Upload 14 files

face_parsing/__init__.py

@@ -0,0 +1,3 @@
+from .swap import init_parser, swap_regions, mask_regions, mask_regions_to_list
+from .model import BiSeNet
+from .parse_mask import init_parsing_model, get_parsed_mask, SoftErosion

face_parsing/model.py

@@ -0,0 +1,283 @@
+#!/usr/bin/python
+# -*- encoding: utf-8 -*-
+
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchvision
+
+from .resnet import Resnet18
+# from modules.bn import InPlaceABNSync as BatchNorm2d
+
+
+class ConvBNReLU(nn.Module):
+    def __init__(self, in_chan, out_chan, ks=3, stride=1, padding=1, *args, **kwargs):
+        super(ConvBNReLU, self).__init__()
+        self.conv = nn.Conv2d(in_chan,
+                out_chan,
+                kernel_size = ks,
+                stride = stride,
+                padding = padding,
+                bias = False)
+        self.bn = nn.BatchNorm2d(out_chan)
+        self.init_weight()
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = F.relu(self.bn(x))
+        return x
+
+    def init_weight(self):
+        for ly in self.children():
+            if isinstance(ly, nn.Conv2d):
+                nn.init.kaiming_normal_(ly.weight, a=1)
+                if not ly.bias is None: nn.init.constant_(ly.bias, 0)
+
+class BiSeNetOutput(nn.Module):
+    def __init__(self, in_chan, mid_chan, n_classes, *args, **kwargs):
+        super(BiSeNetOutput, self).__init__()
+        self.conv = ConvBNReLU(in_chan, mid_chan, ks=3, stride=1, padding=1)
+        self.conv_out = nn.Conv2d(mid_chan, n_classes, kernel_size=1, bias=False)
+        self.init_weight()
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.conv_out(x)
+        return x
+
+    def init_weight(self):
+        for ly in self.children():
+            if isinstance(ly, nn.Conv2d):
+                nn.init.kaiming_normal_(ly.weight, a=1)
+                if not ly.bias is None: nn.init.constant_(ly.bias, 0)
+
+    def get_params(self):
+        wd_params, nowd_params = [], []
+        for name, module in self.named_modules():
+            if isinstance(module, nn.Linear) or isinstance(module, nn.Conv2d):
+                wd_params.append(module.weight)
+                if not module.bias is None:
+                    nowd_params.append(module.bias)
+            elif isinstance(module, nn.BatchNorm2d):
+                nowd_params += list(module.parameters())
+        return wd_params, nowd_params
+
+
+class AttentionRefinementModule(nn.Module):
+    def __init__(self, in_chan, out_chan, *args, **kwargs):
+        super(AttentionRefinementModule, self).__init__()
+        self.conv = ConvBNReLU(in_chan, out_chan, ks=3, stride=1, padding=1)
+        self.conv_atten = nn.Conv2d(out_chan, out_chan, kernel_size= 1, bias=False)
+        self.bn_atten = nn.BatchNorm2d(out_chan)
+        self.sigmoid_atten = nn.Sigmoid()
+        self.init_weight()
+
+    def forward(self, x):
+        feat = self.conv(x)
+        atten = F.avg_pool2d(feat, feat.size()[2:])
+        atten = self.conv_atten(atten)
+        atten = self.bn_atten(atten)
+        atten = self.sigmoid_atten(atten)
+        out = torch.mul(feat, atten)
+        return out
+
+    def init_weight(self):
+        for ly in self.children():
+            if isinstance(ly, nn.Conv2d):
+                nn.init.kaiming_normal_(ly.weight, a=1)
+                if not ly.bias is None: nn.init.constant_(ly.bias, 0)
+
+
+class ContextPath(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super(ContextPath, self).__init__()
+        self.resnet = Resnet18()
+        self.arm16 = AttentionRefinementModule(256, 128)
+        self.arm32 = AttentionRefinementModule(512, 128)
+        self.conv_head32 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1)
+        self.conv_head16 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1)
+        self.conv_avg = ConvBNReLU(512, 128, ks=1, stride=1, padding=0)
+
+        self.init_weight()
+
+    def forward(self, x):
+        H0, W0 = x.size()[2:]
+        feat8, feat16, feat32 = self.resnet(x)
+        H8, W8 = feat8.size()[2:]
+        H16, W16 = feat16.size()[2:]
+        H32, W32 = feat32.size()[2:]
+
+        avg = F.avg_pool2d(feat32, feat32.size()[2:])
+        avg = self.conv_avg(avg)
+        avg_up = F.interpolate(avg, (H32, W32), mode='nearest')
+
+        feat32_arm = self.arm32(feat32)
+        feat32_sum = feat32_arm + avg_up
+        feat32_up = F.interpolate(feat32_sum, (H16, W16), mode='nearest')
+        feat32_up = self.conv_head32(feat32_up)
+
+        feat16_arm = self.arm16(feat16)
+        feat16_sum = feat16_arm + feat32_up
+        feat16_up = F.interpolate(feat16_sum, (H8, W8), mode='nearest')
+        feat16_up = self.conv_head16(feat16_up)
+
+        return feat8, feat16_up, feat32_up  # x8, x8, x16
+
+    def init_weight(self):
+        for ly in self.children():
+            if isinstance(ly, nn.Conv2d):
+                nn.init.kaiming_normal_(ly.weight, a=1)
+                if not ly.bias is None: nn.init.constant_(ly.bias, 0)
+
+    def get_params(self):
+        wd_params, nowd_params = [], []
+        for name, module in self.named_modules():
+            if isinstance(module, (nn.Linear, nn.Conv2d)):
+                wd_params.append(module.weight)
+                if not module.bias is None:
+                    nowd_params.append(module.bias)
+            elif isinstance(module, nn.BatchNorm2d):
+                nowd_params += list(module.parameters())
+        return wd_params, nowd_params
+
+
+### This is not used, since I replace this with the resnet feature with the same size
+class SpatialPath(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super(SpatialPath, self).__init__()
+        self.conv1 = ConvBNReLU(3, 64, ks=7, stride=2, padding=3)
+        self.conv2 = ConvBNReLU(64, 64, ks=3, stride=2, padding=1)
+        self.conv3 = ConvBNReLU(64, 64, ks=3, stride=2, padding=1)
+        self.conv_out = ConvBNReLU(64, 128, ks=1, stride=1, padding=0)
+        self.init_weight()
+
+    def forward(self, x):
+        feat = self.conv1(x)
+        feat = self.conv2(feat)
+        feat = self.conv3(feat)
+        feat = self.conv_out(feat)
+        return feat
+
+    def init_weight(self):
+        for ly in self.children():
+            if isinstance(ly, nn.Conv2d):
+                nn.init.kaiming_normal_(ly.weight, a=1)
+                if not ly.bias is None: nn.init.constant_(ly.bias, 0)
+
+    def get_params(self):
+        wd_params, nowd_params = [], []
+        for name, module in self.named_modules():
+            if isinstance(module, nn.Linear) or isinstance(module, nn.Conv2d):
+                wd_params.append(module.weight)
+                if not module.bias is None:
+                    nowd_params.append(module.bias)
+            elif isinstance(module, nn.BatchNorm2d):
+                nowd_params += list(module.parameters())
+        return wd_params, nowd_params
+
+
+class FeatureFusionModule(nn.Module):
+    def __init__(self, in_chan, out_chan, *args, **kwargs):
+        super(FeatureFusionModule, self).__init__()
+        self.convblk = ConvBNReLU(in_chan, out_chan, ks=1, stride=1, padding=0)
+        self.conv1 = nn.Conv2d(out_chan,
+                out_chan//4,
+                kernel_size = 1,
+                stride = 1,
+                padding = 0,
+                bias = False)
+        self.conv2 = nn.Conv2d(out_chan//4,
+                out_chan,
+                kernel_size = 1,
+                stride = 1,
+                padding = 0,
+                bias = False)
+        self.relu = nn.ReLU(inplace=True)
+        self.sigmoid = nn.Sigmoid()
+        self.init_weight()
+
+    def forward(self, fsp, fcp):
+        fcat = torch.cat([fsp, fcp], dim=1)
+        feat = self.convblk(fcat)
+        atten = F.avg_pool2d(feat, feat.size()[2:])
+        atten = self.conv1(atten)
+        atten = self.relu(atten)
+        atten = self.conv2(atten)
+        atten = self.sigmoid(atten)
+        feat_atten = torch.mul(feat, atten)
+        feat_out = feat_atten + feat
+        return feat_out
+
+    def init_weight(self):
+        for ly in self.children():
+            if isinstance(ly, nn.Conv2d):
+                nn.init.kaiming_normal_(ly.weight, a=1)
+                if not ly.bias is None: nn.init.constant_(ly.bias, 0)
+
+    def get_params(self):
+        wd_params, nowd_params = [], []
+        for name, module in self.named_modules():
+            if isinstance(module, nn.Linear) or isinstance(module, nn.Conv2d):
+                wd_params.append(module.weight)
+                if not module.bias is None:
+                    nowd_params.append(module.bias)
+            elif isinstance(module, nn.BatchNorm2d):
+                nowd_params += list(module.parameters())
+        return wd_params, nowd_params
+
+
+class BiSeNet(nn.Module):
+    def __init__(self, n_classes, *args, **kwargs):
+        super(BiSeNet, self).__init__()
+        self.cp = ContextPath()
+        ## here self.sp is deleted
+        self.ffm = FeatureFusionModule(256, 256)
+        self.conv_out = BiSeNetOutput(256, 256, n_classes)
+        self.conv_out16 = BiSeNetOutput(128, 64, n_classes)
+        self.conv_out32 = BiSeNetOutput(128, 64, n_classes)
+        self.init_weight()
+
+    def forward(self, x):
+        H, W = x.size()[2:]
+        feat_res8, feat_cp8, feat_cp16 = self.cp(x)  # here return res3b1 feature
+        feat_sp = feat_res8  # use res3b1 feature to replace spatial path feature
+        feat_fuse = self.ffm(feat_sp, feat_cp8)
+
+        feat_out = self.conv_out(feat_fuse)
+        feat_out16 = self.conv_out16(feat_cp8)
+        feat_out32 = self.conv_out32(feat_cp16)
+
+        feat_out = F.interpolate(feat_out, (H, W), mode='bilinear', align_corners=True)
+        feat_out16 = F.interpolate(feat_out16, (H, W), mode='bilinear', align_corners=True)
+        feat_out32 = F.interpolate(feat_out32, (H, W), mode='bilinear', align_corners=True)
+        return feat_out, feat_out16, feat_out32
+
+    def init_weight(self):
+        for ly in self.children():
+            if isinstance(ly, nn.Conv2d):
+                nn.init.kaiming_normal_(ly.weight, a=1)
+                if not ly.bias is None: nn.init.constant_(ly.bias, 0)
+
+    def get_params(self):
+        wd_params, nowd_params, lr_mul_wd_params, lr_mul_nowd_params = [], [], [], []
+        for name, child in self.named_children():
+            child_wd_params, child_nowd_params = child.get_params()
+            if isinstance(child, FeatureFusionModule) or isinstance(child, BiSeNetOutput):
+                lr_mul_wd_params += child_wd_params
+                lr_mul_nowd_params += child_nowd_params
+            else:
+                wd_params += child_wd_params
+                nowd_params += child_nowd_params
+        return wd_params, nowd_params, lr_mul_wd_params, lr_mul_nowd_params
+
+
+if __name__ == "__main__":
+    net = BiSeNet(19)
+    net.cuda()
+    net.eval()
+    in_ten = torch.randn(16, 3, 640, 480).cuda()
+    out, out16, out32 = net(in_ten)
+    print(out.shape)
+
+    net.get_params()

face_parsing/parse_mask.py

@@ -0,0 +1,107 @@
+import cv2
+import torch
+import torchvision
+import numpy as np
+import torch.nn as nn
+from PIL import Image
+from tqdm import tqdm
+import torch.nn.functional as F
+import torchvision.transforms as transforms
+
+from . model import BiSeNet
+
+class SoftErosion(nn.Module):
+    def __init__(self, kernel_size=15, threshold=0.6, iterations=1):
+        super(SoftErosion, self).__init__()
+        r = kernel_size // 2
+        self.padding = r
+        self.iterations = iterations
+        self.threshold = threshold
+
+        # Create kernel
+        y_indices, x_indices = torch.meshgrid(torch.arange(0., kernel_size), torch.arange(0., kernel_size))
+        dist = torch.sqrt((x_indices - r) ** 2 + (y_indices - r) ** 2)
+        kernel = dist.max() - dist
+        kernel /= kernel.sum()
+        kernel = kernel.view(1, 1, *kernel.shape)
+        self.register_buffer('weight', kernel)
+
+    def forward(self, x):
+        batch_size = x.size(0)  # Get the batch size
+        output = []
+
+        for i in tqdm(range(batch_size), desc="Soft-Erosion", leave=False):
+            input_tensor = x[i:i+1]  # Take one input tensor from the batch
+            input_tensor = input_tensor.float()  # Convert input to float tensor
+            input_tensor = input_tensor.unsqueeze(1)  # Add a channel dimension
+
+            for _ in range(self.iterations - 1):
+                input_tensor = torch.min(input_tensor, F.conv2d(input_tensor, weight=self.weight,
+                                                                groups=input_tensor.shape[1],
+                                                                padding=self.padding))
+            input_tensor = F.conv2d(input_tensor, weight=self.weight, groups=input_tensor.shape[1],
+                                    padding=self.padding)
+
+            mask = input_tensor >= self.threshold
+            input_tensor[mask] = 1.0
+            input_tensor[~mask] /= input_tensor[~mask].max()
+
+            input_tensor = input_tensor.squeeze(1)  # Remove the extra channel dimension
+            output.append(input_tensor.detach().cpu().numpy())
+
+        return np.array(output)
+
+transform = transforms.Compose([
+    transforms.Resize((512, 512)),
+    transforms.ToTensor(),
+    transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
+])
+
+
+
+def init_parsing_model(model_path, device="cpu"):
+    net = BiSeNet(19)
+    net.to(device)
+    net.load_state_dict(torch.load(model_path))
+    net.eval()
+    return net
+
+def transform_images(imgs):
+    tensor_images = torch.stack([transform(Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))) for img in imgs], dim=0)
+    return tensor_images
+
+def get_parsed_mask(net, imgs, classes=[1, 2, 3, 4, 5, 10, 11, 12, 13], device="cpu", batch_size=8, softness=20):
+    if softness > 0:
+        smooth_mask = SoftErosion(kernel_size=17, threshold=0.9, iterations=softness).to(device)
+
+    masks = []
+    for i in tqdm(range(0, len(imgs), batch_size), total=len(imgs) // batch_size, desc="Face-parsing"):
+        batch_imgs = imgs[i:i + batch_size]
+
+        tensor_images = transform_images(batch_imgs).to(device)
+        with torch.no_grad():
+            out = net(tensor_images)[0]
+        # parsing = out.argmax(dim=1)
+        # arget_classes = torch.tensor(classes).to(device)
+        # batch_masks = torch.isin(parsing, target_classes).to(device)
+        ## torch.isin was slightly slower in my test, so using np.isin
+        parsing = out.argmax(dim=1).detach().cpu().numpy()
+        batch_masks = np.isin(parsing, classes).astype('float32')
+
+        if softness > 0:
+            # batch_masks = smooth_mask(batch_masks).transpose(1,0,2,3)[0]
+            mask_tensor = torch.from_numpy(batch_masks.copy()).float().to(device)
+            batch_masks = smooth_mask(mask_tensor).transpose(1,0,2,3)[0]
+
+        yield batch_masks
+
+        #masks.append(batch_masks)
+
+    #if len(masks) >= 1:
+    #    masks = np.concatenate(masks, axis=0)
+    # masks = np.repeat(np.expand_dims(masks, axis=1), 3, axis=1)
+
+    # for i, mask in enumerate(masks):
+    #    cv2.imwrite(f"mask/{i}.jpg", (mask * 255).astype("uint8"))
+
+    #return masks

face_parsing/resnet.py

@@ -0,0 +1,109 @@
+#!/usr/bin/python
+# -*- encoding: utf-8 -*-
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.model_zoo as modelzoo
+
+# from modules.bn import InPlaceABNSync as BatchNorm2d
+
+resnet18_url = 'https://download.pytorch.org/models/resnet18-5c106cde.pth'
+
+
+def conv3x3(in_planes, out_planes, stride=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
+                     padding=1, bias=False)
+
+
+class BasicBlock(nn.Module):
+    def __init__(self, in_chan, out_chan, stride=1):
+        super(BasicBlock, self).__init__()
+        self.conv1 = conv3x3(in_chan, out_chan, stride)
+        self.bn1 = nn.BatchNorm2d(out_chan)
+        self.conv2 = conv3x3(out_chan, out_chan)
+        self.bn2 = nn.BatchNorm2d(out_chan)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = None
+        if in_chan != out_chan or stride != 1:
+            self.downsample = nn.Sequential(
+                nn.Conv2d(in_chan, out_chan,
+                          kernel_size=1, stride=stride, bias=False),
+                nn.BatchNorm2d(out_chan),
+                )
+
+    def forward(self, x):
+        residual = self.conv1(x)
+        residual = F.relu(self.bn1(residual))
+        residual = self.conv2(residual)
+        residual = self.bn2(residual)
+
+        shortcut = x
+        if self.downsample is not None:
+            shortcut = self.downsample(x)
+
+        out = shortcut + residual
+        out = self.relu(out)
+        return out
+
+
+def create_layer_basic(in_chan, out_chan, bnum, stride=1):
+    layers = [BasicBlock(in_chan, out_chan, stride=stride)]
+    for i in range(bnum-1):
+        layers.append(BasicBlock(out_chan, out_chan, stride=1))
+    return nn.Sequential(*layers)
+
+
+class Resnet18(nn.Module):
+    def __init__(self):
+        super(Resnet18, self).__init__()
+        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
+                               bias=False)
+        self.bn1 = nn.BatchNorm2d(64)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = create_layer_basic(64, 64, bnum=2, stride=1)
+        self.layer2 = create_layer_basic(64, 128, bnum=2, stride=2)
+        self.layer3 = create_layer_basic(128, 256, bnum=2, stride=2)
+        self.layer4 = create_layer_basic(256, 512, bnum=2, stride=2)
+        self.init_weight()
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = F.relu(self.bn1(x))
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        feat8 = self.layer2(x) # 1/8
+        feat16 = self.layer3(feat8) # 1/16
+        feat32 = self.layer4(feat16) # 1/32
+        return feat8, feat16, feat32
+
+    def init_weight(self):
+        state_dict = modelzoo.load_url(resnet18_url)
+        self_state_dict = self.state_dict()
+        for k, v in state_dict.items():
+            if 'fc' in k: continue
+            self_state_dict.update({k: v})
+        self.load_state_dict(self_state_dict)
+
+    def get_params(self):
+        wd_params, nowd_params = [], []
+        for name, module in self.named_modules():
+            if isinstance(module, (nn.Linear, nn.Conv2d)):
+                wd_params.append(module.weight)
+                if not module.bias is None:
+                    nowd_params.append(module.bias)
+            elif isinstance(module,  nn.BatchNorm2d):
+                nowd_params += list(module.parameters())
+        return wd_params, nowd_params
+
+
+if __name__ == "__main__":
+    net = Resnet18()
+    x = torch.randn(16, 3, 224, 224)
+    out = net(x)
+    print(out[0].size())
+    print(out[1].size())
+    print(out[2].size())
+    net.get_params()

face_parsing/swap.py

@@ -0,0 +1,133 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchvision.transforms as transforms
+import cv2
+import numpy as np
+
+from .model import BiSeNet
+
+mask_regions = {
+    "Background":0,
+    "Skin":1,
+    "L-Eyebrow":2,
+    "R-Eyebrow":3,
+    "L-Eye":4,
+    "R-Eye":5,
+    "Eye-G":6,
+    "L-Ear":7,
+    "R-Ear":8,
+    "Ear-R":9,
+    "Nose":10,
+    "Mouth":11,
+    "U-Lip":12,
+    "L-Lip":13,
+    "Neck":14,
+    "Neck-L":15,
+    "Cloth":16,
+    "Hair":17,
+    "Hat":18
+}
+
+# Borrowed from simswap
+# https://github.com/neuralchen/SimSwap/blob/26c84d2901bd56eda4d5e3c5ca6da16e65dc82a6/util/reverse2original.py#L30
+class SoftErosion(nn.Module):
+    def __init__(self, kernel_size=15, threshold=0.6, iterations=1):
+        super(SoftErosion, self).__init__()
+        r = kernel_size // 2
+        self.padding = r
+        self.iterations = iterations
+        self.threshold = threshold
+
+        # Create kernel
+        y_indices, x_indices = torch.meshgrid(torch.arange(0., kernel_size), torch.arange(0., kernel_size))
+        dist = torch.sqrt((x_indices - r) ** 2 + (y_indices - r) ** 2)
+        kernel = dist.max() - dist
+        kernel /= kernel.sum()
+        kernel = kernel.view(1, 1, *kernel.shape)
+        self.register_buffer('weight', kernel)
+
+    def forward(self, x):
+        x = x.float()
+        for i in range(self.iterations - 1):
+            x = torch.min(x, F.conv2d(x, weight=self.weight, groups=x.shape[1], padding=self.padding))
+        x = F.conv2d(x, weight=self.weight, groups=x.shape[1], padding=self.padding)
+
+        mask = x >= self.threshold
+        x[mask] = 1.0
+        x[~mask] /= x[~mask].max()
+
+        return x, mask
+
+device = "cpu"
+
+def init_parser(pth_path, mode="cpu"):
+    global device
+    device = mode
+    n_classes = 19
+    net = BiSeNet(n_classes=n_classes)
+    if device == "cuda":
+        net.cuda()
+        net.load_state_dict(torch.load(pth_path))
+    else:
+        net.load_state_dict(torch.load(pth_path, map_location=torch.device('cpu')))
+    net.eval()
+    return net
+
+
+def image_to_parsing(img, net):
+    img = cv2.resize(img, (512, 512))
+    img = img[:,:,::-1]
+    transform = transforms.Compose([
+        transforms.ToTensor(),
+        transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
+    ])
+    img = transform(img.copy())
+    img = torch.unsqueeze(img, 0)
+
+    with torch.no_grad():
+        img = img.to(device)
+        out = net(img)[0]
+        parsing = out.squeeze(0).cpu().numpy().argmax(0)
+        return parsing
+
+
+def get_mask(parsing, classes):
+    res = parsing == classes[0]
+    for val in classes[1:]:
+        res += parsing == val
+    return res
+
+
+def swap_regions(source, target, net, smooth_mask, includes=[1,2,3,4,5,10,11,12,13], blur=10):
+    parsing = image_to_parsing(source, net)
+
+    if len(includes) == 0:
+        return source, np.zeros_like(source)
+
+    include_mask = get_mask(parsing, includes)
+    mask = np.repeat(include_mask[:, :, np.newaxis], 3, axis=2).astype("float32")
+
+    if smooth_mask is not None:
+        mask_tensor = torch.from_numpy(mask.copy().transpose((2, 0, 1))).float().to(device)
+        face_mask_tensor = mask_tensor[0] + mask_tensor[1]
+        soft_face_mask_tensor, _ = smooth_mask(face_mask_tensor.unsqueeze_(0).unsqueeze_(0))
+        soft_face_mask_tensor.squeeze_()
+        mask = np.repeat(soft_face_mask_tensor.cpu().numpy()[:, :, np.newaxis], 3, axis=2)
+
+    if blur > 0:
+        mask = cv2.GaussianBlur(mask, (0, 0), blur)
+
+    resized_source = cv2.resize((source).astype("float32"), (512, 512))
+    resized_target = cv2.resize((target).astype("float32"), (512, 512))
+    result = mask * resized_source + (1 - mask) * resized_target
+    result = cv2.resize(result.astype("uint8"), (source.shape[1], source.shape[0]))
+
+    return result
+
+def mask_regions_to_list(values):
+    out_ids = []
+    for value in values:
+        if value in mask_regions.keys():
+            out_ids.append(mask_regions.get(value))
+    return out_ids

gfpgan/weights/detection_Resnet50_Final.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9253ceab3578be0efd87eed777820c4a0d7e31a5e2068c3156722f2d6653b73
+size 134

gfpgan/weights/parsing_parsenet.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1df950acea7b00dc68fab8cb603099096d8d2130f6c93650ac5c6ad27f3f009
+size 133

upscaler/RealESRGAN/__init__.py

@@ -0,0 +1 @@
+from .model import RealESRGAN

upscaler/RealESRGAN/arch_utils.py

@@ -0,0 +1,197 @@
+import math
+import torch
+from torch import nn as nn
+from torch.nn import functional as F
+from torch.nn import init as init
+from torch.nn.modules.batchnorm import _BatchNorm
+
+@torch.no_grad()
+def default_init_weights(module_list, scale=1, bias_fill=0, **kwargs):
+    """Initialize network weights.
+
+    Args:
+        module_list (list[nn.Module] | nn.Module): Modules to be initialized.
+        scale (float): Scale initialized weights, especially for residual
+            blocks. Default: 1.
+        bias_fill (float): The value to fill bias. Default: 0
+        kwargs (dict): Other arguments for initialization function.
+    """
+    if not isinstance(module_list, list):
+        module_list = [module_list]
+    for module in module_list:
+        for m in module.modules():
+            if isinstance(m, nn.Conv2d):
+                init.kaiming_normal_(m.weight, **kwargs)
+                m.weight.data *= scale
+                if m.bias is not None:
+                    m.bias.data.fill_(bias_fill)
+            elif isinstance(m, nn.Linear):
+                init.kaiming_normal_(m.weight, **kwargs)
+                m.weight.data *= scale
+                if m.bias is not None:
+                    m.bias.data.fill_(bias_fill)
+            elif isinstance(m, _BatchNorm):
+                init.constant_(m.weight, 1)
+                if m.bias is not None:
+                    m.bias.data.fill_(bias_fill)
+
+
+def make_layer(basic_block, num_basic_block, **kwarg):
+    """Make layers by stacking the same blocks.
+
+    Args:
+        basic_block (nn.module): nn.module class for basic block.
+        num_basic_block (int): number of blocks.
+
+    Returns:
+        nn.Sequential: Stacked blocks in nn.Sequential.
+    """
+    layers = []
+    for _ in range(num_basic_block):
+        layers.append(basic_block(**kwarg))
+    return nn.Sequential(*layers)
+
+
+class ResidualBlockNoBN(nn.Module):
+    """Residual block without BN.
+
+    It has a style of:
+        ---Conv-ReLU-Conv-+-
+         |________________|
+
+    Args:
+        num_feat (int): Channel number of intermediate features.
+            Default: 64.
+        res_scale (float): Residual scale. Default: 1.
+        pytorch_init (bool): If set to True, use pytorch default init,
+            otherwise, use default_init_weights. Default: False.
+    """
+
+    def __init__(self, num_feat=64, res_scale=1, pytorch_init=False):
+        super(ResidualBlockNoBN, self).__init__()
+        self.res_scale = res_scale
+        self.conv1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1, bias=True)
+        self.conv2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1, bias=True)
+        self.relu = nn.ReLU(inplace=True)
+
+        if not pytorch_init:
+            default_init_weights([self.conv1, self.conv2], 0.1)
+
+    def forward(self, x):
+        identity = x
+        out = self.conv2(self.relu(self.conv1(x)))
+        return identity + out * self.res_scale
+
+
+class Upsample(nn.Sequential):
+    """Upsample module.
+
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+    """
+
+    def __init__(self, scale, num_feat):
+        m = []
+        if (scale & (scale - 1)) == 0:  # scale = 2^n
+            for _ in range(int(math.log(scale, 2))):
+                m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
+                m.append(nn.PixelShuffle(2))
+        elif scale == 3:
+            m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
+            m.append(nn.PixelShuffle(3))
+        else:
+            raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.')
+        super(Upsample, self).__init__(*m)
+
+
+def flow_warp(x, flow, interp_mode='bilinear', padding_mode='zeros', align_corners=True):
+    """Warp an image or feature map with optical flow.
+
+    Args:
+        x (Tensor): Tensor with size (n, c, h, w).
+        flow (Tensor): Tensor with size (n, h, w, 2), normal value.
+        interp_mode (str): 'nearest' or 'bilinear'. Default: 'bilinear'.
+        padding_mode (str): 'zeros' or 'border' or 'reflection'.
+            Default: 'zeros'.
+        align_corners (bool): Before pytorch 1.3, the default value is
+            align_corners=True. After pytorch 1.3, the default value is
+            align_corners=False. Here, we use the True as default.
+
+    Returns:
+        Tensor: Warped image or feature map.
+    """
+    assert x.size()[-2:] == flow.size()[1:3]
+    _, _, h, w = x.size()
+    # create mesh grid
+    grid_y, grid_x = torch.meshgrid(torch.arange(0, h).type_as(x), torch.arange(0, w).type_as(x))
+    grid = torch.stack((grid_x, grid_y), 2).float()  # W(x), H(y), 2
+    grid.requires_grad = False
+
+    vgrid = grid + flow
+    # scale grid to [-1,1]
+    vgrid_x = 2.0 * vgrid[:, :, :, 0] / max(w - 1, 1) - 1.0
+    vgrid_y = 2.0 * vgrid[:, :, :, 1] / max(h - 1, 1) - 1.0
+    vgrid_scaled = torch.stack((vgrid_x, vgrid_y), dim=3)
+    output = F.grid_sample(x, vgrid_scaled, mode=interp_mode, padding_mode=padding_mode, align_corners=align_corners)
+
+    # TODO, what if align_corners=False
+    return output
+
+
+def resize_flow(flow, size_type, sizes, interp_mode='bilinear', align_corners=False):
+    """Resize a flow according to ratio or shape.
+
+    Args:
+        flow (Tensor): Precomputed flow. shape [N, 2, H, W].
+        size_type (str): 'ratio' or 'shape'.
+        sizes (list[int | float]): the ratio for resizing or the final output
+            shape.
+            1) The order of ratio should be [ratio_h, ratio_w]. For
+            downsampling, the ratio should be smaller than 1.0 (i.e., ratio
+            < 1.0). For upsampling, the ratio should be larger than 1.0 (i.e.,
+            ratio > 1.0).
+            2) The order of output_size should be [out_h, out_w].
+        interp_mode (str): The mode of interpolation for resizing.
+            Default: 'bilinear'.
+        align_corners (bool): Whether align corners. Default: False.
+
+    Returns:
+        Tensor: Resized flow.
+    """
+    _, _, flow_h, flow_w = flow.size()
+    if size_type == 'ratio':
+        output_h, output_w = int(flow_h * sizes[0]), int(flow_w * sizes[1])
+    elif size_type == 'shape':
+        output_h, output_w = sizes[0], sizes[1]
+    else:
+        raise ValueError(f'Size type should be ratio or shape, but got type {size_type}.')
+
+    input_flow = flow.clone()
+    ratio_h = output_h / flow_h
+    ratio_w = output_w / flow_w
+    input_flow[:, 0, :, :] *= ratio_w
+    input_flow[:, 1, :, :] *= ratio_h
+    resized_flow = F.interpolate(
+        input=input_flow, size=(output_h, output_w), mode=interp_mode, align_corners=align_corners)
+    return resized_flow
+
+
+# TODO: may write a cpp file
+def pixel_unshuffle(x, scale):
+    """ Pixel unshuffle.
+
+    Args:
+        x (Tensor): Input feature with shape (b, c, hh, hw).
+        scale (int): Downsample ratio.
+
+    Returns:
+        Tensor: the pixel unshuffled feature.
+    """
+    b, c, hh, hw = x.size()
+    out_channel = c * (scale**2)
+    assert hh % scale == 0 and hw % scale == 0
+    h = hh // scale
+    w = hw // scale
+    x_view = x.view(b, c, h, scale, w, scale)
+    return x_view.permute(0, 1, 3, 5, 2, 4).reshape(b, out_channel, h, w)

upscaler/RealESRGAN/model.py

@@ -0,0 +1,90 @@
+import os
+import torch
+from torch.nn import functional as F
+from PIL import Image
+import numpy as np
+import cv2
+
+from .rrdbnet_arch import RRDBNet
+from .utils import pad_reflect, split_image_into_overlapping_patches, stich_together, \
+                   unpad_image
+
+
+HF_MODELS = {
+    2: dict(
+        repo_id='sberbank-ai/Real-ESRGAN',
+        filename='RealESRGAN_x2.pth',
+    ),
+    4: dict(
+        repo_id='sberbank-ai/Real-ESRGAN',
+        filename='RealESRGAN_x4.pth',
+    ),
+    8: dict(
+        repo_id='sberbank-ai/Real-ESRGAN',
+        filename='RealESRGAN_x8.pth',
+    ),
+}
+
+
+class RealESRGAN:
+    def __init__(self, device, scale=4):
+        self.device = device
+        self.scale = scale
+        self.model = RRDBNet(
+            num_in_ch=3, num_out_ch=3, num_feat=64,
+            num_block=23, num_grow_ch=32, scale=scale
+        )
+
+    def load_weights(self, model_path, download=True):
+        if not os.path.exists(model_path) and download:
+            from huggingface_hub import hf_hub_url, cached_download
+            assert self.scale in [2,4,8], 'You can download models only with scales: 2, 4, 8'
+            config = HF_MODELS[self.scale]
+            cache_dir = os.path.dirname(model_path)
+            local_filename = os.path.basename(model_path)
+            config_file_url = hf_hub_url(repo_id=config['repo_id'], filename=config['filename'])
+            cached_download(config_file_url, cache_dir=cache_dir, force_filename=local_filename)
+            print('Weights downloaded to:', os.path.join(cache_dir, local_filename))
+
+        loadnet = torch.load(model_path)
+        if 'params' in loadnet:
+            self.model.load_state_dict(loadnet['params'], strict=True)
+        elif 'params_ema' in loadnet:
+            self.model.load_state_dict(loadnet['params_ema'], strict=True)
+        else:
+            self.model.load_state_dict(loadnet, strict=True)
+        self.model.eval()
+        self.model.to(self.device)
+
+    @torch.cuda.amp.autocast()
+    def predict(self, lr_image, batch_size=4, patches_size=192,
+                padding=24, pad_size=15):
+        scale = self.scale
+        device = self.device
+        lr_image = np.array(lr_image)
+        lr_image = pad_reflect(lr_image, pad_size)
+
+        patches, p_shape = split_image_into_overlapping_patches(
+            lr_image, patch_size=patches_size, padding_size=padding
+        )
+        img = torch.FloatTensor(patches/255).permute((0,3,1,2)).to(device).detach()
+
+        with torch.no_grad():
+            res = self.model(img[0:batch_size])
+            for i in range(batch_size, img.shape[0], batch_size):
+                res = torch.cat((res, self.model(img[i:i+batch_size])), 0)
+
+        sr_image = res.permute((0,2,3,1)).clamp_(0, 1).cpu()
+        np_sr_image = sr_image.numpy()
+
+        padded_size_scaled = tuple(np.multiply(p_shape[0:2], scale)) + (3,)
+        scaled_image_shape = tuple(np.multiply(lr_image.shape[0:2], scale)) + (3,)
+        np_sr_image = stich_together(
+            np_sr_image, padded_image_shape=padded_size_scaled,
+            target_shape=scaled_image_shape, padding_size=padding * scale
+        )
+        sr_img = (np_sr_image*255).astype(np.uint8)
+        sr_img = unpad_image(sr_img, pad_size*scale)
+        #sr_img = Image.fromarray(sr_img)
+
+        return sr_img

upscaler/RealESRGAN/rrdbnet_arch.py

@@ -0,0 +1,121 @@
+import torch
+from torch import nn as nn
+from torch.nn import functional as F
+
+from .arch_utils import default_init_weights, make_layer, pixel_unshuffle
+
+
+class ResidualDenseBlock(nn.Module):
+    """Residual Dense Block.
+
+    Used in RRDB block in ESRGAN.
+
+    Args:
+        num_feat (int): Channel number of intermediate features.
+        num_grow_ch (int): Channels for each growth.
+    """
+
+    def __init__(self, num_feat=64, num_grow_ch=32):
+        super(ResidualDenseBlock, self).__init__()
+        self.conv1 = nn.Conv2d(num_feat, num_grow_ch, 3, 1, 1)
+        self.conv2 = nn.Conv2d(num_feat + num_grow_ch, num_grow_ch, 3, 1, 1)
+        self.conv3 = nn.Conv2d(num_feat + 2 * num_grow_ch, num_grow_ch, 3, 1, 1)
+        self.conv4 = nn.Conv2d(num_feat + 3 * num_grow_ch, num_grow_ch, 3, 1, 1)
+        self.conv5 = nn.Conv2d(num_feat + 4 * num_grow_ch, num_feat, 3, 1, 1)
+
+        self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+
+        # initialization
+        default_init_weights([self.conv1, self.conv2, self.conv3, self.conv4, self.conv5], 0.1)
+
+    def forward(self, x):
+        x1 = self.lrelu(self.conv1(x))
+        x2 = self.lrelu(self.conv2(torch.cat((x, x1), 1)))
+        x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), 1)))
+        x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1)))
+        x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1))
+        # Emperically, we use 0.2 to scale the residual for better performance
+        return x5 * 0.2 + x
+
+
+class RRDB(nn.Module):
+    """Residual in Residual Dense Block.
+
+    Used in RRDB-Net in ESRGAN.
+
+    Args:
+        num_feat (int): Channel number of intermediate features.
+        num_grow_ch (int): Channels for each growth.
+    """
+
+    def __init__(self, num_feat, num_grow_ch=32):
+        super(RRDB, self).__init__()
+        self.rdb1 = ResidualDenseBlock(num_feat, num_grow_ch)
+        self.rdb2 = ResidualDenseBlock(num_feat, num_grow_ch)
+        self.rdb3 = ResidualDenseBlock(num_feat, num_grow_ch)
+
+    def forward(self, x):
+        out = self.rdb1(x)
+        out = self.rdb2(out)
+        out = self.rdb3(out)
+        # Emperically, we use 0.2 to scale the residual for better performance
+        return out * 0.2 + x
+
+
+class RRDBNet(nn.Module):
+    """Networks consisting of Residual in Residual Dense Block, which is used
+    in ESRGAN.
+
+    ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks.
+
+    We extend ESRGAN for scale x2 and scale x1.
+    Note: This is one option for scale 1, scale 2 in RRDBNet.
+    We first employ the pixel-unshuffle (an inverse operation of pixelshuffle to reduce the spatial size
+    and enlarge the channel size before feeding inputs into the main ESRGAN architecture.
+
+    Args:
+        num_in_ch (int): Channel number of inputs.
+        num_out_ch (int): Channel number of outputs.
+        num_feat (int): Channel number of intermediate features.
+            Default: 64
+        num_block (int): Block number in the trunk network. Defaults: 23
+        num_grow_ch (int): Channels for each growth. Default: 32.
+    """
+
+    def __init__(self, num_in_ch, num_out_ch, scale=4, num_feat=64, num_block=23, num_grow_ch=32):
+        super(RRDBNet, self).__init__()
+        self.scale = scale
+        if scale == 2:
+            num_in_ch = num_in_ch * 4
+        elif scale == 1:
+            num_in_ch = num_in_ch * 16
+        self.conv_first = nn.Conv2d(num_in_ch, num_feat, 3, 1, 1)
+        self.body = make_layer(RRDB, num_block, num_feat=num_feat, num_grow_ch=num_grow_ch)
+        self.conv_body = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        # upsample
+        self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        if scale == 8:
+            self.conv_up3 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+
+        self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+
+    def forward(self, x):
+        if self.scale == 2:
+            feat = pixel_unshuffle(x, scale=2)
+        elif self.scale == 1:
+            feat = pixel_unshuffle(x, scale=4)
+        else:
+            feat = x
+        feat = self.conv_first(feat)
+        body_feat = self.conv_body(self.body(feat))
+        feat = feat + body_feat
+        # upsample
+        feat = self.lrelu(self.conv_up1(F.interpolate(feat, scale_factor=2, mode='nearest')))
+        feat = self.lrelu(self.conv_up2(F.interpolate(feat, scale_factor=2, mode='nearest')))
+        if self.scale == 8:
+            feat = self.lrelu(self.conv_up3(F.interpolate(feat, scale_factor=2, mode='nearest')))
+        out = self.conv_last(self.lrelu(self.conv_hr(feat)))
+        return out

upscaler/RealESRGAN/utils.py

@@ -0,0 +1,133 @@
+import numpy as np
+import torch
+from PIL import Image
+import os
+import io
+
+def pad_reflect(image, pad_size):
+    imsize = image.shape
+    height, width = imsize[:2]
+    new_img = np.zeros([height+pad_size*2, width+pad_size*2, imsize[2]]).astype(np.uint8)
+    new_img[pad_size:-pad_size, pad_size:-pad_size, :] = image
+    
+    new_img[0:pad_size, pad_size:-pad_size, :] = np.flip(image[0:pad_size, :, :], axis=0) #top
+    new_img[-pad_size:, pad_size:-pad_size, :] = np.flip(image[-pad_size:, :, :], axis=0) #bottom
+    new_img[:, 0:pad_size, :] = np.flip(new_img[:, pad_size:pad_size*2, :], axis=1) #left
+    new_img[:, -pad_size:, :] = np.flip(new_img[:, -pad_size*2:-pad_size, :], axis=1) #right
+    
+    return new_img
+
+def unpad_image(image, pad_size):
+    return image[pad_size:-pad_size, pad_size:-pad_size, :]
+
+
+def process_array(image_array, expand=True):
+    """ Process a 3-dimensional array into a scaled, 4 dimensional batch of size 1. """
+    
+    image_batch = image_array / 255.0
+    if expand:
+        image_batch = np.expand_dims(image_batch, axis=0)
+    return image_batch
+
+
+def process_output(output_tensor):
+    """ Transforms the 4-dimensional output tensor into a suitable image format. """
+    
+    sr_img = output_tensor.clip(0, 1) * 255
+    sr_img = np.uint8(sr_img)
+    return sr_img
+
+
+def pad_patch(image_patch, padding_size, channel_last=True):
+    """ Pads image_patch with with padding_size edge values. """
+    
+    if channel_last:
+        return np.pad(
+            image_patch,
+            ((padding_size, padding_size), (padding_size, padding_size), (0, 0)),
+            'edge',
+        )
+    else:
+        return np.pad(
+            image_patch,
+            ((0, 0), (padding_size, padding_size), (padding_size, padding_size)),
+            'edge',
+        )
+
+
+def unpad_patches(image_patches, padding_size):
+    return image_patches[:, padding_size:-padding_size, padding_size:-padding_size, :]
+
+
+def split_image_into_overlapping_patches(image_array, patch_size, padding_size=2):
+    """ Splits the image into partially overlapping patches.
+    The patches overlap by padding_size pixels.
+    Pads the image twice:
+        - first to have a size multiple of the patch size,
+        - then to have equal padding at the borders.
+    Args:
+        image_array: numpy array of the input image.
+        patch_size: size of the patches from the original image (without padding).
+        padding_size: size of the overlapping area.
+    """
+    
+    xmax, ymax, _ = image_array.shape
+    x_remainder = xmax % patch_size
+    y_remainder = ymax % patch_size
+    
+    # modulo here is to avoid extending of patch_size instead of 0
+    x_extend = (patch_size - x_remainder) % patch_size
+    y_extend = (patch_size - y_remainder) % patch_size
+    
+    # make sure the image is divisible into regular patches
+    extended_image = np.pad(image_array, ((0, x_extend), (0, y_extend), (0, 0)), 'edge')
+    
+    # add padding around the image to simplify computations
+    padded_image = pad_patch(extended_image, padding_size, channel_last=True)
+    
+    xmax, ymax, _ = padded_image.shape
+    patches = []
+    
+    x_lefts = range(padding_size, xmax - padding_size, patch_size)
+    y_tops = range(padding_size, ymax - padding_size, patch_size)
+    
+    for x in x_lefts:
+        for y in y_tops:
+            x_left = x - padding_size
+            y_top = y - padding_size
+            x_right = x + patch_size + padding_size
+            y_bottom = y + patch_size + padding_size
+            patch = padded_image[x_left:x_right, y_top:y_bottom, :]
+            patches.append(patch)
+    
+    return np.array(patches), padded_image.shape
+
+
+def stich_together(patches, padded_image_shape, target_shape, padding_size=4):
+    """ Reconstruct the image from overlapping patches.
+    After scaling, shapes and padding should be scaled too.
+    Args:
+        patches: patches obtained with split_image_into_overlapping_patches
+        padded_image_shape: shape of the padded image contructed in split_image_into_overlapping_patches
+        target_shape: shape of the final image
+        padding_size: size of the overlapping area.
+    """
+    
+    xmax, ymax, _ = padded_image_shape
+    patches = unpad_patches(patches, padding_size)
+    patch_size = patches.shape[1]
+    n_patches_per_row = ymax // patch_size
+    
+    complete_image = np.zeros((xmax, ymax, 3))
+    
+    row = -1
+    col = 0
+    for i in range(len(patches)):
+        if i % n_patches_per_row == 0:
+            row += 1
+            col = 0
+        complete_image[
+        row * patch_size: (row + 1) * patch_size, col * patch_size: (col + 1) * patch_size,:
+        ] = patches[i]
+        col += 1
+    return complete_image[0: target_shape[0], 0: target_shape[1], :]

upscaler/codeformer.py

@@ -0,0 +1,37 @@
+import cv2
+import torch
+import onnx
+import onnxruntime
+import numpy as np
+
+import time
+
+# codeformer converted to onnx
+# using https://github.com/redthing1/CodeFormer
+
+
+class CodeFormerEnhancer:
+    def __init__(self, model_path="codeformer.onnx", device='cpu'):
+        model = onnx.load(model_path)
+        session_options = onnxruntime.SessionOptions()
+        session_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
+        providers = ["CPUExecutionProvider"]
+        if device == 'cuda':
+            providers = [("CUDAExecutionProvider", {"cudnn_conv_algo_search": "DEFAULT"}),"CPUExecutionProvider"]
+        self.session = onnxruntime.InferenceSession(model_path, sess_options=session_options, providers=providers)
+
+    def enhance(self, img, w=0.9):
+        img = cv2.resize(img, (512, 512), interpolation=cv2.INTER_LINEAR)
+        img = img.astype(np.float32)[:,:,::-1] / 255.0
+        img = img.transpose((2, 0, 1))
+        nrm_mean = np.array([0.5, 0.5, 0.5]).reshape((-1, 1, 1))
+        nrm_std = np.array([0.5, 0.5, 0.5]).reshape((-1, 1, 1))
+        img = (img - nrm_mean) / nrm_std
+
+        img = np.expand_dims(img, axis=0)
+
+        out = self.session.run(None, {'x':img.astype(np.float32), 'w':np.array([w], dtype=np.double)})[0]
+        out = (out[0].transpose(1,2,0).clip(-1,1) + 1) * 0.5
+        out = (out * 255)[:,:,::-1]
+
+        return out.astype('uint8')