Pix2pix in Pytorch

This project is based on Image-to-Image Translation with Conditional Adversarial Networks.

Prerequisites

• Python 3.6.x
• PyTorch 0.4.x & torchvision

Dataset

Maps dataset can be downloaded from original project CycleGAN and pix2pix in PyTorch. Should choose the CycleGAN one in the format of

.
├── datasets
|   ├── maps
|   |   ├── trainA             # Training
|   |   ├── trainB
|   |   ├── valA               # Validation
|   |   ├── valB
|   |   ├── testA              # Test
|   |   ├── testB

Training

Train model from scratch

python main.py --mode train --data_dir [data_directory] --n_epoch 200 --G resnet9 --gan_loss MSE

Default data_dir is ./datasets/maps/. Source (A) and target (B) images should be in folders trainA/trainB, valA/valB, testA/testB separately.

Continue train using pretrained model

python main.py --mode train --pretrain_path ./checkpoints/xxx/xxx.pt

Need to provide same configs/options when continue to train a model.

Plot loss stats from train.json

python plot.py --dir ./checkpoints/xxx

It will look for train.json in the directory and output plots as result.png.

Loss:

See more options available

Testing

python main.py --mode test --pretrain_path ./checkpoints/xxx/xxx.pt

This generates all images from test set and save them to ./checkpoints/xxx/images/test/.

t-SNE plot of 5 random generated images from different models and ground truth: