamp);
+                *pv = (T)v; // Write value.
+            }
+        }
+    }
+}
+
+template <class T, bool signWrite, bool signRead> void* choose_filtered_lrelu_act_kernel(void)
+{
+    return (void*)filtered_lrelu_act_kernel<T, signWrite, signRead>;
+}
+
+//------------------------------------------------------------------------
+// CUDA kernel selection.
+
+template <class T, class index_t, bool signWrite, bool signRead> filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel(const filtered_lrelu_kernel_params& p, int sharedKB)
+{
+    filtered_lrelu_kernel_spec s = { 0 };
+
+    // Return the first matching kernel.
+#define CASE(SH, U, FU, D, FD, MODE, TW, TH, W, XR, WS) \
+    if (sharedKB >= SH) \
+    if ((p.fuShape.y == 0 && (MODE == MODE_SUSD || MODE == MODE_SUFD)) || (p.fuShape.y > 0 && (MODE == MODE_FUSD || MODE == MODE_FUFD))) \
+    if ((p.fdShape.y == 0 && (MODE == MODE_SUSD || MODE == MODE_FUSD)) || (p.fdShape.y > 0 && (MODE == MODE_SUFD || MODE == MODE_FUFD))) \
+    if (p.up == U && p.fuShape.x <= FU && p.fuShape.y <= FU && p.down == D && p.fdShape.x <= FD && p.fdShape.y <= FD) \
+    { \
+        static_assert((D*TW % 4) == 0, "down * tileWidth must be divisible by 4"); \
+        static_assert(FU % U == 0, "upscaling filter size must be multiple of upscaling factor"); \
+        static_assert(FD % D == 0, "downscaling filter size must be multiple of downscaling factor"); \
+        s.setup = (void*)setup_filters_kernel; \
+        s.exec = (void*)filtered_lrelu_kernel<T, index_t, SH, signWrite, signRead, MODE, U, FU, D, FD, TW, TH, W*32, !!XR, !!WS>; \
+        s.tileOut = make_int2(TW, TH); \
+        s.numWarps = W; \
+        s.xrep = XR; \
+        s.dynamicSharedKB = (SH == 48) ? 0 : SH; \
+        return s; \
+    }
+
+    // Launch parameters for various kernel specializations.
+    // Small filters must be listed before large filters, otherwise the kernel for larger filter will always match first.
+    // Kernels that use more shared memory must be listed before those that use less, for the same reason.
+
+    CASE(/*sharedKB*/48, /*up,fu*/1,1,  /*down,fd*/1,1,  /*mode*/MODE_FUFD, /*tw,th,warps,xrep,wskip*/64,  178, 32,  0,  0) // 1t-upf1-downf1
+    CASE(/*sharedKB*/48, /*up,fu*/2,8,  /*down,fd*/1,1,  /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/152, 95,  16,  0,  0) // 4t-ups2-downf1
+    CASE(/*sharedKB*/48, /*up,fu*/1,1,  /*down,fd*/2,8,  /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/56,  22,  16,  0,  0) // 4t-upf1-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,8,  /*down,fd*/2,8,  /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/56,  29,  16,  11, 0) // 4t-ups2-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,8,  /*down,fd*/2,8,  /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/60,  28,  16,  0,  0) // 4t-upf2-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,8,  /*down,fd*/2,8,  /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/56,  28,  16,  0,  0) // 4t-ups2-downf2
+    CASE(/*sharedKB*/48, /*up,fu*/4,16, /*down,fd*/2,8,  /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/56,  31,  16,  11, 0) // 4t-ups4-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/4,16, /*down,fd*/2,8,  /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/56,  36,  16,  0,  0) // 4t-ups4-downf2
+    CASE(/*sharedKB*/48, /*up,fu*/2,8,  /*down,fd*/4,16, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/16,  22,  16,  12, 0) // 4t-ups2-downs4
+    CASE(/*sharedKB*/48, /*up,fu*/2,8,  /*down,fd*/4,16, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/29,  15,  16,  0,  0) // 4t-upf2-downs4
+    CASE(/*sharedKB*/48, /*up,fu*/2,12, /*down,fd*/1,1,  /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/96,  150, 28,  0,  0) // 6t-ups2-downf1
+    CASE(/*sharedKB*/48, /*up,fu*/1,1,  /*down,fd*/2,12, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/32,  35,  24,  0,  0) // 6t-upf1-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,12, /*down,fd*/2,12, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/32,  46,  16,  10, 0) // 6t-ups2-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,12, /*down,fd*/2,12, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/58,  28,  24,  8,  0) // 6t-upf2-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,12, /*down,fd*/2,12, /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/52,  28,  16,  0,  0) // 6t-ups2-downf2
+    CASE(/*sharedKB*/48, /*up,fu*/4,24, /*down,fd*/2,12, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/32,  51,  16,  5,  0) // 6t-ups4-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/4,24, /*down,fd*/2,12, /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/32,  56,  16,  6,  0) // 6t-ups4-downf2
+    CASE(/*sharedKB*/48, /*up,fu*/2,12, /*down,fd*/4,24, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/16,  18,  16,  12, 0) // 6t-ups2-downs4
+    CASE(/*sharedKB*/96, /*up,fu*/2,12, /*down,fd*/4,24, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/27,  31,  32,  6,  0) // 6t-upf2-downs4 96kB
+    CASE(/*sharedKB*/48, /*up,fu*/2,12, /*down,fd*/4,24, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/27,  13,  24,  0,  0) // 6t-upf2-downs4
+    CASE(/*sharedKB*/48, /*up,fu*/2,16, /*down,fd*/1,1,  /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/148, 89,  24,  0,  0) // 8t-ups2-downf1
+    CASE(/*sharedKB*/48, /*up,fu*/1,1,  /*down,fd*/2,16, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/32,  31,  16,  5,  0) // 8t-upf1-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,16, /*down,fd*/2,16, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/32,  41,  16,  9,  0) // 8t-ups2-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,16, /*down,fd*/2,16, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/56,  26,  24,  0,  0) // 8t-upf2-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,16, /*down,fd*/2,16, /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/32,  40,  16,  0,  0) // 8t-ups2-downf2
+    CASE(/*sharedKB*/48, /*up,fu*/4,32, /*down,fd*/2,16, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/32,  46,  24,  5,  0) // 8t-ups4-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/4,32, /*down,fd*/2,16, /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/32,  50,  16,  0,  0) // 8t-ups4-downf2
+    CASE(/*sharedKB*/96, /*up,fu*/2,16, /*down,fd*/4,32, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/24,  24,  32,  12, 1) // 8t-ups2-downs4 96kB
+    CASE(/*sharedKB*/48, /*up,fu*/2,16, /*down,fd*/4,32, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/16,  13,  16,  10, 1) // 8t-ups2-downs4
+    CASE(/*sharedKB*/96, /*up,fu*/2,16, /*down,fd*/4,32, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/25,  28,  28,  4,  0) // 8t-upf2-downs4 96kB
+    CASE(/*sharedKB*/48, /*up,fu*/2,16, /*down,fd*/4,32, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/25,  10,  24,  0,  0) // 8t-upf2-downs4
+
+    #undef CASE
+    return s; // No kernel found.
+}
+
+//------------------------------------------------------------------------

models/torch_utils/ops/filtered_lrelu.h

@@ -0,0 +1,90 @@
+// Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#include <cuda_runtime.h>
+
+//------------------------------------------------------------------------
+// CUDA kernel parameters.
+
+struct filtered_lrelu_kernel_params
+{
+    // These parameters decide which kernel to use.
+    int             up;         // upsampling ratio (1, 2, 4)
+    int             down;       // downsampling ratio (1, 2, 4)
+    int2            fuShape;    // [size, 1] | [size, size]
+    int2            fdShape;    // [size, 1] | [size, size]
+
+    int             _dummy;     // Alignment.
+
+    // Rest of the parameters.
+    const void*     x;          // Input tensor.
+    void*           y;          // Output tensor.
+    const void*     b;          // Bias tensor.
+    unsigned char*  s;          // Sign tensor in/out. NULL if unused.
+    const float*    fu;         // Upsampling filter.
+    const float*    fd;         // Downsampling filter.
+
+    int2            pad0;       // Left/top padding.
+    float           gain;       // Additional gain factor.
+    float           slope;      // Leaky ReLU slope on negative side.
+    float           clamp;      // Clamp after nonlinearity.
+    int             flip;       // Filter kernel flip for gradient computation.
+
+    int             tilesXdim;  // Original number of horizontal output tiles.
+    int             tilesXrep;  // Number of horizontal tiles per CTA.
+    int             blockZofs;  // Block z offset to support large minibatch, channel dimensions.
+
+    int4            xShape;     // [width, height, channel, batch]
+    int4            yShape;     // [width, height, channel, batch]
+    int2            sShape;     // [width, height] - width is in bytes. Contiguous. Zeros if unused.
+    int2            sOfs;       // [ofs_x, ofs_y] - offset between upsampled data and sign tensor.
+    int             swLimit;    // Active width of sign tensor in bytes.
+
+    longlong4       xStride;    // Strides of all tensors except signs, same component order as shapes.
+    longlong4       yStride;    //
+    int64_t         bStride;    //
+    longlong3       fuStride;   //
+    longlong3       fdStride;   //
+};
+
+struct filtered_lrelu_act_kernel_params
+{
+    void*           x;          // Input/output, modified in-place.
+    unsigned char*  s;          // Sign tensor in/out. NULL if unused.
+
+    float           gain;       // Additional gain factor.
+    float           slope;      // Leaky ReLU slope on negative side.
+    float           clamp;      // Clamp after nonlinearity.
+
+    int4            xShape;     // [width, height, channel, batch]
+    longlong4       xStride;    // Input/output tensor strides, same order as in shape.
+    int2            sShape;     // [width, height] - width is in elements. Contiguous. Zeros if unused.
+    int2            sOfs;       // [ofs_x, ofs_y] - offset between upsampled data and sign tensor.
+};
+
+//------------------------------------------------------------------------
+// CUDA kernel specialization.
+
+struct filtered_lrelu_kernel_spec
+{
+    void*   setup;              // Function for filter kernel setup.
+    void*   exec;               // Function for main operation.
+    int2    tileOut;            // Width/height of launch tile.
+    int     numWarps;           // Number of warps per thread block, determines launch block size.
+    int     xrep;               // For processing multiple horizontal tiles per thread block.
+    int     dynamicSharedKB;    // How much dynamic shared memory the exec kernel wants.
+};
+
+//------------------------------------------------------------------------
+// CUDA kernel selection.
+
+template <class T, class index_t, bool signWrite, bool signRead> filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel(const filtered_lrelu_kernel_params& p, int sharedKB);
+template <class T, bool signWrite, bool signRead> void* choose_filtered_lrelu_act_kernel(void);
+template <bool signWrite, bool signRead> cudaError_t copy_filters(cudaStream_t stream);
+
+//------------------------------------------------------------------------

models/torch_utils/ops/filtered_lrelu.py

@@ -0,0 +1,274 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import os
+import numpy as np
+import torch
+import warnings
+
+from .. import custom_ops
+from .. import misc
+from . import upfirdn2d
+from . import bias_act
+
+#----------------------------------------------------------------------------
+
+_plugin = None
+
+def _init():
+    global _plugin
+    if _plugin is None:
+        _plugin = custom_ops.get_plugin(
+            module_name='filtered_lrelu_plugin',
+            sources=['filtered_lrelu.cpp', 'filtered_lrelu_wr.cu', 'filtered_lrelu_rd.cu', 'filtered_lrelu_ns.cu'],
+            headers=['filtered_lrelu.h', 'filtered_lrelu.cu'],
+            source_dir=os.path.dirname(__file__),
+            extra_cuda_cflags=['--use_fast_math'],
+        )
+    return True
+
+def _get_filter_size(f):
+    if f is None:
+        return 1, 1
+    assert isinstance(f, torch.Tensor)
+    assert 1 <= f.ndim <= 2
+    return f.shape[-1], f.shape[0] # width, height
+
+def _parse_padding(padding):
+    if isinstance(padding, int):
+        padding = [padding, padding]
+    assert isinstance(padding, (list, tuple))
+    assert all(isinstance(x, (int, np.integer)) for x in padding)
+    padding = [int(x) for x in padding]
+    if len(padding) == 2:
+        px, py = padding
+        padding = [px, px, py, py]
+    px0, px1, py0, py1 = padding
+    return px0, px1, py0, py1
+
+#----------------------------------------------------------------------------
+
+def filtered_lrelu(x, fu=None, fd=None, b=None, up=1, down=1, padding=0, gain=np.sqrt(2), slope=0.2, clamp=None, flip_filter=False, impl='cuda'):
+    r"""Filtered leaky ReLU for a batch of 2D images.
+
+    Performs the following sequence of operations for each channel:
+
+    1. Add channel-specific bias if provided (`b`).
+
+    2. Upsample the image by inserting N-1 zeros after each pixel (`up`).
+
+    3. Pad the image with the specified number of zeros on each side (`padding`).
+       Negative padding corresponds to cropping the image.
+
+    4. Convolve the image with the specified upsampling FIR filter (`fu`), shrinking it
+       so that the footprint of all output pixels lies within the input image.
+
+    5. Multiply each value by the provided gain factor (`gain`).
+
+    6. Apply leaky ReLU activation function to each value.
+
+    7. Clamp each value between -clamp and +clamp, if `clamp` parameter is provided.
+
+    8. Convolve the image with the specified downsampling FIR filter (`fd`), shrinking
+       it so that the footprint of all output pixels lies within the input image.
+
+    9. Downsample the image by keeping every Nth pixel (`down`).
+
+    The fused op is considerably more efficient than performing the same calculation
+    using standard PyTorch ops. It supports gradients of arbitrary order.
+
+    Args:
+        x:           Float32/float16/float64 input tensor of the shape
+                     `[batch_size, num_channels, in_height, in_width]`.
+        fu:          Float32 upsampling FIR filter of the shape
+                     `[filter_height, filter_width]` (non-separable),
+                     `[filter_taps]` (separable), or
+                     `None` (identity).
+        fd:          Float32 downsampling FIR filter of the shape
+                     `[filter_height, filter_width]` (non-separable),
+                     `[filter_taps]` (separable), or
+                     `None` (identity).
+        b:           Bias vector, or `None` to disable. Must be a 1D tensor of the same type
+                     as `x`. The length of vector must must match the channel dimension of `x`.
+        up:          Integer upsampling factor (default: 1).
+        down:        Integer downsampling factor. (default: 1).
+        padding:     Padding with respect to the upsampled image. Can be a single number
+                     or a list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
+                     (default: 0).
+        gain:        Overall scaling factor for signal magnitude (default: sqrt(2)).
+        slope:       Slope on the negative side of leaky ReLU (default: 0.2).
+        clamp:       Maximum magnitude for leaky ReLU output (default: None).
+        flip_filter: False = convolution, True = correlation (default: False).
+        impl:        Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
+
+    Returns:
+        Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
+    """
+    assert isinstance(x, torch.Tensor)
+    assert impl in ['ref', 'cuda']
+    if impl == 'cuda' and x.device.type == 'cuda' and _init():
+        return _filtered_lrelu_cuda(up=up, down=down, padding=padding, gain=gain, slope=slope, clamp=clamp, flip_filter=flip_filter).apply(x, fu, fd, b, None, 0, 0)
+    return _filtered_lrelu_ref(x, fu=fu, fd=fd, b=b, up=up, down=down, padding=padding, gain=gain, slope=slope, clamp=clamp, flip_filter=flip_filter)
+
+#----------------------------------------------------------------------------
+
+@misc.profiled_function
+def _filtered_lrelu_ref(x, fu=None, fd=None, b=None, up=1, down=1, padding=0, gain=np.sqrt(2), slope=0.2, clamp=None, flip_filter=False):
+    """Slow and memory-inefficient reference implementation of `filtered_lrelu()` using
+    existing `upfirdn2n()` and `bias_act()` ops.
+    """
+    assert isinstance(x, torch.Tensor) and x.ndim == 4
+    fu_w, fu_h = _get_filter_size(fu)
+    fd_w, fd_h = _get_filter_size(fd)
+    if b is not None:
+        assert isinstance(b, torch.Tensor) and b.dtype == x.dtype
+        misc.assert_shape(b, [x.shape[1]])
+    assert isinstance(up, int) and up >= 1
+    assert isinstance(down, int) and down >= 1
+    px0, px1, py0, py1 = _parse_padding(padding)
+    assert gain == float(gain) and gain > 0
+    assert slope == float(slope) and slope >= 0
+    assert clamp is None or (clamp == float(clamp) and clamp >= 0)
+
+    # Calculate output size.
+    batch_size, channels, in_h, in_w = x.shape
+    in_dtype = x.dtype
+    out_w = (in_w * up + (px0 + px1) - (fu_w - 1) - (fd_w - 1) + (down - 1)) // down
+    out_h = (in_h * up + (py0 + py1) - (fu_h - 1) - (fd_h - 1) + (down - 1)) // down
+
+    # Compute using existing ops.
+    x = bias_act.bias_act(x=x, b=b) # Apply bias.
+    x = upfirdn2d.upfirdn2d(x=x, f=fu, up=up, padding=[px0, px1, py0, py1], gain=up**2, flip_filter=flip_filter) # Upsample.
+    x = bias_act.bias_act(x=x, act='lrelu', alpha=slope, gain=gain, clamp=clamp) # Bias, leaky ReLU, clamp.
+    x = upfirdn2d.upfirdn2d(x=x, f=fd, down=down, flip_filter=flip_filter) # Downsample.
+
+    # Check output shape & dtype.
+    misc.assert_shape(x, [batch_size, channels, out_h, out_w])
+    assert x.dtype == in_dtype
+    return x
+
+#----------------------------------------------------------------------------
+
+_filtered_lrelu_cuda_cache = dict()
+
+def _filtered_lrelu_cuda(up=1, down=1, padding=0, gain=np.sqrt(2), slope=0.2, clamp=None, flip_filter=False):
+    """Fast CUDA implementation of `filtered_lrelu()` using custom ops.
+    """
+    assert isinstance(up, int) and up >= 1
+    assert isinstance(down, int) and down >= 1
+    px0, px1, py0, py1 = _parse_padding(padding)
+    assert gain == float(gain) and gain > 0
+    gain = float(gain)
+    assert slope == float(slope) and slope >= 0
+    slope = float(slope)
+    assert clamp is None or (clamp == float(clamp) and clamp >= 0)
+    clamp = float(clamp if clamp is not None else 'inf')
+
+    # Lookup from cache.
+    key = (up, down, px0, px1, py0, py1, gain, slope, clamp, flip_filter)
+    if key in _filtered_lrelu_cuda_cache:
+        return _filtered_lrelu_cuda_cache[key]
+
+    # Forward op.
+    class FilteredLReluCuda(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, x, fu, fd, b, si, sx, sy): # pylint: disable=arguments-differ
+            assert isinstance(x, torch.Tensor) and x.ndim == 4
+
+            # Replace empty up/downsample kernels with full 1x1 kernels (faster than separable).
+            if fu is None:
+                fu = torch.ones([1, 1], dtype=torch.float32, device=x.device)
+            if fd is None:
+                fd = torch.ones([1, 1], dtype=torch.float32, device=x.device)
+            assert 1 <= fu.ndim <= 2
+            assert 1 <= fd.ndim <= 2
+
+            # Replace separable 1x1 kernels with full 1x1 kernels when scale factor is 1.
+            if up == 1 and fu.ndim == 1 and fu.shape[0] == 1:
+                fu = fu.square()[None]
+            if down == 1 and fd.ndim == 1 and fd.shape[0] == 1:
+                fd = fd.square()[None]
+
+            # Missing sign input tensor.
+            if si is None:
+                si = torch.empty([0])
+
+            # Missing bias tensor.
+            if b is None:
+                b = torch.zeros([x.shape[1]], dtype=x.dtype, device=x.device)
+
+            # Construct internal sign tensor only if gradients are needed.
+            write_signs = (si.numel() == 0) and (x.requires_grad or b.requires_grad)
+
+            # Warn if input storage strides are not in decreasing order due to e.g. channels-last layout.
+            strides = [x.stride(i) for i in range(x.ndim) if x.size(i) > 1]
+            if any(a < b for a, b in zip(strides[:-1], strides[1:])):
+                warnings.warn("low-performance memory layout detected in filtered_lrelu input", RuntimeWarning)
+
+            # Call C++/Cuda plugin if datatype is supported.
+            if x.dtype in [torch.float16, torch.float32]:
+                if torch.cuda.current_stream(x.device) != torch.cuda.default_stream(x.device):
+                    warnings.warn("filtered_lrelu called with non-default cuda stream but concurrent execution is not supported", RuntimeWarning)
+                y, so, return_code = _plugin.filtered_lrelu(x, fu, fd, b, si, up, down, px0, px1, py0, py1, sx, sy, gain, slope, clamp, flip_filter, write_signs)
+            else:
+                return_code = -1
+
+            # No Cuda kernel found? Fall back to generic implementation. Still more memory efficient than the reference implementation because
+            # only the bit-packed sign tensor is retained for gradient computation.
+            if return_code < 0:
+                warnings.warn("filtered_lrelu called with parameters that have no optimized CUDA kernel, using generic fallback", RuntimeWarning)
+
+                y = x.add(b.unsqueeze(-1).unsqueeze(-1)) # Add bias.
+                y = upfirdn2d.upfirdn2d(x=y, f=fu, up=up, padding=[px0, px1, py0, py1], gain=up**2, flip_filter=flip_filter) # Upsample.
+                so = _plugin.filtered_lrelu_act_(y, si, sx, sy, gain, slope, clamp, write_signs) # Activation function and sign handling. Modifies y in-place.
+                y = upfirdn2d.upfirdn2d(x=y, f=fd, down=down, flip_filter=flip_filter) # Downsample.
+
+            # Prepare for gradient computation.
+            ctx.save_for_backward(fu, fd, (si if si.numel() else so))
+            ctx.x_shape = x.shape
+            ctx.y_shape = y.shape
+            ctx.s_ofs = sx, sy
+            return y
+
+        @staticmethod
+        def backward(ctx, dy): # pylint: disable=arguments-differ
+            fu, fd, si = ctx.saved_tensors
+            _, _, xh, xw = ctx.x_shape
+            _, _, yh, yw = ctx.y_shape
+            sx, sy = ctx.s_ofs
+            dx  = None # 0
+            dfu = None; assert not ctx.needs_input_grad[1]
+            dfd = None; assert not ctx.needs_input_grad[2]
+            db  = None # 3
+            dsi = None; assert not ctx.needs_input_grad[4]
+            dsx = None; assert not ctx.needs_input_grad[5]
+            dsy = None; assert not ctx.needs_input_grad[6]
+
+            if ctx.needs_input_grad[0] or ctx.needs_input_grad[3]:
+                pp = [
+                    (fu.shape[-1] - 1) + (fd.shape[-1] - 1) - px0,
+                    xw * up - yw * down + px0 - (up - 1),
+                    (fu.shape[0] - 1) + (fd.shape[0] - 1) - py0,
+                    xh * up - yh * down + py0 - (up - 1),
+                ]
+                gg = gain * (up ** 2) / (down ** 2)
+                ff = (not flip_filter)
+                sx = sx - (fu.shape[-1] - 1) + px0
+                sy = sy - (fu.shape[0]  - 1) + py0
+                dx = _filtered_lrelu_cuda(up=down, down=up, padding=pp, gain=gg, slope=slope, clamp=None, flip_filter=ff).apply(dy, fd, fu, None, si, sx, sy)
+
+            if ctx.needs_input_grad[3]:
+                db = dx.sum([0, 2, 3])
+
+            return dx, dfu, dfd, db, dsi, dsx, dsy
+
+    # Add to cache.
+    _filtered_lrelu_cuda_cache[key] = FilteredLReluCuda
+    return FilteredLReluCuda
+
+#----------------------------------------------------------------------------

models/torch_utils/ops/filtered_lrelu_ns.cu

@@ -0,0 +1,27 @@
+// Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#include "filtered_lrelu.cu"
+
+// Template/kernel specializations for no signs mode (no gradients required).
+
+// Full op, 32-bit indexing.
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<c10::Half, int32_t, false, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<float,     int32_t, false, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+
+// Full op, 64-bit indexing.
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<c10::Half, int64_t, false, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<float,     int64_t, false, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+
+// Activation/signs only for generic variant. 64-bit indexing.
+template void* choose_filtered_lrelu_act_kernel<c10::Half, false, false>(void);
+template void* choose_filtered_lrelu_act_kernel<float,     false, false>(void);
+template void* choose_filtered_lrelu_act_kernel<double,    false, false>(void);
+
+// Copy filters to constant memory.
+template cudaError_t copy_filters<false, false>(cudaStream_t stream);

models/torch_utils/ops/filtered_lrelu_rd.cu

@@ -0,0 +1,27 @@
+// Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#include "filtered_lrelu.cu"
+
+// Template/kernel specializations for sign read mode.
+
+// Full op, 32-bit indexing.
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<c10::Half, int32_t, false, true>(const filtered_lrelu_kernel_params& p, int sharedKB);
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<float,     int32_t, false, true>(const filtered_lrelu_kernel_params& p, int sharedKB);
+
+// Full op, 64-bit indexing.
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<c10::Half, int64_t, false, true>(const filtered_lrelu_kernel_params& p, int sharedKB);
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<float,     int64_t, false, true>(const filtered_lrelu_kernel_params& p, int sharedKB);
+
+// Activation/signs only for generic variant. 64-bit indexing.
+template void* choose_filtered_lrelu_act_kernel<c10::Half, false, true>(void);
+template void* choose_filtered_lrelu_act_kernel<float,     false, true>(void);
+template void* choose_filtered_lrelu_act_kernel<double,    false, true>(void);
+
+// Copy filters to constant memory.
+template cudaError_t copy_filters<false, true>(cudaStream_t stream);

models/torch_utils/ops/filtered_lrelu_wr.cu

@@ -0,0 +1,27 @@
+// Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#include "filtered_lrelu.cu"
+
+// Template/kernel specializations for sign write mode.
+
+// Full op, 32-bit indexing.
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<c10::Half, int32_t, true, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<float,     int32_t, true, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+
+// Full op, 64-bit indexing.
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<c10::Half, int64_t, true, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<float,     int64_t, true, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+
+// Activation/signs only for generic variant. 64-bit indexing.
+template void* choose_filtered_lrelu_act_kernel<c10::Half, true, false>(void);
+template void* choose_filtered_lrelu_act_kernel<float,     true, false>(void);
+template void* choose_filtered_lrelu_act_kernel<double,    true, false>(void);
+
+// Copy filters to constant memory.
+template cudaError_t copy_filters<true, false>(cudaStream_t stream);

models/torch_utils/ops/fma.py

@@ -0,0 +1,60 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Fused multiply-add, with slightly faster gradients than `torch.addcmul()`."""
+
+import torch
+
+#----------------------------------------------------------------------------
+
+def fma(a, b, c): # => a * b + c
+    return _FusedMultiplyAdd.apply(a, b, c)
+
+#----------------------------------------------------------------------------
+
+class _FusedMultiplyAdd(torch.autograd.Function): # a * b + c
+    @staticmethod
+    def forward(ctx, a, b, c): # pylint: disable=arguments-differ
+        out = torch.addcmul(c, a, b)
+        ctx.save_for_backward(a, b)
+        ctx.c_shape = c.shape
+        return out
+
+    @staticmethod
+    def backward(ctx, dout): # pylint: disable=arguments-differ
+        a, b = ctx.saved_tensors
+        c_shape = ctx.c_shape
+        da = None
+        db = None
+        dc = None
+
+        if ctx.needs_input_grad[0]:
+            da = _unbroadcast(dout * b, a.shape)
+
+        if ctx.needs_input_grad[1]:
+            db = _unbroadcast(dout * a, b.shape)
+
+        if ctx.needs_input_grad[2]:
+            dc = _unbroadcast(dout, c_shape)
+
+        return da, db, dc
+
+#----------------------------------------------------------------------------
+
+def _unbroadcast(x, shape):
+    extra_dims = x.ndim - len(shape)
+    assert extra_dims >= 0
+    dim = [i for i in range(x.ndim) if x.shape[i] > 1 and (i < extra_dims or shape[i - extra_dims] == 1)]
+    if len(dim):
+        x = x.sum(dim=dim, keepdim=True)
+    if extra_dims:
+        x = x.reshape(-1, *x.shape[extra_dims+1:])
+    assert x.shape == shape
+    return x
+
+#----------------------------------------------------------------------------

models/torch_utils/ops/grid_sample_gradfix.py

@@ -0,0 +1,77 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Custom replacement for `torch.nn.functional.grid_sample` that
+supports arbitrarily high order gradients between the input and output.
+Only works on 2D images and assumes
+`mode='bilinear'`, `padding_mode='zeros'`, `align_corners=False`."""
+
+import torch
+
+# pylint: disable=redefined-builtin
+# pylint: disable=arguments-differ
+# pylint: disable=protected-access
+
+#----------------------------------------------------------------------------
+
+enabled = False  # Enable the custom op by setting this to true.
+
+#----------------------------------------------------------------------------
+
+def grid_sample(input, grid):
+    if _should_use_custom_op():
+        return _GridSample2dForward.apply(input, grid)
+    return torch.nn.functional.grid_sample(input=input, grid=grid, mode='bilinear', padding_mode='zeros', align_corners=False)
+
+#----------------------------------------------------------------------------
+
+def _should_use_custom_op():
+    return enabled
+
+#----------------------------------------------------------------------------
+
+class _GridSample2dForward(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, input, grid):
+        assert input.ndim == 4
+        assert grid.ndim == 4
+        output = torch.nn.functional.grid_sample(input=input, grid=grid, mode='bilinear', padding_mode='zeros', align_corners=False)
+        ctx.save_for_backward(input, grid)
+        return output
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        input, grid = ctx.saved_tensors
+        grad_input, grad_grid = _GridSample2dBackward.apply(grad_output, input, grid)
+        return grad_input, grad_grid
+
+#----------------------------------------------------------------------------
+
+class _GridSample2dBackward(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, grad_output, input, grid):
+        op = torch._C._jit_get_operation('aten::grid_sampler_2d_backward')
+        grad_input, grad_grid = op(grad_output, input, grid, 0, 0, False)
+        ctx.save_for_backward(grid)
+        return grad_input, grad_grid
+
+    @staticmethod
+    def backward(ctx, grad2_grad_input, grad2_grad_grid):
+        _ = grad2_grad_grid # unused
+        grid, = ctx.saved_tensors
+        grad2_grad_output = None
+        grad2_input = None
+        grad2_grid = None
+
+        if ctx.needs_input_grad[0]:
+            grad2_grad_output = _GridSample2dForward.apply(grad2_grad_input, grid)
+
+        assert not ctx.needs_input_grad[2]
+        return grad2_grad_output, grad2_input, grad2_grid
+
+#----------------------------------------------------------------------------

models/torch_utils/ops/upfirdn2d.cpp

@@ -0,0 +1,107 @@
+// Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#include <torch/extension.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include "upfirdn2d.h"
+
+//------------------------------------------------------------------------
+
+static torch::Tensor upfirdn2d(torch::Tensor x, torch::Tensor f, int upx, int upy, int downx, int downy, int padx0, int padx1, int pady0, int pady1, bool flip, float gain)
+{
+    // Validate arguments.
+    TORCH_CHECK(x.is_cuda(), "x must reside on CUDA device");
+    TORCH_CHECK(f.device() == x.device(), "f must reside on the same device as x");
+    TORCH_CHECK(f.dtype() == torch::kFloat, "f must be float32");
+    TORCH_CHECK(x.numel() <= INT_MAX, "x is too large");
+    TORCH_CHECK(f.numel() <= INT_MAX, "f is too large");
+    TORCH_CHECK(x.numel() > 0, "x has zero size");
+    TORCH_CHECK(f.numel() > 0, "f has zero size");
+    TORCH_CHECK(x.dim() == 4, "x must be rank 4");
+    TORCH_CHECK(f.dim() == 2, "f must be rank 2");
+    TORCH_CHECK((x.size(0)-1)*x.stride(0) + (x.size(1)-1)*x.stride(1) + (x.size(2)-1)*x.stride(2) + (x.size(3)-1)*x.stride(3) <= INT_MAX, "x memory footprint is too large");
+    TORCH_CHECK(f.size(0) >= 1 && f.size(1) >= 1, "f must be at least 1x1");
+    TORCH_CHECK(upx >= 1 && upy >= 1, "upsampling factor must be at least 1");
+    TORCH_CHECK(downx >= 1 && downy >= 1, "downsampling factor must be at least 1");
+
+    // Create output tensor.
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
+    int outW = ((int)x.size(3) * upx + padx0 + padx1 - (int)f.size(1) + downx) / downx;
+    int outH = ((int)x.size(2) * upy + pady0 + pady1 - (int)f.size(0) + downy) / downy;
+    TORCH_CHECK(outW >= 1 && outH >= 1, "output must be at least 1x1");
+    torch::Tensor y = torch::empty({x.size(0), x.size(1), outH, outW}, x.options(), x.suggest_memory_format());
+    TORCH_CHECK(y.numel() <= INT_MAX, "output is too large");
+    TORCH_CHECK((y.size(0)-1)*y.stride(0) + (y.size(1)-1)*y.stride(1) + (y.size(2)-1)*y.stride(2) + (y.size(3)-1)*y.stride(3) <= INT_MAX, "output memory footprint is too large");
+
+    // Initialize CUDA kernel parameters.
+    upfirdn2d_kernel_params p;
+    p.x             = x.data_ptr();
+    p.f             = f.data_ptr<float>();
+    p.y             = y.data_ptr();
+    p.up            = make_int2(upx, upy);
+    p.down          = make_int2(downx, downy);
+    p.pad0          = make_int2(padx0, pady0);
+    p.flip          = (flip) ? 1 : 0;
+    p.gain          = gain;
+    p.inSize        = make_int4((int)x.size(3), (int)x.size(2), (int)x.size(1), (int)x.size(0));
+    p.inStride      = make_int4((int)x.stride(3), (int)x.stride(2), (int)x.stride(1), (int)x.stride(0));
+    p.filterSize    = make_int2((int)f.size(1), (int)f.size(0));
+    p.filterStride  = make_int2((int)f.stride(1), (int)f.stride(0));
+    p.outSize       = make_int4((int)y.size(3), (int)y.size(2), (int)y.size(1), (int)y.size(0));
+    p.outStride     = make_int4((int)y.stride(3), (int)y.stride(2), (int)y.stride(1), (int)y.stride(0));
+    p.sizeMajor     = (p.inStride.z == 1) ? p.inSize.w : p.inSize.w * p.inSize.z;
+    p.sizeMinor     = (p.inStride.z == 1) ? p.inSize.z : 1;
+
+    // Choose CUDA kernel.
+    upfirdn2d_kernel_spec spec;
+    AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), "upfirdn2d_cuda", [&]
+    {
+        spec = choose_upfirdn2d_kernel<scalar_t>(p);
+    });
+
+    // Set looping options.
+    p.loopMajor     = (p.sizeMajor - 1) / 16384 + 1;
+    p.loopMinor     = spec.loopMinor;
+    p.loopX         = spec.loopX;
+    p.launchMinor   = (p.sizeMinor - 1) / p.loopMinor + 1;
+    p.launchMajor   = (p.sizeMajor - 1) / p.loopMajor + 1;
+
+    // Compute grid size.
+    dim3 blockSize, gridSize;
+    if (spec.tileOutW < 0) // large
+    {
+        blockSize = dim3(4, 32, 1);
+        gridSize = dim3(
+            ((p.outSize.y - 1) / blockSize.x + 1) * p.launchMinor,
+            (p.outSize.x - 1) / (blockSize.y * p.loopX) + 1,
+            p.launchMajor);
+    }
+    else // small
+    {
+        blockSize = dim3(256, 1, 1);
+        gridSize = dim3(
+            ((p.outSize.y - 1) / spec.tileOutH + 1) * p.launchMinor,
+            (p.outSize.x - 1) / (spec.tileOutW * p.loopX) + 1,
+            p.launchMajor);
+    }
+
+    // Launch CUDA kernel.
+    void* args[] = {&p};
+    AT_CUDA_CHECK(cudaLaunchKernel(spec.kernel, gridSize, blockSize, args, 0, at::cuda::getCurrentCUDAStream()));
+    return y;
+}
+
+//------------------------------------------------------------------------
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m)
+{
+    m.def("upfirdn2d", &upfirdn2d);
+}
+
+//------------------------------------------------------------------------

models/torch_utils/ops/upfirdn2d.cu

@@ -0,0 +1,384 @@
+// Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#include <c10/util/Half.h>
+#include "upfirdn2d.h"
+
+//------------------------------------------------------------------------
+// Helpers.
+
+template <class T> struct InternalType;
+template <> struct InternalType<double>     { typedef double scalar_t; };
+template <> struct InternalType<float>      { typedef float  scalar_t; };
+template <> struct InternalType<c10::Half>  { typedef float  scalar_t; };
+
+static __device__ __forceinline__ int floor_div(int a, int b)
+{
+    int t = 1 - a / b;
+    return (a + t * b) / b - t;
+}
+
+//------------------------------------------------------------------------
+// Generic CUDA implementation for large filters.
+
+template <class T> static __global__ void upfirdn2d_kernel_large(upfirdn2d_kernel_params p)
+{
+    typedef typename InternalType<T>::scalar_t scalar_t;
+
+    // Calculate thread index.
+    int minorBase = blockIdx.x * blockDim.x + threadIdx.x;
+    int outY = minorBase / p.launchMinor;
+    minorBase -= outY * p.launchMinor;
+    int outXBase = blockIdx.y * p.loopX * blockDim.y + threadIdx.y;
+    int majorBase = blockIdx.z * p.loopMajor;
+    if (outXBase >= p.outSize.x | outY >= p.outSize.y | majorBase >= p.sizeMajor)
+        return;
+
+    // Setup Y receptive field.
+    int midY = outY * p.down.y + p.up.y - 1 - p.pad0.y;
+    int inY = min(max(floor_div(midY, p.up.y), 0), p.inSize.y);
+    int h = min(max(floor_div(midY + p.filterSize.y, p.up.y), 0), p.inSize.y) - inY;
+    int filterY = midY + p.filterSize.y - (inY + 1) * p.up.y;
+    if (p.flip)
+        filterY = p.filterSize.y - 1 - filterY;
+
+    // Loop over major, minor, and X.
+    for (int majorIdx = 0, major = majorBase; majorIdx < p.loopMajor & major < p.sizeMajor; majorIdx++, major++)
+    for (int minorIdx = 0, minor = minorBase; minorIdx < p.loopMinor & minor < p.sizeMinor; minorIdx++, minor += p.launchMinor)
+    {
+        int nc = major * p.sizeMinor + minor;
+        int n = nc / p.inSize.z;
+        int c = nc - n * p.inSize.z;
+        for (int loopX = 0, outX = outXBase; loopX < p.loopX & outX < p.outSize.x; loopX++, outX += blockDim.y)
+        {
+            // Setup X receptive field.
+            int midX = outX * p.down.x + p.up.x - 1 - p.pad0.x;
+            int inX = min(max(floor_div(midX, p.up.x), 0), p.inSize.x);
+            int w = min(max(floor_div(midX + p.filterSize.x, p.up.x), 0), p.inSize.x) - inX;
+            int filterX = midX + p.filterSize.x - (inX + 1) * p.up.x;
+            if (p.flip)
+                filterX = p.filterSize.x - 1 - filterX;
+
+            // Initialize pointers.
+            const T* xp = &((const T*)p.x)[inX * p.inStride.x + inY * p.inStride.y + c * p.inStride.z + n * p.inStride.w];
+            const float* fp = &p.f[filterX * p.filterStride.x + filterY * p.filterStride.y];
+            int filterStepX = ((p.flip) ? p.up.x : -p.up.x) * p.filterStride.x;
+            int filterStepY = ((p.flip) ? p.up.y : -p.up.y) * p.filterStride.y;
+
+            // Inner loop.
+            scalar_t v = 0;
+            for (int y = 0; y < h; y++)
+            {
+                for (int x = 0; x < w; x++)
+                {
+                    v += (scalar_t)(*xp) * (scalar_t)(*fp);
+                    xp += p.inStride.x;
+                    fp += filterStepX;
+                }
+                xp += p.inStride.y - w * p.inStride.x;
+                fp += filterStepY - w * filterStepX;
+            }
+
+            // Store result.
+            v *= p.gain;
+            ((T*)p.y)[outX * p.outStride.x + outY * p.outStride.y + c * p.outStride.z + n * p.outStride.w] = (T)v;
+        }
+    }
+}
+
+//------------------------------------------------------------------------
+// Specialized CUDA implementation for small filters.
+
+template <class T, int upx, int upy, int downx, int downy, int filterW, int filterH, int tileOutW, int tileOutH, int loopMinor>
+static __global__ void upfirdn2d_kernel_small(upfirdn2d_kernel_params p)
+{
+    typedef typename InternalType<T>::scalar_t scalar_t;
+    const int tileInW = ((tileOutW - 1) * downx + filterW - 1) / upx + 1;
+    const int tileInH = ((tileOutH - 1) * downy + filterH - 1) / upy + 1;
+    __shared__ volatile scalar_t sf[filterH][filterW];
+    __shared__ volatile scalar_t sx[tileInH][tileInW][loopMinor];
+
+    // Calculate tile index.
+    int minorBase = blockIdx.x;
+    int tileOutY = minorBase / p.launchMinor;
+    minorBase -= tileOutY * p.launchMinor;
+    minorBase *= loopMinor;
+    tileOutY *= tileOutH;
+    int tileOutXBase = blockIdx.y * p.loopX * tileOutW;
+    int majorBase = blockIdx.z * p.loopMajor;
+    if (tileOutXBase >= p.outSize.x | tileOutY >= p.outSize.y | majorBase >= p.sizeMajor)
+        return;
+
+    // Load filter (flipped).
+    for (int tapIdx = threadIdx.x; tapIdx < filterH * filterW; tapIdx += blockDim.x)
+    {
+        int fy = tapIdx / filterW;
+        int fx = tapIdx - fy * filterW;
+        scalar_t v = 0;
+        if (fx < p.filterSize.x & fy < p.filterSize.y)
+        {
+            int ffx = (p.flip) ? fx : p.filterSize.x - 1 - fx;
+            int ffy = (p.flip) ? fy : p.filterSize.y - 1 - fy;
+            v = (scalar_t)p.f[ffx * p.filterStride.x + ffy * p.filterStride.y];
+        }
+        sf[fy][fx] = v;
+    }
+
+    // Loop over major and X.
+    for (int majorIdx = 0, major = majorBase; majorIdx < p.loopMajor & major < p.sizeMajor; majorIdx++, major++)
+    {
+        int baseNC = major * p.sizeMinor + minorBase;
+        int n = baseNC / p.inSize.z;
+        int baseC = baseNC - n * p.inSize.z;
+        for (int loopX = 0, tileOutX = tileOutXBase; loopX < p.loopX & tileOutX < p.outSize.x; loopX++, tileOutX += tileOutW)
+        {
+            // Load input pixels.
+            int tileMidX = tileOutX * downx + upx - 1 - p.pad0.x;
+            int tileMidY = tileOutY * downy + upy - 1 - p.pad0.y;
+            int tileInX = floor_div(tileMidX, upx);
+            int tileInY = floor_div(tileMidY, upy);
+            __syncthreads();
+            for (int inIdx = threadIdx.x; inIdx < tileInH * tileInW * loopMinor; inIdx += blockDim.x)
+            {
+                int relC = inIdx;
+                int relInX = relC / loopMinor;
+                int relInY = relInX / tileInW;
+                relC -= relInX * loopMinor;
+                relInX -= relInY * tileInW;
+                int c = baseC + relC;
+                int inX = tileInX + relInX;
+                int inY = tileInY + relInY;
+                scalar_t v = 0;
+                if (inX >= 0 & inY >= 0 & inX < p.inSize.x & inY < p.inSize.y & c < p.inSize.z)
+                    v = (scalar_t)((const T*)p.x)[inX * p.inStride.x + inY * p.inStride.y + c * p.inStride.z + n * p.inStride.w];
+                sx[relInY][relInX][relC] = v;
+            }
+
+            // Loop over output pixels.
+            __syncthreads();
+            for (int outIdx = threadIdx.x; outIdx < tileOutH * tileOutW * loopMinor; outIdx += blockDim.x)
+            {
+                int relC = outIdx;
+                int relOutX = relC / loopMinor;
+                int relOutY = relOutX / tileOutW;
+                relC -= relOutX * loopMinor;
+                relOutX -= relOutY * tileOutW;
+                int c = baseC + relC;
+                int outX = tileOutX + relOutX;
+                int outY = tileOutY + relOutY;
+
+                // Setup receptive field.
+                int midX = tileMidX + relOutX * downx;
+                int midY = tileMidY + relOutY * downy;
+                int inX = floor_div(midX, upx);
+                int inY = floor_div(midY, upy);
+                int relInX = inX - tileInX;
+                int relInY = inY - tileInY;
+                int filterX = (inX + 1) * upx - midX - 1; // flipped
+                int filterY = (inY + 1) * upy - midY - 1; // flipped
+
+                // Inner loop.
+                if (outX < p.outSize.x & outY < p.outSize.y & c < p.outSize.z)
+                {
+                    scalar_t v = 0;
+                    #pragma unroll
+                    for (int y = 0; y < filterH / upy; y++)
+                        #pragma unroll
+                        for (int x = 0; x < filterW / upx; x++)
+                            v += sx[relInY + y][relInX + x][relC] * sf[filterY + y * upy][filterX + x * upx];
+                    v *= p.gain;
+                    ((T*)p.y)[outX * p.outStride.x + outY * p.outStride.y + c * p.outStride.z + n * p.outStride.w] = (T)v;
+                }
+            }
+        }
+    }
+}
+
+//------------------------------------------------------------------------
+// CUDA kernel selection.
+
+template <class T> upfirdn2d_kernel_spec choose_upfirdn2d_kernel(const upfirdn2d_kernel_params& p)
+{
+    int s = p.inStride.z, fx = p.filterSize.x, fy = p.filterSize.y;
+    upfirdn2d_kernel_spec spec = {(void*)upfirdn2d_kernel_large<T>, -1,-1,1, 4}; // contiguous
+    if (s == 1)           spec = {(void*)upfirdn2d_kernel_large<T>, -1,-1,4, 1}; // channels_last
+
+    // No up/downsampling.
+    if (p.up.x == 1 && p.up.y == 1 && p.down.x == 1 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 24 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 24,24, 64,32,1>, 64,32,1, 1};
+        if (s != 1 && fx <= 16 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 16,16, 64,32,1>, 64,32,1, 1};
+        if (s != 1 && fx <= 7  && fy <= 7 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 7,7,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 6,6,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 5  && fy <= 5 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 5,5,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 4,4,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 3  && fy <= 3 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 3,3,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 24 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 24,1,  128,8,1>, 128,8,1, 1};
+        if (s != 1 && fx <= 16 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 16,1,  128,8,1>, 128,8,1, 1};
+        if (s != 1 && fx <= 8  && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 8,1,   128,8,1>, 128,8,1, 1};
+        if (s != 1 && fx <= 1  && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,24,  32,32,1>, 32,32,1, 1};
+        if (s != 1 && fx <= 1  && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,16,  32,32,1>, 32,32,1, 1};
+        if (s != 1 && fx <= 1  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,8,   32,32,1>, 32,32,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 24 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 24,24, 32,32,1>,  32,32,1,  1};
+        if (s == 1 && fx <= 16 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 16,16, 32,32,1>,  32,32,1,  1};
+        if (s == 1 && fx <= 7  && fy <= 7 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 7,7,   16,16,8>,  16,16,8,  1};
+        if (s == 1 && fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 6,6,   16,16,8>,  16,16,8,  1};
+        if (s == 1 && fx <= 5  && fy <= 5 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 5,5,   16,16,8>,  16,16,8,  1};
+        if (s == 1 && fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 4,4,   16,16,8>,  16,16,8,  1};
+        if (s == 1 && fx <= 3  && fy <= 3 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 3,3,   16,16,8>,  16,16,8,  1};
+        if (s == 1 && fx <= 24 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 24,1,  128,1,16>, 128,1,16, 1};
+        if (s == 1 && fx <= 16 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 16,1,  128,1,16>, 128,1,16, 1};
+        if (s == 1 && fx <= 8  && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 8,1,   128,1,16>, 128,1,16, 1};
+        if (s == 1 && fx <= 1  && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,24,  1,128,16>, 1,128,16, 1};
+        if (s == 1 && fx <= 1  && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,16,  1,128,16>, 1,128,16, 1};
+        if (s == 1 && fx <= 1  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,8,   1,128,16>, 1,128,16, 1};
+    }
+
+    // 2x upsampling.
+    if (p.up.x == 2 && p.up.y == 2 && p.down.x == 1 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 24 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 24,24, 64,32,1>, 64,32,1, 1};
+        if (s != 1 && fx <= 16 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 16,16, 64,32,1>, 64,32,1, 1};
+        if (s != 1 && fx <= 8  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 8,8,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 6,6,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 4,4,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 2  && fy <= 2 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 2,2,   64,16,1>, 64,16,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 24 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 24,24, 32,32,1>, 32,32,1, 1};
+        if (s == 1 && fx <= 16 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 16,16, 32,32,1>, 32,32,1, 1};
+        if (s == 1 && fx <= 8  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 8,8,   16,16,8>, 16,16,8, 1};
+        if (s == 1 && fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 6,6,   16,16,8>, 16,16,8, 1};
+        if (s == 1 && fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 4,4,   16,16,8>, 16,16,8, 1};
+        if (s == 1 && fx <= 2  && fy <= 2 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 2,2,   16,16,8>, 16,16,8, 1};
+    }
+    if (p.up.x == 2 && p.up.y == 1 && p.down.x == 1 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 24 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 24,1, 128,8,1>, 128,8,1, 1};
+        if (s != 1 && fx <= 16 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 16,1, 128,8,1>, 128,8,1, 1};
+        if (s != 1 && fx <= 8  && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 8,1,  128,8,1>, 128,8,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 24 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 24,1, 128,1,16>, 128,1,16, 1};
+        if (s == 1 && fx <= 16 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 16,1, 128,1,16>, 128,1,16, 1};
+        if (s == 1 && fx <= 8  && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 8,1,  128,1,16>, 128,1,16, 1};
+    }
+    if (p.up.x == 1 && p.up.y == 2 && p.down.x == 1 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 1 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,24, 32,32,1>, 32,32,1, 1};
+        if (s != 1 && fx <= 1 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,16, 32,32,1>, 32,32,1, 1};
+        if (s != 1 && fx <= 1 && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,8,  32,32,1>, 32,32,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 1 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,24, 1,128,16>, 1,128,16, 1};
+        if (s == 1 && fx <= 1 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,16, 1,128,16>, 1,128,16, 1};
+        if (s == 1 && fx <= 1 && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,8,  1,128,16>, 1,128,16, 1};
+    }
+
+    // 2x downsampling.
+    if (p.up.x == 1 && p.up.y == 1 && p.down.x == 2 && p.down.y == 2)
+    {
+        // contiguous
+        if (s != 1 && fx <= 24 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 24,24, 32,16,1>, 32,16,1, 1};
+        if (s != 1 && fx <= 16 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 16,16, 32,16,1>, 32,16,1, 1};
+        if (s != 1 && fx <= 8  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 8,8,   32,8,1>,  32,8,1,  1};
+        if (s != 1 && fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 6,6,   32,8,1>,  32,8,1,  1};
+        if (s != 1 && fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 4,4,   32,8,1>,  32,8,1,  1};
+        if (s != 1 && fx <= 2  && fy <= 2 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 2,2,   32,8,1>,  32,8,1,  1};
+        // channels_last
+        if (s == 1 && fx <= 24 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 24,24, 16,16,1>, 16,16,1, 1};
+        if (s == 1 && fx <= 16 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 16,16, 16,16,1>, 16,16,1, 1};
+        if (s == 1 && fx <= 8  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 8,8,   8,8,8>,   8,8,8,   1};
+        if (s == 1 && fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 6,6,   8,8,8>,   8,8,8,   1};
+        if (s == 1 && fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 4,4,   8,8,8>,   8,8,8,   1};
+        if (s == 1 && fx <= 2  && fy <= 2 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 2,2,   8,8,8>,   8,8,8,   1};
+    }
+    if (p.up.x == 1 && p.up.y == 1 && p.down.x == 2 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 24 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 24,1, 64,8,1>, 64,8,1, 1};
+        if (s != 1 && fx <= 16 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 16,1, 64,8,1>, 64,8,1, 1};
+        if (s != 1 && fx <= 8  && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 8,1,  64,8,1>, 64,8,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 24 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 24,1, 64,1,8>, 64,1,8, 1};
+        if (s == 1 && fx <= 16 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 16,1, 64,1,8>, 64,1,8, 1};
+        if (s == 1 && fx <= 8  && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 8,1,  64,1,8>, 64,1,8, 1};
+    }
+    if (p.up.x == 1 && p.up.y == 1 && p.down.x == 1 && p.down.y == 2)
+    {
+        // contiguous
+        if (s != 1 && fx <= 1 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,24, 32,16,1>, 32,16,1, 1};
+        if (s != 1 && fx <= 1 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,16, 32,16,1>, 32,16,1, 1};
+        if (s != 1 && fx <= 1 && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,8,  32,16,1>, 32,16,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 1  && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,24, 1,64,8>, 1,64,8, 1};
+        if (s == 1 && fx <= 1  && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,16, 1,64,8>, 1,64,8, 1};
+        if (s == 1 && fx <= 1  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,8,  1,64,8>, 1,64,8, 1};
+    }
+
+    // 4x upsampling.
+    if (p.up.x == 4 && p.up.y == 4 && p.down.x == 1 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 48 && fy <= 48) spec = {(void*)upfirdn2d_kernel_small<T, 4,4, 1,1, 48,48, 64,32,1>, 64,32,1, 1};
+        if (s != 1 && fx <= 32 && fy <= 32) spec = {(void*)upfirdn2d_kernel_small<T, 4,4, 1,1, 32,32, 64,32,1>, 64,32,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 48 && fy <= 48) spec = {(void*)upfirdn2d_kernel_small<T, 4,4, 1,1, 48,48, 32,32,1>, 32,32,1, 1};
+        if (s == 1 && fx <= 32 && fy <= 32) spec = {(void*)upfirdn2d_kernel_small<T, 4,4, 1,1, 32,32, 32,32,1>, 32,32,1, 1};
+    }
+    if (p.up.x == 4 && p.up.y == 1 && p.down.x == 1 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 48 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 4,1, 1,1, 48,1, 128,8,1>, 128,8,1, 1};
+        if (s != 1 && fx <= 32 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 4,1, 1,1, 32,1, 128,8,1>, 128,8,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 48 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 4,1, 1,1, 48,1, 128,1,16>, 128,1,16, 1};
+        if (s == 1 && fx <= 32 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 4,1, 1,1, 32,1, 128,1,16>, 128,1,16, 1};
+    }
+    if (p.up.x == 1 && p.up.y == 4 && p.down.x == 1 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 1 && fy <= 48) spec = {(void*)upfirdn2d_kernel_small<T, 1,4, 1,1, 1,48, 32,32,1>, 32,32,1, 1};
+        if (s != 1 && fx <= 1 && fy <= 32) spec = {(void*)upfirdn2d_kernel_small<T, 1,4, 1,1, 1,32, 32,32,1>, 32,32,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 1 && fy <= 48) spec = {(void*)upfirdn2d_kernel_small<T, 1,4, 1,1, 1,48, 1,128,16>, 1,128,16, 1};
+        if (s == 1 && fx <= 1 && fy <= 32) spec = {(void*)upfirdn2d_kernel_small<T, 1,4, 1,1, 1,32, 1,128,16>, 1,128,16, 1};
+    }
+
+    // 4x downsampling (inefficient).
+    if (p.up.x == 1 && p.up.y == 1 && p.down.x == 4 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 48 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 4,1, 48,1, 32,8,1>, 32,8,1, 1};
+        if (s != 1 && fx <= 32 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 4,1, 32,1, 32,8,1>, 32,8,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 48 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 4,1, 48,1, 32,1,8>, 32,1,8, 1};
+        if (s == 1 && fx <= 32 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 4,1, 32,1, 32,1,8>, 32,1,8, 1};
+    }
+    if (p.up.x == 1 && p.up.y == 1 && p.down.x == 1 && p.down.y == 4)
+    {
+        // contiguous
+        if (s != 1 && fx <= 1 && fy <= 48) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,4, 1,48, 32,8,1>, 32,8,1, 1};
+        if (s != 1 && fx <= 1 && fy <= 32) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,4, 1,32, 32,8,1>, 32,8,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 1  && fy <= 48) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,4, 1,48, 1,32,8>, 1,32,8, 1};
+        if (s == 1 && fx <= 1  && fy <= 32) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,4, 1,32, 1,32,8>, 1,32,8, 1};
+    }
+    return spec;
+}
+
+//------------------------------------------------------------------------
+// Template specializations.
+
+template upfirdn2d_kernel_spec choose_upfirdn2d_kernel<double>   (const upfirdn2d_kernel_params& p);
+template upfirdn2d_kernel_spec choose_upfirdn2d_kernel<float>    (const upfirdn2d_kernel_params& p);
+template upfirdn2d_kernel_spec choose_upfirdn2d_kernel<c10::Half>(const upfirdn2d_kernel_params& p);
+
+//------------------------------------------------------------------------

models/torch_utils/ops/upfirdn2d.h

@@ -0,0 +1,59 @@
+// Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#include <cuda_runtime.h>
+
+//------------------------------------------------------------------------
+// CUDA kernel parameters.
+
+struct upfirdn2d_kernel_params
+{
+    const void*     x;
+    const float*    f;
+    void*           y;
+
+    int2            up;
+    int2            down;
+    int2            pad0;
+    int             flip;
+    float           gain;
+
+    int4            inSize;         // [width, height, channel, batch]
+    int4            inStride;
+    int2            filterSize;     // [width, height]
+    int2            filterStride;
+    int4            outSize;        // [width, height, channel, batch]
+    int4            outStride;
+    int             sizeMinor;
+    int             sizeMajor;
+
+    int             loopMinor;
+    int             loopMajor;
+    int             loopX;
+    int             launchMinor;
+    int             launchMajor;
+};
+
+//------------------------------------------------------------------------
+// CUDA kernel specialization.
+
+struct upfirdn2d_kernel_spec
+{
+    void*   kernel;
+    int     tileOutW;
+    int     tileOutH;
+    int     loopMinor;
+    int     loopX;
+};
+
+//------------------------------------------------------------------------
+// CUDA kernel selection.
+
+template <class T> upfirdn2d_kernel_spec choose_upfirdn2d_kernel(const upfirdn2d_kernel_params& p);
+
+//------------------------------------------------------------------------

models/torch_utils/ops/upfirdn2d.py

@@ -0,0 +1,389 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Custom PyTorch ops for efficient resampling of 2D images."""
+
+import os
+import numpy as np
+import torch
+
+from .. import custom_ops
+from .. import misc
+from . import conv2d_gradfix
+
+#----------------------------------------------------------------------------
+
+_plugin = None
+
+def _init():
+    global _plugin
+    if _plugin is None:
+        _plugin = custom_ops.get_plugin(
+            module_name='upfirdn2d_plugin',
+            sources=['upfirdn2d.cpp', 'upfirdn2d.cu'],
+            headers=['upfirdn2d.h'],
+            source_dir=os.path.dirname(__file__),
+            extra_cuda_cflags=['--use_fast_math'],
+        )
+    return True
+
+def _parse_scaling(scaling):
+    if isinstance(scaling, int):
+        scaling = [scaling, scaling]
+    assert isinstance(scaling, (list, tuple))
+    assert all(isinstance(x, int) for x in scaling)
+    sx, sy = scaling
+    assert sx >= 1 and sy >= 1
+    return sx, sy
+
+def _parse_padding(padding):
+    if isinstance(padding, int):
+        padding = [padding, padding]
+    assert isinstance(padding, (list, tuple))
+    assert all(isinstance(x, int) for x in padding)
+    if len(padding) == 2:
+        padx, pady = padding
+        padding = [padx, padx, pady, pady]
+    padx0, padx1, pady0, pady1 = padding
+    return padx0, padx1, pady0, pady1
+
+def _get_filter_size(f):
+    if f is None:
+        return 1, 1
+    assert isinstance(f, torch.Tensor) and f.ndim in [1, 2]
+    fw = f.shape[-1]
+    fh = f.shape[0]
+    with misc.suppress_tracer_warnings():
+        fw = int(fw)
+        fh = int(fh)
+    misc.assert_shape(f, [fh, fw][:f.ndim])
+    assert fw >= 1 and fh >= 1
+    return fw, fh
+
+#----------------------------------------------------------------------------
+
+def setup_filter(f, device=torch.device('cpu'), normalize=True, flip_filter=False, gain=1, separable=None):
+    r"""Convenience function to setup 2D FIR filter for `upfirdn2d()`.
+
+    Args:
+        f:           Torch tensor, numpy array, or python list of the shape
+                     `[filter_height, filter_width]` (non-separable),
+                     `[filter_taps]` (separable),
+                     `[]` (impulse), or
+                     `None` (identity).
+        device:      Result device (default: cpu).
+        normalize:   Normalize the filter so that it retains the magnitude
+                     for constant input signal (DC)? (default: True).
+        flip_filter: Flip the filter? (default: False).
+        gain:        Overall scaling factor for signal magnitude (default: 1).
+        separable:   Return a separable filter? (default: select automatically).
+
+    Returns:
+        Float32 tensor of the shape
+        `[filter_height, filter_width]` (non-separable) or
+        `[filter_taps]` (separable).
+    """
+    # Validate.
+    if f is None:
+        f = 1
+    f = torch.as_tensor(f, dtype=torch.float32)
+    assert f.ndim in [0, 1, 2]
+    assert f.numel() > 0
+    if f.ndim == 0:
+        f = f[np.newaxis]
+
+    # Separable?
+    if separable is None:
+        separable = (f.ndim == 1 and f.numel() >= 8)
+    if f.ndim == 1 and not separable:
+        f = f.ger(f)
+    assert f.ndim == (1 if separable else 2)
+
+    # Apply normalize, flip, gain, and device.
+    if normalize:
+        f /= f.sum()
+    if flip_filter:
+        f = f.flip(list(range(f.ndim)))
+    f = f * (gain ** (f.ndim / 2))
+    f = f.to(device=device)
+    return f
+
+#----------------------------------------------------------------------------
+
+def upfirdn2d(x, f, up=1, down=1, padding=0, flip_filter=False, gain=1, impl='cuda'):
+    r"""Pad, upsample, filter, and downsample a batch of 2D images.
+
+    Performs the following sequence of operations for each channel:
+
+    1. Upsample the image by inserting N-1 zeros after each pixel (`up`).
+
+    2. Pad the image with the specified number of zeros on each side (`padding`).
+       Negative padding corresponds to cropping the image.
+
+    3. Convolve the image with the specified 2D FIR filter (`f`), shrinking it
+       so that the footprint of all output pixels lies within the input image.
+
+    4. Downsample the image by keeping every Nth pixel (`down`).
+
+    This sequence of operations bears close resemblance to scipy.signal.upfirdn().
+    The fused op is considerably more efficient than performing the same calculation
+    using standard PyTorch ops. It supports gradients of arbitrary order.
+
+    Args:
+        x:           Float32/float64/float16 input tensor of the shape
+                     `[batch_size, num_channels, in_height, in_width]`.
+        f:           Float32 FIR filter of the shape
+                     `[filter_height, filter_width]` (non-separable),
+                     `[filter_taps]` (separable), or
+                     `None` (identity).
+        up:          Integer upsampling factor. Can be a single int or a list/tuple
+                     `[x, y]` (default: 1).
+        down:        Integer downsampling factor. Can be a single int or a list/tuple
+                     `[x, y]` (default: 1).
+        padding:     Padding with respect to the upsampled image. Can be a single number
+                     or a list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
+                     (default: 0).
+        flip_filter: False = convolution, True = correlation (default: False).
+        gain:        Overall scaling factor for signal magnitude (default: 1).
+        impl:        Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
+
+    Returns:
+        Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
+    """
+    assert isinstance(x, torch.Tensor)
+    assert impl in ['ref', 'cuda']
+    if impl == 'cuda' and x.device.type == 'cuda' and _init():
+        return _upfirdn2d_cuda(up=up, down=down, padding=padding, flip_filter=flip_filter, gain=gain).apply(x, f)
+    return _upfirdn2d_ref(x, f, up=up, down=down, padding=padding, flip_filter=flip_filter, gain=gain)
+
+#----------------------------------------------------------------------------
+
+@misc.profiled_function
+def _upfirdn2d_ref(x, f, up=1, down=1, padding=0, flip_filter=False, gain=1):
+    """Slow reference implementation of `upfirdn2d()` using standard PyTorch ops.
+    """
+    # Validate arguments.
+    assert isinstance(x, torch.Tensor) and x.ndim == 4
+    if f is None:
+        f = torch.ones([1, 1], dtype=torch.float32, device=x.device)
+    assert isinstance(f, torch.Tensor) and f.ndim in [1, 2]
+    assert f.dtype == torch.float32 and not f.requires_grad
+    batch_size, num_channels, in_height, in_width = x.shape
+    upx, upy = _parse_scaling(up)
+    downx, downy = _parse_scaling(down)
+    padx0, padx1, pady0, pady1 = _parse_padding(padding)
+
+    # Check that upsampled buffer is not smaller than the filter.
+    upW = in_width * upx + padx0 + padx1
+    upH = in_height * upy + pady0 + pady1
+    assert upW >= f.shape[-1] and upH >= f.shape[0]
+
+    # Upsample by inserting zeros.
+    x = x.reshape([batch_size, num_channels, in_height, 1, in_width, 1])
+    x = torch.nn.functional.pad(x, [0, upx - 1, 0, 0, 0, upy - 1])
+    x = x.reshape([batch_size, num_channels, in_height * upy, in_width * upx])
+
+    # Pad or crop.
+    x = torch.nn.functional.pad(x, [max(padx0, 0), max(padx1, 0), max(pady0, 0), max(pady1, 0)])
+    x = x[:, :, max(-pady0, 0) : x.shape[2] - max(-pady1, 0), max(-padx0, 0) : x.shape[3] - max(-padx1, 0)]
+
+    # Setup filter.
+    f = f * (gain ** (f.ndim / 2))
+    f = f.to(x.dtype)
+    if not flip_filter:
+        f = f.flip(list(range(f.ndim)))
+
+    # Convolve with the filter.
+    f = f[np.newaxis, np.newaxis].repeat([num_channels, 1] + [1] * f.ndim)
+    if f.ndim == 4:
+        x = conv2d_gradfix.conv2d(input=x, weight=f, groups=num_channels)
+    else:
+        x = conv2d_gradfix.conv2d(input=x, weight=f.unsqueeze(2), groups=num_channels)
+        x = conv2d_gradfix.conv2d(input=x, weight=f.unsqueeze(3), groups=num_channels)
+
+    # Downsample by throwing away pixels.
+    x = x[:, :, ::downy, ::downx]
+    return x
+
+#----------------------------------------------------------------------------
+
+_upfirdn2d_cuda_cache = dict()
+
+def _upfirdn2d_cuda(up=1, down=1, padding=0, flip_filter=False, gain=1):
+    """Fast CUDA implementation of `upfirdn2d()` using custom ops.
+    """
+    # Parse arguments.
+    upx, upy = _parse_scaling(up)
+    downx, downy = _parse_scaling(down)
+    padx0, padx1, pady0, pady1 = _parse_padding(padding)
+
+    # Lookup from cache.
+    key = (upx, upy, downx, downy, padx0, padx1, pady0, pady1, flip_filter, gain)
+    if key in _upfirdn2d_cuda_cache:
+        return _upfirdn2d_cuda_cache[key]
+
+    # Forward op.
+    class Upfirdn2dCuda(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, x, f): # pylint: disable=arguments-differ
+            assert isinstance(x, torch.Tensor) and x.ndim == 4
+            if f is None:
+                f = torch.ones([1, 1], dtype=torch.float32, device=x.device)
+            if f.ndim == 1 and f.shape[0] == 1:
+                f = f.square().unsqueeze(0) # Convert separable-1 into full-1x1.
+            assert isinstance(f, torch.Tensor) and f.ndim in [1, 2]
+            y = x
+            if f.ndim == 2:
+                y = _plugin.upfirdn2d(y, f, upx, upy, downx, downy, padx0, padx1, pady0, pady1, flip_filter, gain)
+            else:
+                y = _plugin.upfirdn2d(y, f.unsqueeze(0), upx, 1, downx, 1, padx0, padx1, 0, 0, flip_filter, 1.0)
+                y = _plugin.upfirdn2d(y, f.unsqueeze(1), 1, upy, 1, downy, 0, 0, pady0, pady1, flip_filter, gain)
+            ctx.save_for_backward(f)
+            ctx.x_shape = x.shape
+            return y
+
+        @staticmethod
+        def backward(ctx, dy): # pylint: disable=arguments-differ
+            f, = ctx.saved_tensors
+            _, _, ih, iw = ctx.x_shape
+            _, _, oh, ow = dy.shape
+            fw, fh = _get_filter_size(f)
+            p = [
+                fw - padx0 - 1,
+                iw * upx - ow * downx + padx0 - upx + 1,
+                fh - pady0 - 1,
+                ih * upy - oh * downy + pady0 - upy + 1,
+            ]
+            dx = None
+            df = None
+
+            if ctx.needs_input_grad[0]:
+                dx = _upfirdn2d_cuda(up=down, down=up, padding=p, flip_filter=(not flip_filter), gain=gain).apply(dy, f)
+
+            assert not ctx.needs_input_grad[1]
+            return dx, df
+
+    # Add to cache.
+    _upfirdn2d_cuda_cache[key] = Upfirdn2dCuda
+    return Upfirdn2dCuda
+
+#----------------------------------------------------------------------------
+
+def filter2d(x, f, padding=0, flip_filter=False, gain=1, impl='cuda'):
+    r"""Filter a batch of 2D images using the given 2D FIR filter.
+
+    By default, the result is padded so that its shape matches the input.
+    User-specified padding is applied on top of that, with negative values
+    indicating cropping. Pixels outside the image are assumed to be zero.
+
+    Args:
+        x:           Float32/float64/float16 input tensor of the shape
+                     `[batch_size, num_channels, in_height, in_width]`.
+        f:           Float32 FIR filter of the shape
+                     `[filter_height, filter_width]` (non-separable),
+                     `[filter_taps]` (separable), or
+                     `None` (identity).
+        padding:     Padding with respect to the output. Can be a single number or a
+                     list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
+                     (default: 0).
+        flip_filter: False = convolution, True = correlation (default: False).
+        gain:        Overall scaling factor for signal magnitude (default: 1).
+        impl:        Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
+
+    Returns:
+        Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
+    """
+    padx0, padx1, pady0, pady1 = _parse_padding(padding)
+    fw, fh = _get_filter_size(f)
+    p = [
+        padx0 + fw // 2,
+        padx1 + (fw - 1) // 2,
+        pady0 + fh // 2,
+        pady1 + (fh - 1) // 2,
+    ]
+    return upfirdn2d(x, f, padding=p, flip_filter=flip_filter, gain=gain, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def upsample2d(x, f, up=2, padding=0, flip_filter=False, gain=1, impl='cuda'):
+    r"""Upsample a batch of 2D images using the given 2D FIR filter.
+
+    By default, the result is padded so that its shape is a multiple of the input.
+    User-specified padding is applied on top of that, with negative values
+    indicating cropping. Pixels outside the image are assumed to be zero.
+
+    Args:
+        x:           Float32/float64/float16 input tensor of the shape
+                     `[batch_size, num_channels, in_height, in_width]`.
+        f:           Float32 FIR filter of the shape
+                     `[filter_height, filter_width]` (non-separable),
+                     `[filter_taps]` (separable), or
+                     `None` (identity).
+        up:          Integer upsampling factor. Can be a single int or a list/tuple
+                     `[x, y]` (default: 1).
+        padding:     Padding with respect to the output. Can be a single number or a
+                     list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
+                     (default: 0).
+        flip_filter: False = convolution, True = correlation (default: False).
+        gain:        Overall scaling factor for signal magnitude (default: 1).
+        impl:        Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
+
+    Returns:
+        Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
+    """
+    upx, upy = _parse_scaling(up)
+    padx0, padx1, pady0, pady1 = _parse_padding(padding)
+    fw, fh = _get_filter_size(f)
+    p = [
+        padx0 + (fw + upx - 1) // 2,
+        padx1 + (fw - upx) // 2,
+        pady0 + (fh + upy - 1) // 2,
+        pady1 + (fh - upy) // 2,
+    ]
+    return upfirdn2d(x, f, up=up, padding=p, flip_filter=flip_filter, gain=gain*upx*upy, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def downsample2d(x, f, down=2, padding=0, flip_filter=False, gain=1, impl='cuda'):
+    r"""Downsample a batch of 2D images using the given 2D FIR filter.
+
+    By default, the result is padded so that its shape is a fraction of the input.
+    User-specified padding is applied on top of that, with negative values
+    indicating cropping. Pixels outside the image are assumed to be zero.
+
+    Args:
+        x:           Float32/float64/float16 input tensor of the shape
+                     `[batch_size, num_channels, in_height, in_width]`.
+        f:           Float32 FIR filter of the shape
+                     `[filter_height, filter_width]` (non-separable),
+                     `[filter_taps]` (separable), or
+                     `None` (identity).
+        down:        Integer downsampling factor. Can be a single int or a list/tuple
+                     `[x, y]` (default: 1).
+        padding:     Padding with respect to the input. Can be a single number or a
+                     list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
+                     (default: 0).
+        flip_filter: False = convolution, True = correlation (default: False).
+        gain:        Overall scaling factor for signal magnitude (default: 1).
+        impl:        Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
+
+    Returns:
+        Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
+    """
+    downx, downy = _parse_scaling(down)
+    padx0, padx1, pady0, pady1 = _parse_padding(padding)
+    fw, fh = _get_filter_size(f)
+    p = [
+        padx0 + (fw - downx + 1) // 2,
+        padx1 + (fw - downx) // 2,
+        pady0 + (fh - downy + 1) // 2,
+        pady1 + (fh - downy) // 2,
+    ]
+    return upfirdn2d(x, f, down=down, padding=p, flip_filter=flip_filter, gain=gain, impl=impl)
+
+#----------------------------------------------------------------------------

models/torch_utils/persistence.py

@@ -0,0 +1,251 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Facilities for pickling Python code alongside other data.
+
+The pickled code is automatically imported into a separate Python module
+during unpickling. This way, any previously exported pickles will remain
+usable even if the original code is no longer available, or if the current
+version of the code is not consistent with what was originally pickled."""
+
+import sys
+import pickle
+import io
+import inspect
+import copy
+import uuid
+import types
+# import dnnlib
+
+#----------------------------------------------------------------------------
+
+_version            = 6         # internal version number
+_decorators         = set()     # {decorator_class, ...}
+_import_hooks       = []        # [hook_function, ...]
+_module_to_src_dict = dict()    # {module: src, ...}
+_src_to_module_dict = dict()    # {src: module, ...}
+
+#----------------------------------------------------------------------------
+
+def persistent_class(orig_class):
+    r"""Class decorator that extends a given class to save its source code
+    when pickled.
+
+    Example:
+
+        from torch_utils import persistence
+
+        @persistence.persistent_class
+        class MyNetwork(torch.nn.Module):
+            def __init__(self, num_inputs, num_outputs):
+                super().__init__()
+                self.fc = MyLayer(num_inputs, num_outputs)
+                ...
+
+        @persistence.persistent_class
+        class MyLayer(torch.nn.Module):
+            ...
+
+    When pickled, any instance of `MyNetwork` and `MyLayer` will save its
+    source code alongside other internal state (e.g., parameters, buffers,
+    and submodules). This way, any previously exported pickle will remain
+    usable even if the class definitions have been modified or are no
+    longer available.
+
+    The decorator saves the source code of the entire Python module
+    containing the decorated class. It does *not* save the source code of
+    any imported modules. Thus, the imported modules must be available
+    during unpickling, also including `torch_utils.persistence` itself.
+
+    It is ok to call functions defined in the same module from the
+    decorated class. However, if the decorated class depends on other
+    classes defined in the same module, they must be decorated as well.
+    This is illustrated in the above example in the case of `MyLayer`.
+
+    It is also possible to employ the decorator just-in-time before
+    calling the constructor. For example:
+
+        cls = MyLayer
+        if want_to_make_it_persistent:
+            cls = persistence.persistent_class(cls)
+        layer = cls(num_inputs, num_outputs)
+
+    As an additional feature, the decorator also keeps track of the
+    arguments that were used to construct each instance of the decorated
+    class. The arguments can be queried via `obj.init_args` and
+    `obj.init_kwargs`, and they are automatically pickled alongside other
+    object state. A typical use case is to first unpickle a previous
+    instance of a persistent class, and then upgrade it to use the latest
+    version of the source code:
+
+        with open('old_pickle.pkl', 'rb') as f:
+            old_net = pickle.load(f)
+        new_net = MyNetwork(*old_obj.init_args, **old_obj.init_kwargs)
+        misc.copy_params_and_buffers(old_net, new_net, require_all=True)
+    """
+    assert isinstance(orig_class, type)
+    if is_persistent(orig_class):
+        return orig_class
+
+    assert orig_class.__module__ in sys.modules
+    orig_module = sys.modules[orig_class.__module__]
+    orig_module_src = _module_to_src(orig_module)
+
+    class Decorator(orig_class):
+        _orig_module_src = orig_module_src
+        _orig_class_name = orig_class.__name__
+
+        def __init__(self, *args, **kwargs):
+            super().__init__(*args, **kwargs)
+            self._init_args = copy.deepcopy(args)
+            self._init_kwargs = copy.deepcopy(kwargs)
+            assert orig_class.__name__ in orig_module.__dict__
+            _check_pickleable(self.__reduce__())
+
+        @property
+        def init_args(self):
+            return copy.deepcopy(self._init_args)
+
+        @property
+        def init_kwargs(self):
+            return dnnlib.EasyDict(copy.deepcopy(self._init_kwargs))
+
+        def __reduce__(self):
+            fields = list(super().__reduce__())
+            fields += [None] * max(3 - len(fields), 0)
+            if fields[0] is not _reconstruct_persistent_obj:
+                meta = dict(type='class', version=_version, module_src=self._orig_module_src, class_name=self._orig_class_name, state=fields[2])
+                fields[0] = _reconstruct_persistent_obj # reconstruct func
+                fields[1] = (meta,) # reconstruct args
+                fields[2] = None # state dict
+            return tuple(fields)
+
+    Decorator.__name__ = orig_class.__name__
+    _decorators.add(Decorator)
+    return Decorator
+
+#----------------------------------------------------------------------------
+
+def is_persistent(obj):
+    r"""Test whether the given object or class is persistent, i.e.,
+    whether it will save its source code when pickled.
+    """
+    try:
+        if obj in _decorators:
+            return True
+    except TypeError:
+        pass
+    return type(obj) in _decorators # pylint: disable=unidiomatic-typecheck
+
+#----------------------------------------------------------------------------
+
+def import_hook(hook):
+    r"""Register an import hook that is called whenever a persistent object
+    is being unpickled. A typical use case is to patch the pickled source
+    code to avoid errors and inconsistencies when the API of some imported
+    module has changed.
+
+    The hook should have the following signature:
+
+        hook(meta) -> modified meta
+
+    `meta` is an instance of `dnnlib.EasyDict` with the following fields:
+
+        type:       Type of the persistent object, e.g. `'class'`.
+        version:    Internal version number of `torch_utils.persistence`.
+        module_src  Original source code of the Python module.
+        class_name: Class name in the original Python module.
+        state:      Internal state of the object.
+
+    Example:
+
+        @persistence.import_hook
+        def wreck_my_network(meta):
+            if meta.class_name == 'MyNetwork':
+                print('MyNetwork is being imported. I will wreck it!')
+                meta.module_src = meta.module_src.replace("True", "False")
+            return meta
+    """
+    assert callable(hook)
+    _import_hooks.append(hook)
+
+#----------------------------------------------------------------------------
+
+def _reconstruct_persistent_obj(meta):
+    r"""Hook that is called internally by the `pickle` module to unpickle
+    a persistent object.
+    """
+    meta = dnnlib.EasyDict(meta)
+    meta.state = dnnlib.EasyDict(meta.state)
+    for hook in _import_hooks:
+        meta = hook(meta)
+        assert meta is not None
+
+    assert meta.version == _version
+    module = _src_to_module(meta.module_src)
+
+    assert meta.type == 'class'
+    orig_class = module.__dict__[meta.class_name]
+    decorator_class = persistent_class(orig_class)
+    obj = decorator_class.__new__(decorator_class)
+
+    setstate = getattr(obj, '__setstate__', None)
+    if callable(setstate):
+        setstate(meta.state) # pylint: disable=not-callable
+    else:
+        obj.__dict__.update(meta.state)
+    return obj
+
+#----------------------------------------------------------------------------
+
+def _module_to_src(module):
+    r"""Query the source code of a given Python module.
+    """
+    src = _module_to_src_dict.get(module, None)
+    if src is None:
+        src = inspect.getsource(module)
+        _module_to_src_dict[module] = src
+        _src_to_module_dict[src] = module
+    return src
+
+def _src_to_module(src):
+    r"""Get or create a Python module for the given source code.
+    """
+    module = _src_to_module_dict.get(src, None)
+    if module is None:
+        module_name = "_imported_module_" + uuid.uuid4().hex
+        module = types.ModuleType(module_name)
+        sys.modules[module_name] = module
+        _module_to_src_dict[module] = src
+        _src_to_module_dict[src] = module
+        exec(src, module.__dict__) # pylint: disable=exec-used
+    return module
+
+#----------------------------------------------------------------------------
+
+def _check_pickleable(obj):
+    r"""Check that the given object is pickleable, raising an exception if
+    it is not. This function is expected to be considerably more efficient
+    than actually pickling the object.
+    """
+    def recurse(obj):
+        if isinstance(obj, (list, tuple, set)):
+            return [recurse(x) for x in obj]
+        if isinstance(obj, dict):
+            return [[recurse(x), recurse(y)] for x, y in obj.items()]
+        if isinstance(obj, (str, int, float, bool, bytes, bytearray)):
+            return None # Python primitive types are pickleable.
+        if f'{type(obj).__module__}.{type(obj).__name__}' in ['numpy.ndarray', 'torch.Tensor', 'torch.nn.parameter.Parameter']:
+            return None # NumPy arrays and PyTorch tensors are pickleable.
+        if is_persistent(obj):
+            return None # Persistent objects are pickleable, by virtue of the constructor check.
+        return obj
+    with io.BytesIO() as f:
+        pickle.dump(recurse(obj), f)
+
+#----------------------------------------------------------------------------

models/torch_utils/training_stats.py

@@ -0,0 +1,268 @@
+# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Facilities for reporting and collecting training statistics across
+multiple processes and devices. The interface is designed to minimize
+synchronization overhead as well as the amount of boilerplate in user
+code."""
+
+import re
+import numpy as np
+import torch
+import dnnlib
+
+from . import misc
+
+#----------------------------------------------------------------------------
+
+_num_moments    = 3             # [num_scalars, sum_of_scalars, sum_of_squares]
+_reduce_dtype   = torch.float32 # Data type to use for initial per-tensor reduction.
+_counter_dtype  = torch.float64 # Data type to use for the internal counters.
+_rank           = 0             # Rank of the current process.
+_sync_device    = None          # Device to use for multiprocess communication. None = single-process.
+_sync_called    = False         # Has _sync() been called yet?
+_counters       = dict()        # Running counters on each device, updated by report(): name => device => torch.Tensor
+_cumulative     = dict()        # Cumulative counters on the CPU, updated by _sync(): name => torch.Tensor
+
+#----------------------------------------------------------------------------
+
+def init_multiprocessing(rank, sync_device):
+    r"""Initializes `torch_utils.training_stats` for collecting statistics
+    across multiple processes.
+
+    This function must be called after
+    `torch.distributed.init_process_group()` and before `Collector.update()`.
+    The call is not necessary if multi-process collection is not needed.
+
+    Args:
+        rank:           Rank of the current process.
+        sync_device:    PyTorch device to use for inter-process
+                        communication, or None to disable multi-process
+                        collection. Typically `torch.device('cuda', rank)`.
+    """
+    global _rank, _sync_device
+    assert not _sync_called
+    _rank = rank
+    _sync_device = sync_device
+
+#----------------------------------------------------------------------------
+
+@misc.profiled_function
+def report(name, value):
+    r"""Broadcasts the given set of scalars to all interested instances of
+    `Collector`, across device and process boundaries.
+
+    This function is expected to be extremely cheap and can be safely
+    called from anywhere in the training loop, loss function, or inside a
+    `torch.nn.Module`.
+
+    Warning: The current implementation expects the set of unique names to
+    be consistent across processes. Please make sure that `report()` is
+    called at least once for each unique name by each process, and in the
+    same order. If a given process has no scalars to broadcast, it can do
+    `report(name, [])` (empty list).
+
+    Args:
+        name:   Arbitrary string specifying the name of the statistic.
+                Averages are accumulated separately for each unique name.
+        value:  Arbitrary set of scalars. Can be a list, tuple,
+                NumPy array, PyTorch tensor, or Python scalar.
+
+    Returns:
+        The same `value` that was passed in.
+    """
+    if name not in _counters:
+        _counters[name] = dict()
+
+    elems = torch.as_tensor(value)
+    if elems.numel() == 0:
+        return value
+
+    elems = elems.detach().flatten().to(_reduce_dtype)
+    moments = torch.stack([
+        torch.ones_like(elems).sum(),
+        elems.sum(),
+        elems.square().sum(),
+    ])
+    assert moments.ndim == 1 and moments.shape[0] == _num_moments
+    moments = moments.to(_counter_dtype)
+
+    device = moments.device
+    if device not in _counters[name]:
+        _counters[name][device] = torch.zeros_like(moments)
+    _counters[name][device].add_(moments)
+    return value
+
+#----------------------------------------------------------------------------
+
+def report0(name, value):
+    r"""Broadcasts the given set of scalars by the first process (`rank = 0`),
+    but ignores any scalars provided by the other processes.
+    See `report()` for further details.
+    """
+    report(name, value if _rank == 0 else [])
+    return value
+
+#----------------------------------------------------------------------------
+
+class Collector:
+    r"""Collects the scalars broadcasted by `report()` and `report0()` and
+    computes their long-term averages (mean and standard deviation) over
+    user-defined periods of time.
+
+    The averages are first collected into internal counters that are not
+    directly visible to the user. They are then copied to the user-visible
+    state as a result of calling `update()` and can then be queried using
+    `mean()`, `std()`, `as_dict()`, etc. Calling `update()` also resets the
+    internal counters for the next round, so that the user-visible state
+    effectively reflects averages collected between the last two calls to
+    `update()`.
+
+    Args:
+        regex:          Regular expression defining which statistics to
+                        collect. The default is to collect everything.
+        keep_previous:  Whether to retain the previous averages if no
+                        scalars were collected on a given round
+                        (default: True).
+    """
+    def __init__(self, regex='.*', keep_previous=True):
+        self._regex = re.compile(regex)
+        self._keep_previous = keep_previous
+        self._cumulative = dict()
+        self._moments = dict()
+        self.update()
+        self._moments.clear()
+
+    def names(self):
+        r"""Returns the names of all statistics broadcasted so far that
+        match the regular expression specified at construction time.
+        """
+        return [name for name in _counters if self._regex.fullmatch(name)]
+
+    def update(self):
+        r"""Copies current values of the internal counters to the
+        user-visible state and resets them for the next round.
+
+        If `keep_previous=True` was specified at construction time, the
+        operation is skipped for statistics that have received no scalars
+        since the last update, retaining their previous averages.
+
+        This method performs a number of GPU-to-CPU transfers and one
+        `torch.distributed.all_reduce()`. It is intended to be called
+        periodically in the main training loop, typically once every
+        N training steps.
+        """
+        if not self._keep_previous:
+            self._moments.clear()
+        for name, cumulative in _sync(self.names()):
+            if name not in self._cumulative:
+                self._cumulative[name] = torch.zeros([_num_moments], dtype=_counter_dtype)
+            delta = cumulative - self._cumulative[name]
+            self._cumulative[name].copy_(cumulative)
+            if float(delta[0]) != 0:
+                self._moments[name] = delta
+
+    def _get_delta(self, name):
+        r"""Returns the raw moments that were accumulated for the given
+        statistic between the last two calls to `update()`, or zero if
+        no scalars were collected.
+        """
+        assert self._regex.fullmatch(name)
+        if name not in self._moments:
+            self._moments[name] = torch.zeros([_num_moments], dtype=_counter_dtype)
+        return self._moments[name]
+
+    def num(self, name):
+        r"""Returns the number of scalars that were accumulated for the given
+        statistic between the last two calls to `update()`, or zero if
+        no scalars were collected.
+        """
+        delta = self._get_delta(name)
+        return int(delta[0])
+
+    def mean(self, name):
+        r"""Returns the mean of the scalars that were accumulated for the
+        given statistic between the last two calls to `update()`, or NaN if
+        no scalars were collected.
+        """
+        delta = self._get_delta(name)
+        if int(delta[0]) == 0:
+            return float('nan')
+        return float(delta[1] / delta[0])
+
+    def std(self, name):
+        r"""Returns the standard deviation of the scalars that were
+        accumulated for the given statistic between the last two calls to
+        `update()`, or NaN if no scalars were collected.
+        """
+        delta = self._get_delta(name)
+        if int(delta[0]) == 0 or not np.isfinite(float(delta[1])):
+            return float('nan')
+        if int(delta[0]) == 1:
+            return float(0)
+        mean = float(delta[1] / delta[0])
+        raw_var = float(delta[2] / delta[0])
+        return np.sqrt(max(raw_var - np.square(mean), 0))
+
+    def as_dict(self):
+        r"""Returns the averages accumulated between the last two calls to
+        `update()` as an `dnnlib.EasyDict`. The contents are as follows:
+
+            dnnlib.EasyDict(
+                NAME = dnnlib.EasyDict(num=FLOAT, mean=FLOAT, std=FLOAT),
+                ...
+            )
+        """
+        stats = dnnlib.EasyDict()
+        for name in self.names():
+            stats[name] = dnnlib.EasyDict(num=self.num(name), mean=self.mean(name), std=self.std(name))
+        return stats
+
+    def __getitem__(self, name):
+        r"""Convenience getter.
+        `collector[name]` is a synonym for `collector.mean(name)`.
+        """
+        return self.mean(name)
+
+#----------------------------------------------------------------------------
+
+def _sync(names):
+    r"""Synchronize the global cumulative counters across devices and
+    processes. Called internally by `Collector.update()`.
+    """
+    if len(names) == 0:
+        return []
+    global _sync_called
+    _sync_called = True
+
+    # Collect deltas within current rank.
+    deltas = []
+    device = _sync_device if _sync_device is not None else torch.device('cpu')
+    for name in names:
+        delta = torch.zeros([_num_moments], dtype=_counter_dtype, device=device)
+        for counter in _counters[name].values():
+            delta.add_(counter.to(device))
+            counter.copy_(torch.zeros_like(counter))
+        deltas.append(delta)
+    deltas = torch.stack(deltas)
+
+    # Sum deltas across ranks.
+    if _sync_device is not None:
+        torch.distributed.all_reduce(deltas)
+
+    # Update cumulative values.
+    deltas = deltas.cpu()
+    for idx, name in enumerate(names):
+        if name not in _cumulative:
+            _cumulative[name] = torch.zeros([_num_moments], dtype=_counter_dtype)
+        _cumulative[name].add_(deltas[idx])
+
+    # Return name-value pairs.
+    return [(name, _cumulative[name]) for name in names]
+
+#----------------------------------------------------------------------------