Study Friendly Implementation of DiscoGAN in Pytorch

More Information: Original Paper

Implemenation based on Official Implementation, but Simplified.

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

• Ubuntu 16.04
• Python 3.6 (Anaconda)
• Pytorch 0.2.0
• Torchvision 0.1.9
• PIL
• cv2 (OpenCV) (pip install python-opencv)

• discogan.py: Main Code
• discogan_test.py: Test Code after Training
• network.py: Generator and Discriminator
• db/download.sh: DB Downloader (Edges/Shoes/Handbags)
• db/download.py: DB Downloader (Facescrub)

• Image Size = 64x64
• Batch Size = 64
• Learning Rate = 0.0002
• Weight Decay = 0.00001
• Adam_beta1 = 0.5
• Loss Weights: See the code
• Scheduling: See the code

./db/download.sh dataset_name

dataset_name can be one of [edges2shoes, edges2handbags]

You can do handbags2shoes using both datasets.

• edges2shoes : 600x500, 1096 for Train, 1098 for Val
• edges2handbags : 256x256, 138567 for Train, 200 for Val

python ./db/download.py

This code downloads face image independently. So there are some problems.

After download images 10~20k,

You should remove some broken images MANUALLY. :<

• facescrub/actors/face : Various Size
• facescrub/actresses/face : Various Size

python discogan.py --task edges2shoes #(or handbags)

python discogan.py --task handbags2shoes --starting_rate 0.5

python discogan.py --task facescrubs

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

python discogan_test.py --task taskname --num_epochs N --batctSize M

• batchSize means test sample size.
• num_epochs is the parameter which model will be used for test.

Test results will be saved in ./test_result