• Pose Guided Person Image Generation (NIPS2017)
• Disentangled Person Image Generation (CVPR2018 Spotlight)
• Progressive Pose Attention Transfer for Person Image Generation (CVPR2019 Oral)
• Pose-normalized image generation for person re-identification

• PyTorch 1.1+(if you use Pytorch 1.0, tensorboardX is also needed.)
• ignite
• torchvision
• TensorBoard
• numpy
• tqdm

For fair comparison, all implementation use 263,632 training pairs and 12,000 testing pairs from Market-1501 as in PATN

copy&rename above pair and annotation file to ./data

Finally, your ./data folder looks like:

data
├── market
│   ├── annotation-test.csv
│   ├── annotation-train.csv
│   ├── pairs-test.csv
│   ├── pairs-train.csv
│   ├── attribute
│   │   ├── evaluate_market_attribute.m
│   │   ├── gallery_market.mat
│   │   ├── market_attribute.mat
│   │   ├── README.md
│   │   └── sample_image.jpg
│   ├── test # WILL BE GENERATED IN NEXT STEP
│   │   ├── pose_map_image
│   │   └── pose_mask_image
│   └── train # WILL BE GENERATED IN NEXT STEP
│       ├── pose_map_image
│       └── pose_mask_image

1. python3 tool/generate_pose_map_add_mask.py --type train
2. python3 tool/generate_pose_map_add_mask.py --type test

Obviously, there is a lot of duplicate code between different implementations. In order to avoid repeating myself, I introduced a concept called engine. Whether it is training or testing, it can be seen as an engine, and some repetitive tasks (like loading config, preparing GPU) are done in the run.py. run.py is the start point to train or test.

the usage of run.py:

$./run.py -h
usage: Train [-h] -g GPU_ID -c CONFIG -o OUTPUT [-t TOML]
             {PG2-1,PG2-2,PG2-Generator}

positional arguments:
  {PG2-1,PG2-2,PG2-Generator}
                        run which?

optional arguments:
  -h, --help            show this help message and exit
  -g GPU_ID, --gpu_id GPU_ID
                        gpu_id: e.g. 0
  -c CONFIG, --config CONFIG
                        config file path
  -o OUTPUT, --output OUTPUT
                        output path
  -t TOML, --toml TOML  overwrite toml config use cli arg

During training, you can inspect log/generated_images/model_weights with tensorboard:

# tensorboard --logdir </path/to/checkpoint> --port <port>
tensorboard --logdir ./checkpoints/PG2-1 --port 8000

# ./run.py <engine_name> -g <gpu_id> -c </path/to/config> -o </path/to/checkpoint>
# start to train PG2 stage1.
./run.py PG2-1 -g 1 -c ./implementations/PG2/stage1.toml -o ./checkpoints/PG2-1
# the way to generated images
./run.py PG2-Generator -g 1 -c ./implementations/PG2/stage2.toml -o ./generated_images/PG2

I use TOML as the config format, and you can overwrite the config file with cli arg like this:

./run.py PG2-1 -g 1 -c ./implementations/PG2/stage1.toml -o ./checkpoints/PG2-1 \
        -t "loss.mask_l1.weight=20" -t "train.data.replacement=true"

So, if you wang to specified generated images amount(default: full/12000), you can add a option: -t "generated_limit=100"

the full command example:

./run.py PG2-Generator -g 1 -c ./implementations/PG2/stage2.toml -t "generated_limit=100"  -t "model.generator1.pretrained_path='./checkpoint/PG2-1/network_G1_26000.pth'" -t "model.generator2.pretrained_path='./checkpoint/PG2-2_26000/network_G2_13000.pth'" -o generated_images

First, please change dataset.path.train.image in ./implementations/PG2/stage[1|2].toml

train stage 1: ./run.py PG2-1 --gpu_id 0 -c ./implementations/PG2/stage1.toml -o ./checkpoints/PG2-1

train stage 2: ./run.py PG2-2 -g 2 -c ./implementations/PG2/stage2.toml -o ./checkpoints/PG2-2

generate images: ./run.py PG2-Generator -c ./implementations/PG2/stage2.toml -o ./generated_images -g 3

generate a grid image as example: python tool/generate_grid.py -r ./generated_images -o images.jpg

Please note that this is NOT the best result, the result is just not bad. it can not be used in academic papers.

Pre-trained model can be downloaded Google Drive or Baidu Disk 提取码(code): n9nj

I also provided the tensorboard log file.

# download&extract files above.
unzip weights.zip
# generate all test images
./run.py PG2-Generator -c ./implementations/PG2/stage2.toml -o ./generated_images -g 3  -t "model.generator1.pretrained_path='path/to/weights/G1.pth'" -t "model.generator2.pretrained_path='path/to/weights/G2.pth'"
# random select some images to display
`python tool/generate_grid.py -r ./generated_images -o images.jpg`
# see training logs and images.
tensorboard --logdir path/to/weights/train2 --port 8080

For fair comparisons, I just copy&use the same evaluation codes in previous works, like Deform, PG2 and PATN.

I recommend using docker to evaluate the result because evaluation codes use some outdated frameworks(Tensorflow 1.4.1).

So, next:

1. build docker image with ./evaluate/Dockerfile
2. run evaluate script

$ cd evaluate
$ docker build -t hpt_evaluate . 
$  # For user in China, you can build docker image like this:
$ docker build -t hpt_evaluate . --build-arg PIP_PYPI="https://pypi.tuna.tsinghua.edu.cn/simple"
$ cd ..
$ docker run -v $(pwd):/tmp -e NVIDIA_VISIBLE_DEVICES=0 -w /tmp --runtime=nvidia -it --rm hpt_evaluate:latest python evaluate/getMetrics_market.py

Or use image tensorflow/tensorflow:1.4.1-gpu-py3 to evaluate in docker bash:

docker run -v $(pwd):/tmp -w /tmp --runtime=nvidia -it --rm tensorflow/tensorflow:1.4.1-gpu-py3 bash
# now in docker:
$ pip install scikit-image tqdm 
$ python evaluate/getMetrics_market.py

Liqian Ma - PG2's Tensorflow implementation Thanks for his patience. (￣▽￣)"

@tengteng95 - Pose-Transfer for clear code structure and his great paper.