- All of Monet's paintings parsed from wikiart.org
- Stanford dog dataset from Kaggle
Details
- for the dog dataset, images that meet both of the following conditions are used during training
- object size to image size ratio greater than 0.8
- both image hight and width are greater than 256
- however, while training, images with one side greater than 512 are likely to cause memory issues.
- therefore, images are resized to be less than or equal to 512x512 in the data generator
- in the end 1362 pairs of images are used in the training with monet's paintings being the bottleneck
# build the container
docker build -f Dockerfile.train -t monet_cyclegan .
# start the container
docker run --gpus all -v $(pwd):/work -p 8888:8888 -p 6006:6006 -it monet_cyclegan
# start jupyter notebook
jupyter notebook --ip=0.0.0.0 --allow-root
# see details about the following steps in the notebook folder
# data preparation
# model training
# model preparation for Tensorflow Serving- an image buffer that stores previously generated images. this is used to update discriminators using a history of generated images
- linearly decreases the learning rate to 0 only after the first 100 epochs
- BCE and MSE loss
- U-Net and ResNet generator
- ResNet based generator and 3-layer PatchGAN discriminator
- LS GAN loss
- image buffer with pool size: 50
- learning rate decay (linearly towards 0 after the first 100 epochs)
- 200 epochs with batch size 1
click here
Check it out !
