manicman1999 / bicyclegan-keras

BicycleGAN Implemented In Keras

BicycleGAN-Keras

BicycleGAN Implemented In Keras

Paper: https://papers.nips.cc/paper/6650-toward-multimodal-image-to-image-translation.pdf

Github: https://github.com/junyanz/BicycleGAN

Slight Modifications:

• Only 1 patch discriminator, with 16 8x8 patches
• Uses CNNs for both E and D
• Uses Pooling and Upsampling instead of strides
• Modified learning rates

Results

(1 x Labels, 1 x Ground Truth, 6 x Generated Image)

Dependencies:

• Numpy
• Pillow
• Matplotlib (Somewhat redundant)
• Tensorflow
• Keras

Be sure to load data into /data/ folder.

data/DomainA and data/DomainB are for image pairs. data/DomainAs and data/DomainBs are for non-paired images.

Run pretrained model: Run main.py, 100 sample sheets will be generated.

Train a new model: Clear parameters in declaration of BicycleGAN near the end of main.py. Comment or delete the line where an old model is loaded near the end of main.py - "model.load(6)". Set train_model to true near the end of main.py. Run main.py.

bicyclegan-keras's Issues

Charging images with BicycleGAN

Dear @manicman1999 ,

I am interested in your implementation of BicycleGAN because it seems simple and accessible to me.
At the moment, I'm trying to test your model but I have a little problem.
I get the following error:

'line 50, in
tempA = Image.open("data/DomainA/im ("+str(n)+").png")
File "/usr/local/lib/python3.5/dist-packages/PIL/Image.py", line 2652, in open
fp = builtins.open(filename, "rb")
FileNotFoundError: [Errno 2] No such file or directory: 'data/DomainA/im (2).png'

I think this error comes from the fact that there is only one image in each domain.
Speaking of which, I saw that there were paired images and images that were not. I didn't understand their usefulness in training since it seems to me that paired images are used in this model.

Rekkouza