Pix2Pix in Pytorch

Study Friendly Implementation of Pix2Pix in Pytorch

More Information: Original Paper

Identical Tensorflow Implemenation will be uploaded on HyeongminLEE's Github

GAN: [Pytorch][Tensorflow]
DCGAN: [Pytorch][Tensorflow]
InfoGAN: [Pytorch][Tensorflow]
Pix2Pix: [Pytorch][Tensorflow]
DiscoGAN: [Pytorch][Tensorflow]

1. Environments

Ubuntu 16.04
Python 3.6 (Anaconda)
Pytorch 0.2.0
Torchvision 0.1.9
PIL / random

2. Code Description

pix2pix.py: Main Code
pix2pix_test.py: Test Code after Training
network.py: Generator and Discriminator
db/download.sh: DB Downloader

3. Networks and Parameters

3.1 Hyper-Parameters

Image Size = 256x256 (Randomly Resized and Cropped)
Batch Size = 1 or 4
Learning Rate = 0.0002
Adam_beta1 = 0.5
Lambda_A = 100 (Weight of L1-Loss)

Use Batch Normalization when Batch Size = 4

Use Instance Normalization when Batch Size = 1

3.2 Generator Networks (network.py)

3.3 Discriminator Networks (network.py)

4. DB Download

./db/download.sh dataset_name

dataset_name can be one of [cityscapes, maps, edges2handbags]

cityscapes : 256x256, 2975 for Train, 500 for Val
maps : 600x500, 1096 for Train, 1098 for Val
edges2handbags : 256x256, 138567 for Train, 200 for Val

5. Train

5.1 edges2handbags

python pix2pix.py --dataroot ./datasets/edges2handbags --which_direction AtoB --num_epochs 15 --batchSize 4 --no_resize_or_crop --no_flip

5.2 maps

python pix2pix.py --dataroot ./datasets/maps --which_direction BtoA --num_epochs 100 --batchSize 1

5.3 cityscapes

python pix2pix.py --dataroot ./datasets/cityscapes --which_direction BtoA --num_epochs 100 --batchSize 1

6. Test

After finish training, saved models are in the ./models directory.

which_direction , no_resize_or_crop and no_flip options must be same with Training
batchSize means test sample size
num_epochs is the parameter which model will be used for test

6.1 edges2handbags

python pix2pix_test.py --dataroot ./datasets/edges2handbags --which_direction AtoB --num_epochs 15 --batchSize 4 --no_resize_or_crop --no_flip

6.2 maps

python pix2pix_test.py --dataroot ./datasets/maps --which_direction BtoA --num_epochs 100 --batchSize 1

6.3 cityscapes

python pix2pix_test.py --dataroot ./datasets/cityscapes --which_direction BtoA --num_epochs 100 --batchSize 1

Test results will be saved in ./test_result

7. Results

[Input | Generated | Ground Truth]

Edges to Handbags

Maps to Aerials (100 Epochs)

200 Epochs (which is in the Paper) will give better results

Semantic Labels to Street Scenes (100 Epochs)