first

README.md

@@ -1,12 +1,255 @@
----
-title: Text2Human API
-emoji: 📈
-colorFrom: gray
-colorTo: gray
-sdk: gradio
-sdk_version: 3.0.22
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Text2Human - Official PyTorch Implementation
+
+<!-- <img src="./doc_images/overview.jpg" width="96%" height="96%"> -->
+
+This repository provides the official PyTorch implementation for the following paper:
+
+**Text2Human: Text-Driven Controllable Human Image Generation**</br>
+[Yuming Jiang](https://yumingj.github.io/), [Shuai Yang](https://williamyang1991.github.io/), [Haonan Qiu](http://haonanqiu.com/), [Wayne Wu](https://dblp.org/pid/50/8731.html), [Chen Change Loy](https://www.mmlab-ntu.com/person/ccloy/) and [Ziwei Liu](https://liuziwei7.github.io/)</br>
+In ACM Transactions on Graphics (Proceedings of SIGGRAPH), 2022.
+
+From [MMLab@NTU](https://www.mmlab-ntu.com/index.html) affliated with S-Lab, Nanyang Technological University and SenseTime Research.
+
+<table>
+<tr>
+    <td><img src="assets/1.png" width="100%"/></td>
+    <td><img src="assets/2.png" width="100%"/></td>
+    <td><img src="assets/3.png" width="100%"/></td>
+    <td><img src="assets/4.png" width="100%"/></td>
+</tr>
+<tr>
+    <td align='center' width='24%'>The lady wears a short-sleeve T-shirt with pure color pattern, and a short and denim skirt.</td>
+    <td align='center' width='24%'>The man wears a long and floral shirt, and long pants with the pure color pattern.</td>
+    <td align='center' width='24%'>A lady is wearing a sleeveless pure-color shirt and long jeans</td>
+    <td align='center' width='24%'>The man wears a short-sleeve T-shirt with the pure color pattern and a short pants with the pure color pattern.</td>
+<tr>
+</table>
+
+[**[Project Page]**](https://yumingj.github.io/projects/Text2Human.html) | [**[Paper]**](https://arxiv.org/pdf/2205.15996.pdf) | [**[Dataset]**](https://github.com/yumingj/DeepFashion-MultiModal) | [**[Demo Video]**](https://youtu.be/yKh4VORA_E0)
+
+
+## Updates
+
+- [05/2022] Paper and demo video are released.
+- [05/2022] Code is released.
+- [05/2022] This website is created.
+
+## Installation
+**Clone this repo:**
+```bash
+git clone https://github.com/yumingj/Text2Human.git
+cd Text2Human
+```
+**Dependencies:**
+
+All dependencies for defining the environment are provided in `environment/text2human_env.yaml`.
+We recommend using [Anaconda](https://docs.anaconda.com/anaconda/install/) to manage the python environment:
+```bash
+conda env create -f ./environment/text2human_env.yaml
+conda activate text2human
+conda install -c huggingface tokenizers=0.9.4
+conda install -c huggingface transformers=4.0.0
+conda install -c conda-forge sentence-transformers=2.0.0
+```
+
+If it doesn't work, you may need to install the following packages on your own:
+  - Python 3.6
+  - PyTorch 1.7.1
+  - CUDA 10.1
+  - [sentence-transformers](https://huggingface.co/sentence-transformers) 2.0.0
+  - [tokenizers](https://pypi.org/project/tokenizers/) 0.9.4
+  - [transformers](https://huggingface.co/docs/transformers/installation) 4.0.0
+
+## (1) Dataset Preparation
+
+In this work, we contribute a large-scale high-quality dataset with rich multi-modal annotations named [DeepFashion-MultiModal](https://github.com/yumingj/DeepFashion-MultiModal) Dataset.
+Here we pre-processed the raw annotations of the original dataset for the task of text-driven controllable human image generation. The pre-processing pipeline consists of:
+  - align the human body in the center of the images according to the human pose
+  - fuse the clothing color and clothing fabric annotations into one texture annotation
+  - do some annotation cleaning and image filtering
+  - split the whole dataset into the training set and testing set
+
+You can download our processed dataset from this [Google Drive](https://drive.google.com/file/d/1KIoFfRZNQVn6RV_wTxG2wZmY8f2T_84B/view?usp=sharing). If you want to access the raw annotations, please refer to the [DeepFashion-MultiModal](https://github.com/yumingj/DeepFashion-MultiModal) Dataset.
+
+After downloading the dataset, unzip the file and put them under the dataset folder with the following structure:
+```
+./datasets
+├── train_images
+    ├── xxx.png
+    ...
+    ├── xxx.png
+    └── xxx.png
+├── test_images
+    % the same structure as in train_images
+├── densepose
+    % the same structure as in train_images
+├── segm
+    % the same structure as in train_images
+├── shape_ann
+    ├── test_ann_file.txt
+    ├── train_ann_file.txt
+    └── val_ann_file.txt
+└── texture_ann
+    ├── test
+        ├── lower_fused.txt
+        ├── outer_fused.txt
+        └── upper_fused.txt
+    ├── train
+        % the same files as in test
+    └── val
+        % the same files as in test
+```
+
+## (2) Sampling
+
+### Inference Notebook
+<img src="https://colab.research.google.com/assets/colab-badge.svg" height=22.5></a></br>
+Coming soon.
+
+
+### Pretrained Models
+
+Pretrained models can be downloaded from this [Google Drive](https://drive.google.com/file/d/1VyI8_AbPwAUaZJPaPba8zxsFIWumlDen/view?usp=sharing). Unzip the file and put them under the dataset folder with the following structure:
+```
+pretrained_models
+├── index_pred_net.pth
+├── parsing_gen.pth
+├── parsing_token.pth
+├── sampler.pth
+├── vqvae_bottom.pth
+└── vqvae_top.pth
+```
+
+### Generation from Paring Maps
+You can generate images from given parsing maps and pre-defined texture annotations:
+```python
+python sample_from_parsing.py -opt ./configs/sample_from_parsing.yml
+```
+The results are saved in the folder `./results/sampling_from_parsing`.
+
+### Generation from Poses
+You can generate images from given human poses and pre-defined clothing shape and texture annotations:
+```python
+python sample_from_pose.py -opt ./configs/sample_from_pose.yml
+```
+
+**Remarks**: The above two scripts generate images without language interactions. If you want to generate images using texts, you can use the notebook or our user interface.
+
+### User Interface
+
+```python
+python ui_demo.py
+```
+<img src="./assets/ui.png" width="100%">
+
+The descriptions for shapes should follow the following format:
+```
+<gender>, <sleeve length>, <length of lower clothing>, <outer clothing type>, <other accessories1>, ...
+
+Note: The outer clothing type and accessories can be omitted.
+
+Examples:
+man, sleeveless T-shirt, long pants
+woman, short-sleeve T-shirt, short jeans
+```
+
+The descriptions for textures should follow the following format:
+```
+<upper clothing texture>, <lower clothing texture>, <outer clothing texture>
+
+Note: Currently, we only support 5 types of textures, i.e., pure color, stripe/spline, plaid/lattice, 
+    floral, denim. Your inputs should be restricted to these textures.
+```
+
+## (3) Training Text2Human
+
+### Stage I: Pose to Parsing
+Train the parsing generation network. If you want to skip the training of this network, you can download our pretrained model from [here](https://drive.google.com/file/d/1MNyFLGqIQcOMg_HhgwCmKqdwfQSjeg_6/view?usp=sharing).
+```python
+python train_parsing_gen.py -opt ./configs/parsing_gen.yml
+```
+
+### Stage II: Parsing to Human
+
+**Step 1: Train the top level of the hierarchical VQVAE.**
+We provide our pretrained model [here](https://drive.google.com/file/d/1TwypUg85gPFJtMwBLUjVS66FKR3oaTz8/view?usp=sharing). This model is trained by:
+```python
+python train_vqvae.py -opt ./configs/vqvae_top.yml
+```
+
+**Step 2: Train the bottom level of the hierarchical VQVAE.**
+We provide our pretrained model [here](https://drive.google.com/file/d/15hzbY-RG-ILgzUqqGC0qMzlS4OayPdRH/view?usp=sharing). This model is trained by:
+```python
+python train_vqvae.py -opt ./configs/vqvae_bottom.yml
+```
+
+**Stage 3 & 4: Train the sampler with mixture-of-experts.** To train the sampler, we first need to train a model to tokenize the parsing maps. You can access our pretrained parsing maps [here](https://drive.google.com/file/d/1GLHoOeCP6sMao1-R63ahJMJF7-J00uir/view?usp=sharing).
+```python
+python train_parsing_token.py -opt ./configs/parsing_token.yml
+```
+
+With the parsing tokenization model, the sampler is trained by:
+```python
+python train_sampler.py -opt ./configs/sampler.yml
+```
+Our pretrained sampler is provided [here](https://drive.google.com/file/d/1OQO_kG2fK7eKiG1VJH1OL782X71UQAmS/view?usp=sharing).
+
+**Stage 5: Train the index prediction network.**
+We provide our pretrained index prediction network [here](https://drive.google.com/file/d/1rqhkQD-JGd7YBeIfDvMV-vjfbNHpIhYm/view?usp=sharing). It is trained by:
+```python
+python train_index_prediction.py -opt ./configs/index_pred_net.yml
+```
+
+
+**Remarks**: In the config files, we use the path to our models as the required pretrained models. If you want to train the models from scratch, please replace the path to your own one. We set the numbers of the training epochs as large numbers and you can choose the best epoch for each model. For your reference, our pretrained parsing generation network is trained for 50 epochs, top-level VQVAE is trained for 135 epochs, bottom-level VQVAE is trained for 70 epochs, parsing tokenization network is trained for 20 epochs, sampler is trained for 95 epochs, and the index prediction network is trained for 70 epochs.
+
+## (4) Results
+
+Please visit our [Project Page](https://yumingj.github.io/projects/Text2Human.html#results) to view more results.</br>
+You can select the attribtues to customize the desired human images.
+[<img src="./assets/results.png" width="90%">
+](https://yumingj.github.io/projects/Text2Human.html#results)
+
+## DeepFashion-MultiModal Dataset
+
+<img src="./assets/dataset_logo.png" width="90%">
+
+In this work, we also propose **DeepFashion-MultiModal**, a large-scale high-quality human dataset with rich multi-modal annotations. It has the following properties:
+1. It contains 44,096 high-resolution human images, including 12,701 full body human images.
+2. For each full body images, we **manually annotate** the human parsing labels of 24 classes.
+3. For each full body images, we **manually annotate** the keypoints.
+4. We extract DensePose for each human image.
+5. Each image is **manually annotated** with attributes for both clothes shapes and textures.
+6. We provide a textual description for each image.
+
+<img src="./assets/dataset_overview.png" width="100%">
+
+Please refer to [this repo](https://github.com/yumingj/DeepFashion-MultiModal) for more details about our proposed dataset.
+
+## TODO List
+
+- [ ] Release 1024x512 version of Text2Human.
+- [ ] Train the Text2Human using [SHHQ dataset](https://stylegan-human.github.io/).
+
+## Citation
+
+If you find this work useful for your research, please consider citing our paper:
+
+```bibtex
+@article{jiang2022text2human,
+  title={Text2Human: Text-Driven Controllable Human Image Generation},
+  author={Jiang, Yuming and Yang, Shuai and Qiu, Haonan and Wu, Wayne and Loy, Chen Change and Liu, Ziwei},
+  journal={ACM Transactions on Graphics (TOG)},
+  volume={41},
+  number={4},
+  articleno={162},
+  pages={1--11},
+  year={2022},
+  publisher={ACM New York, NY, USA},
+  doi={10.1145/3528223.3530104},
+}
+```
+
+## Acknowledgments
+
+Part of the code is borrowed from [unleashing-transformers](https://github.com/samb-t/unleashing-transformers), [taming-transformers](https://github.com/CompVis/taming-transformers) and [mmsegmentation](https://github.com/open-mmlab/mmsegmentation).

app.py

@@ -0,0 +1,158 @@
+#!/usr/bin/env python
+
+from __future__ import annotations
+
+import argparse
+import os
+import pathlib
+import subprocess
+
+import gradio as gr
+
+if os.getenv('SYSTEM') == 'spaces':
+    import mim
+
+    mim.uninstall('mmcv-full', confirm_yes=True)
+    mim.install('mmcv-full==1.5.2', is_yes=True)
+
+    with open('patch') as f:
+        subprocess.run('patch -p1'.split(), cwd='Text2Human', stdin=f)
+
+from model import Model
+
+DESCRIPTION = '''# Text2Human
+
+This is an unofficial demo for <a href="https://github.com/yumingj/Text2Human">https://github.com/yumingj/Text2Human</a> made by <a href="https://huggingface.co/spaces/hysts/Text2Human">@hysts</a>.
+You can modify sample steps and seeds. By varying seeds, you can sample different human images under the same pose, shape description, and texture description. The larger the sample steps, the better quality of the generated images. (The default value of sample steps is 256 in the original repo.)
+
+Label image generation step can be skipped. However, in that case, the input label image must be 512x256 in size and must contain only the specified colors.
+'''
+FOOTER = '<img id="visitor-badge" alt="visitor badge" src="https://visitor-badge.glitch.me/badge?page_id=hysts.text2human" />'
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--device', type=str, default='cpu')
+    parser.add_argument('--theme', type=str)
+    parser.add_argument('--share', action='store_true')
+    parser.add_argument('--port', type=int)
+    parser.add_argument('--disable-queue',
+                        dest='enable_queue',
+                        action='store_false')
+    return parser.parse_args()
+
+
+def set_example_image(example: list) -> dict:
+    return gr.Image.update(value=example[0])
+
+
+def set_example_text(example: list) -> dict:
+    return gr.Textbox.update(value=example[0])
+
+
+def main():
+    args = parse_args()
+    model = Model(args.device)
+
+    with gr.Blocks(theme=args.theme, css='style.css') as demo:
+        gr.Markdown(DESCRIPTION)
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Row():
+                    input_image = gr.Image(label='Input Pose Image',
+                                           type='pil',
+                                           elem_id='input-image')
+                    pose_data = gr.Variable()
+                with gr.Row():
+                    paths = sorted(pathlib.Path('pose_images').glob('*.png'))
+                    example_images = gr.Dataset(components=[input_image],
+                                                samples=[[path.as_posix()]
+                                                         for path in paths])
+
+                with gr.Row():
+                    shape_text = gr.Textbox(
+                        label='Shape Description',
+                        placeholder=
+                        '''<gender>, <sleeve length>, <length of lower clothing>, <outer clothing type>, <other accessories1>, ...
+Note: The outer clothing type and accessories can be omitted.''')
+                with gr.Row():
+                    shape_example_texts = gr.Dataset(
+                        components=[shape_text],
+                        samples=[['man, sleeveless T-shirt, long pants'],
+                                 ['woman, short-sleeve T-shirt, short jeans']])
+                with gr.Row():
+                    generate_label_button = gr.Button('Generate Label Image')
+
+            with gr.Column():
+                with gr.Row():
+                    label_image = gr.Image(label='Label Image',
+                                           type='numpy',
+                                           elem_id='label-image')
+
+                with gr.Row():
+                    texture_text = gr.Textbox(
+                        label='Texture Description',
+                        placeholder=
+                        '''<upper clothing texture>, <lower clothing texture>, <outer clothing texture>
+Note: Currently, only 5 types of textures are supported, i.e., pure color, stripe/spline, plaid/lattice, floral, denim.'''
+                    )
+                with gr.Row():
+                    texture_example_texts = gr.Dataset(
+                        components=[texture_text],
+                        samples=[['pure color, denim'], ['floral, stripe']])
+                with gr.Row():
+                    sample_steps = gr.Slider(10,
+                                             300,
+                                             value=10,
+                                             step=10,
+                                             label='Sample Steps')
+                with gr.Row():
+                    seed = gr.Slider(0, 1000000, value=0, step=1, label='Seed')
+                with gr.Row():
+                    generate_human_button = gr.Button('Generate Human')
+
+            with gr.Column():
+                with gr.Row():
+                    result = gr.Image(label='Result',
+                                      type='numpy',
+                                      elem_id='result-image')
+
+        gr.Markdown(FOOTER)
+
+        input_image.change(fn=model.process_pose_image,
+                           inputs=input_image,
+                           outputs=pose_data)
+        generate_label_button.click(fn=model.generate_label_image,
+                                    inputs=[
+                                        pose_data,
+                                        shape_text,
+                                    ],
+                                    outputs=label_image)
+        generate_human_button.click(fn=model.generate_human,
+                                    inputs=[
+                                        label_image,
+                                        texture_text,
+                                        sample_steps,
+                                        seed,
+                                    ],
+                                    outputs=result)
+        example_images.click(fn=set_example_image,
+                             inputs=example_images,
+                             outputs=example_images.components)
+        shape_example_texts.click(fn=set_example_text,
+                                  inputs=shape_example_texts,
+                                  outputs=shape_example_texts.components)
+        texture_example_texts.click(fn=set_example_text,
+                                    inputs=texture_example_texts,
+                                    outputs=texture_example_texts.components)
+
+    demo.launch(
+        enable_queue=args.enable_queue,
+        server_port=args.port,
+        share=args.share,
+    )
+
+
+if __name__ == '__main__':
+    main()

model.py

@@ -0,0 +1,147 @@
+from __future__ import annotations
+
+import os
+import pathlib
+import sys
+import zipfile
+
+import huggingface_hub
+import numpy as np
+import PIL.Image
+import torch
+
+sys.path.insert(0, 'Text2Human')
+
+from models.sample_model import SampleFromPoseModel
+from utils.language_utils import (generate_shape_attributes,
+                                  generate_texture_attributes)
+from utils.options import dict_to_nonedict, parse
+from utils.util import set_random_seed
+
+COLOR_LIST = [
+    (0, 0, 0),
+    (255, 250, 250),
+    (220, 220, 220),
+    (250, 235, 215),
+    (255, 250, 205),
+    (211, 211, 211),
+    (70, 130, 180),
+    (127, 255, 212),
+    (0, 100, 0),
+    (50, 205, 50),
+    (255, 255, 0),
+    (245, 222, 179),
+    (255, 140, 0),
+    (255, 0, 0),
+    (16, 78, 139),
+    (144, 238, 144),
+    (50, 205, 174),
+    (50, 155, 250),
+    (160, 140, 88),
+    (213, 140, 88),
+    (90, 140, 90),
+    (185, 210, 205),
+    (130, 165, 180),
+    (225, 141, 151),
+]
+
+
+class Model:
+    def __init__(self, device: str):
+        self.config = self._load_config()
+        self.config['device'] = device
+        self._download_models()
+        self.model = SampleFromPoseModel(self.config)
+        self.model.batch_size = 1
+
+    def _load_config(self) -> dict:
+        path = 'Text2Human/configs/sample_from_pose.yml'
+        config = parse(path, is_train=False)
+        config = dict_to_nonedict(config)
+        return config
+
+    def _download_models(self) -> None:
+        model_dir = pathlib.Path('pretrained_models')
+        if model_dir.exists():
+            return
+        token = os.getenv('HF_TOKEN')
+        path = huggingface_hub.hf_hub_download('yumingj/Text2Human',
+                                               'pretrained_models.zip',
+                                               use_auth_token=token)
+        model_dir.mkdir()
+        with zipfile.ZipFile(path) as f:
+            f.extractall(model_dir)
+
+    @staticmethod
+    def preprocess_pose_image(image: PIL.Image.Image) -> torch.Tensor:
+        image = np.array(
+            image.resize(
+                size=(256, 512),
+                resample=PIL.Image.Resampling.LANCZOS))[:, :, 2:].transpose(
+                    2, 0, 1).astype(np.float32)
+        image = image / 12. - 1
+        data = torch.from_numpy(image).unsqueeze(1)
+        return data
+
+    @staticmethod
+    def process_mask(mask: np.ndarray) -> np.ndarray:
+        if mask.shape != (512, 256, 3):
+            return None
+        seg_map = np.full(mask.shape[:-1], -1)
+        for index, color in enumerate(COLOR_LIST):
+            seg_map[np.sum(mask == color, axis=2) == 3] = index
+        if not (seg_map != -1).all():
+            return None
+        return seg_map
+
+    @staticmethod
+    def postprocess(result: torch.Tensor) -> np.ndarray:
+        result = result.permute(0, 2, 3, 1)
+        result = result.detach().cpu().numpy()
+        result = result * 255
+        result = np.asarray(result[0, :, :, :], dtype=np.uint8)
+        return result
+
+    def process_pose_image(self, pose_image: PIL.Image.Image) -> torch.Tensor:
+        if pose_image is None:
+            return
+        data = self.preprocess_pose_image(pose_image)
+        self.model.feed_pose_data(data)
+        return data
+
+    def generate_label_image(self, pose_data: torch.Tensor,
+                             shape_text: str) -> np.ndarray:
+        if pose_data is None:
+            return
+        self.model.feed_pose_data(pose_data)
+        shape_attributes = generate_shape_attributes(shape_text)
+        shape_attributes = torch.LongTensor(shape_attributes).unsqueeze(0)
+        self.model.feed_shape_attributes(shape_attributes)
+        self.model.generate_parsing_map()
+        self.model.generate_quantized_segm()
+        colored_segm = self.model.palette_result(self.model.segm[0].cpu())
+        return colored_segm
+
+    def generate_human(self, label_image: np.ndarray, texture_text: str,
+                       sample_steps: int, seed: int) -> np.ndarray:
+        if label_image is None:
+            return
+        mask = label_image.copy()
+        seg_map = self.process_mask(mask)
+        if seg_map is None:
+            return
+        self.model.segm = torch.from_numpy(seg_map).unsqueeze(0).unsqueeze(
+            0).to(self.model.device)
+        self.model.generate_quantized_segm()
+
+        set_random_seed(seed)
+
+        texture_attributes = generate_texture_attributes(texture_text)
+        texture_attributes = torch.LongTensor(texture_attributes)
+        self.model.feed_texture_attributes(texture_attributes)
+        self.model.generate_texture_map()
+
+        self.model.sample_steps = sample_steps
+        out = self.model.sample_and_refine()
+        res = self.postprocess(out)
+        return res

patch

@@ -0,0 +1,169 @@
+diff --git a/models/hierarchy_inference_model.py b/models/hierarchy_inference_model.py
+index 3116307..5de661d 100644
+--- a/models/hierarchy_inference_model.py
++++ b/models/hierarchy_inference_model.py
+@@ -21,7 +21,7 @@ class VQGANTextureAwareSpatialHierarchyInferenceModel():
+ 
+     def __init__(self, opt):
+         self.opt = opt
+-        self.device = torch.device('cuda')
++        self.device = torch.device(opt['device'])
+         self.is_train = opt['is_train']
+ 
+         self.top_encoder = Encoder(
+diff --git a/models/hierarchy_vqgan_model.py b/models/hierarchy_vqgan_model.py
+index 4b0d657..0bf4712 100644
+--- a/models/hierarchy_vqgan_model.py
++++ b/models/hierarchy_vqgan_model.py
+@@ -20,7 +20,7 @@ class HierarchyVQSpatialTextureAwareModel():
+ 
+     def __init__(self, opt):
+         self.opt = opt
+-        self.device = torch.device('cuda')
++        self.device = torch.device(opt['device'])
+         self.top_encoder = Encoder(
+             ch=opt['top_ch'],
+             num_res_blocks=opt['top_num_res_blocks'],
+diff --git a/models/parsing_gen_model.py b/models/parsing_gen_model.py
+index 9440345..15a1ecb 100644
+--- a/models/parsing_gen_model.py
++++ b/models/parsing_gen_model.py
+@@ -22,7 +22,7 @@ class ParsingGenModel():
+ 
+     def __init__(self, opt):
+         self.opt = opt
+-        self.device = torch.device('cuda')
++        self.device = torch.device(opt['device'])
+         self.is_train = opt['is_train']
+ 
+         self.attr_embedder = ShapeAttrEmbedding(
+diff --git a/models/sample_model.py b/models/sample_model.py
+index 4c60e3f..5265cd0 100644
+--- a/models/sample_model.py
++++ b/models/sample_model.py
+@@ -23,7 +23,7 @@ class BaseSampleModel():
+ 
+     def __init__(self, opt):
+         self.opt = opt
+-        self.device = torch.device('cuda')
++        self.device = torch.device(opt['device'])
+ 
+         # hierarchical VQVAE
+         self.decoder = Decoder(
+@@ -123,7 +123,7 @@ class BaseSampleModel():
+ 
+     def load_top_pretrain_models(self):
+         # load pretrained vqgan
+-        top_vae_checkpoint = torch.load(self.opt['top_vae_path'])
++        top_vae_checkpoint = torch.load(self.opt['top_vae_path'], map_location=self.device)
+ 
+         self.decoder.load_state_dict(
+             top_vae_checkpoint['decoder'], strict=True)
+@@ -137,7 +137,7 @@ class BaseSampleModel():
+         self.top_post_quant_conv.eval()
+ 
+     def load_bot_pretrain_network(self):
+-        checkpoint = torch.load(self.opt['bot_vae_path'])
++        checkpoint = torch.load(self.opt['bot_vae_path'], map_location=self.device)
+         self.bot_decoder_res.load_state_dict(
+             checkpoint['bot_decoder_res'], strict=True)
+         self.decoder.load_state_dict(checkpoint['decoder'], strict=True)
+@@ -153,7 +153,7 @@ class BaseSampleModel():
+ 
+     def load_pretrained_segm_token(self):
+         # load pretrained vqgan for segmentation mask
+-        segm_token_checkpoint = torch.load(self.opt['segm_token_path'])
++        segm_token_checkpoint = torch.load(self.opt['segm_token_path'], map_location=self.device)
+         self.segm_encoder.load_state_dict(
+             segm_token_checkpoint['encoder'], strict=True)
+         self.segm_quantizer.load_state_dict(
+@@ -166,7 +166,7 @@ class BaseSampleModel():
+         self.segm_quant_conv.eval()
+ 
+     def load_index_pred_network(self):
+-        checkpoint = torch.load(self.opt['pretrained_index_network'])
++        checkpoint = torch.load(self.opt['pretrained_index_network'], map_location=self.device)
+         self.index_pred_guidance_encoder.load_state_dict(
+             checkpoint['guidance_encoder'], strict=True)
+         self.index_pred_decoder.load_state_dict(
+@@ -176,7 +176,7 @@ class BaseSampleModel():
+         self.index_pred_decoder.eval()
+ 
+     def load_sampler_pretrained_network(self):
+-        checkpoint = torch.load(self.opt['pretrained_sampler'])
++        checkpoint = torch.load(self.opt['pretrained_sampler'], map_location=self.device)
+         self.sampler_fn.load_state_dict(checkpoint, strict=True)
+         self.sampler_fn.eval()
+ 
+@@ -397,7 +397,7 @@ class SampleFromPoseModel(BaseSampleModel):
+                         [185, 210, 205], [130, 165, 180], [225, 141, 151]]
+ 
+     def load_shape_generation_models(self):
+-        checkpoint = torch.load(self.opt['pretrained_parsing_gen'])
++        checkpoint = torch.load(self.opt['pretrained_parsing_gen'], map_location=self.device)
+ 
+         self.shape_attr_embedder.load_state_dict(
+             checkpoint['embedder'], strict=True)
+diff --git a/models/transformer_model.py b/models/transformer_model.py
+index 7db0f3e..4523d17 100644
+--- a/models/transformer_model.py
++++ b/models/transformer_model.py
+@@ -21,7 +21,7 @@ class TransformerTextureAwareModel():
+ 
+     def __init__(self, opt):
+         self.opt = opt
+-        self.device = torch.device('cuda')
++        self.device = torch.device(opt['device'])
+         self.is_train = opt['is_train']
+ 
+         # VQVAE for image
+@@ -317,10 +317,10 @@ class TransformerTextureAwareModel():
+     def sample_fn(self, temp=1.0, sample_steps=None):
+         self._denoise_fn.eval()
+ 
+-        b, device = self.image.size(0), 'cuda'
++        b = self.image.size(0)
+         x_t = torch.ones(
+-            (b, np.prod(self.shape)), device=device).long() * self.mask_id
+-        unmasked = torch.zeros_like(x_t, device=device).bool()
++            (b, np.prod(self.shape)), device=self.device).long() * self.mask_id
++        unmasked = torch.zeros_like(x_t, device=self.device).bool()
+         sample_steps = list(range(1, sample_steps + 1))
+ 
+         texture_mask_flatten = self.texture_tokens.view(-1)
+@@ -336,11 +336,11 @@ class TransformerTextureAwareModel():
+ 
+         for t in reversed(sample_steps):
+             print(f'Sample timestep {t:4d}', end='\r')
+-            t = torch.full((b, ), t, device=device, dtype=torch.long)
++            t = torch.full((b, ), t, device=self.device, dtype=torch.long)
+ 
+             # where to unmask
+             changes = torch.rand(
+-                x_t.shape, device=device) < 1 / t.float().unsqueeze(-1)
++                x_t.shape, device=self.device) < 1 / t.float().unsqueeze(-1)
+             # don't unmask somewhere already unmasked
+             changes = torch.bitwise_xor(changes,
+                                         torch.bitwise_and(changes, unmasked))
+diff --git a/models/vqgan_model.py b/models/vqgan_model.py
+index 13a2e70..9c840f1 100644
+--- a/models/vqgan_model.py
++++ b/models/vqgan_model.py
+@@ -20,7 +20,7 @@ class VQModel():
+     def __init__(self, opt):
+         super().__init__()
+         self.opt = opt
+-        self.device = torch.device('cuda')
++        self.device = torch.device(opt['device'])
+         self.encoder = Encoder(
+             ch=opt['ch'],
+             num_res_blocks=opt['num_res_blocks'],
+@@ -390,7 +390,7 @@ class VQImageSegmTextureModel(VQImageModel):
+ 
+     def __init__(self, opt):
+         self.opt = opt
+-        self.device = torch.device('cuda')
++        self.device = torch.device(opt['device'])
+         self.encoder = Encoder(
+             ch=opt['ch'],
+             num_res_blocks=opt['num_res_blocks'],

requirements.txt

@@ -0,0 +1,12 @@
+einops==0.4.1
+lpips==0.1.4
+mmcv-full==1.5.2
+mmsegmentation==0.24.1
+numpy==1.22.3
+openmim==0.1.5
+Pillow==9.1.1
+sentence-transformers==2.2.0
+tokenizers==0.12.1
+torch==1.11.0
+torchvision==0.12.0
+transformers==4.19.2

style.css

@@ -0,0 +1,16 @@
+h1 {
+  text-align: center;
+}
+#input-image {
+  max-height: 300px;
+}
+#label-image {
+  height: 300px;
+}
+#result-image {
+  height: 300px;
+}
+img#visitor-badge {
+  display: block;
+  margin: auto;
+}