Official Pytorch implementation for Guided Image-to-Image Translation with Bi-Directional Feature Transformation. Please contact Badour AlBahar ([email protected]) if you have any questions.
This codebase was developed and tested with:
- Python2.7
- Pytorch 0.4.1.post2
- CUDA 8.0
-
Pose transfer:
We use DeepFashion dataset. We follow the train/test splits provided by Pose guided person image generation. -
Texture transfer: We use the dataset provided by textureGAN.
-
Depth Upsampling:
We use the NYU v2 dataset.
1. Pose transfer:
python train.py --dataroot /root/DeepFashion/ --name exp_name --netG bFT_resnet --dataset_mode pose --input_nc 3 --guide_nc 18 --output_nc 3 --lr 0.0002 --niter 100 --niter_decay 0 --batch_size 8 --use_GAN --netD basic --beta1 0.9 --checkpoints_dir ./pose_checkpoints2. Texture transfer:
python train.py --dataroot /root/training_handbags_pretrain/ --name exp_name --netG bFT_unet --dataset_mode texture --input_nc 1 --guide_nc 4 --output_nc 3 --niter 100 --niter_decay 0 --batch_size 256 --lr 0.0002 --use_GAN --netD basic --n_layers 7 --beta1 .9 --checkpoints_dir ./texture_checkpoints3. Depth Upsampling:
python train.py --dataroot /root/NYU_RGBD_matfiles/ --name exp_name --netG bFT_resnet --dataset_mode depth --input_nc 1 --guide_nc 3 --output_nc 1 --lr 0.0002 --niter 500 --niter_decay 0 --batch_size 2 --checkpoints_dir ./depth_checkpoints --depthTask_scale [4, 8, or 16]You can specify which epoch to test by specifying --epoch or use the default which is the latest epoch. Results will be saved in --results_dir.
1. Pose transfer:
python test.py --dataroot /root/DeepFashion/ --name exp_name --netG bFT_resnet --dataset_mode pose --input_nc 3 --guide_nc 18 --output_nc 3 --checkpoints_dir ./pose_checkpoints --task pose --results_dir ./pose_results2. Texture transfer:
python test.py --dataroot /root/training_handbags_pretrain/ --name exp_name --netG bFT_unet --n_layers 7 --dataset_mode texture --input_nc 1 --guide_nc 4 --output_nc 3 --checkpoints_dir ./texture_checkpoints --task texture --results_dir ./texture_results3. Depth Upsampling:
python test.py --dataroot /root/NYU_RGBD_matfiles/ --name exp_name --netG bFT_resnet --dataset_mode depth --input_nc 1 --guide_nc 3 --output_nc 1 --checkpoints_dir ./depth_checkpoints --task depth --depthTask_scale [4, 8, or 16] --results_dir ./depth_results-
Download the pretrained checkpoints here.
-
Test: For example, to test the depth upsampling task with scale 16:
python test.py --dataroot /root/NYU_RGBD_matfiles/ --name depth_16 --netG bFT_resnet --dataset_mode depth --input_nc 1 --guide_nc 3 --output_nc 1 --checkpoints_dir ./checkpoints/pretrained/ --task depth --depthTask_scale 16 --results_dir ./depth_resultsYou can specify which epoch to evaluate by specifying --epoch or use the default which is the latest epoch. Results will be saved in --results_dir.
1. Pose transfer:
Please note that the inception score evaluation requires tensorflow. We evaluate with tensorflow 1.4.0.
python evaluate.py --dataroot /root/DeepFashion/ --name pose --netG bFT_resnet --dataset_mode pose --input_nc 3 --guide_nc 18 --output_nc 3 --checkpoints_dir ./checkpoints/pretrained/ --task pose --results_dir ./pose_resultsThis will save the results in --results_dir and compute both SSIM and IS metrics.
2. Texture transfer:
Please download the pretrained model of textureGAN in ./resources from bags, shoes, and clothes. For example, to test the pretrained texture transfer model for the bags dataset:
python evaluate.py --dataroot /root/training_handbags_pretrain/ --name texture_bags --netG bFT_unet --n_layers 7 --dataset_mode texture --input_nc 1 --guide_nc 4 --output_nc 3 --checkpoints_dir ./checkpoints/pretrained/ --task texture --results_dir ./texture_resultsThis will save the output of bFT and textureGAN in --results_dir for 10 random input texture patches per test image. The results can then be used to compute FID and LPIPS.
3. Depth Upsampling:
python evaluate.py --dataroot /root/NYU_RGBD_matfiles/ --name depth_16 --netG bFT_resnet --dataset_mode depth --input_nc 1 --guide_nc 3 --output_nc 1 --checkpoints_dir ./checkpoints/pretrained/ --task depth --depthTask_scale 16 --results_dir ./depth_resultsThis will save the results in --results_dir and compute their RMSE metric.
This code is heavily borrowed from CycleGAN and pix2pix in PyTorch. We thank Shih-Yang Su for the code review.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent