| Monet's painting | Generated Photo* |
|---|---|
 |
 |
| Checkout more examples here! | *weights were downloaded from here |
A Pytorch Lightning implementation of the research paper Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. Given two groups of images (we use Monet's paintings and a bunch of real images here), we want a function that takes an image from one domain and translates it to another domain. A vanilla generator maps a latent distribution to a desired output distribution. Instead of a vector sampled from a normal distribution, the cyclegan generator uses an image sampled from the input domain. The generator tries to fool the discriminator, a network that is being trained to classify if an image belongs to the target domain.
Pytorch Lightning simplifies the development of complex models. Personally, I find it easier to write code for new models in pytorch because it is easier to debug. The downside is that I end up spending too much time in writing training loops and checkpointing functionality. Enter pytorch-lightning. All I have to do now is to write a function for calculating loss from a single training batch (precisely which is the novelty presented in the paper). Everything else, the boilerplate - checkpointing, progressbar and logging is already implemented.
# Pytorch
At <https://pytorch.org>, enter your preferences and run the command that shows up.
# Pytorch Lightning
$ pip install pytorch-lightning
# Others
$ pip install torchsummary Pillow tqdm unittest flask- Cycle Gan research paper
- pytorch-CycleGAN-and-pix2pix, the original repository
- Tensorflow tutorial on cyclegan
- To reduce model oscillation while training, the authors use the strategy from Shrivastava et al.. A pool of 50 latest generated images is stored. An image sampled from this pool is set as input to the discriminator during training. The code for maintaining a pool of images is at utils.py and was taken directly from this file in the original repository. The weight initialization code has been taken from networks.py.
- To verify my implementation of the network architecture, I cloned the original pytorch cyclegan repository, sampled a random tensor input and made a forward pass on their generator and discriminator networks. These tensors have been saved here folder as .pt files. I made a forward pass on the network I implemented and made sure that the output was the same. For reproducibility of network initialization, torch seed was set to zero.
- For generating the above image, I initialized the network defined at ResnetGenerator.py with weights from this website - I do not own multiple GPUs, I am training 5 epochs a day on colab xP