Awesome

img2ImgGAN

Implementation of the paper : Toward Multimodal Image-to-Image Translation

• Link to the paper : arXiv:1711.11586
• PyTorch implementation: Link
• Summary of the paper: Gist

Results

Contents:

• Tensorboard Visualization of models
  • cVAE-GAN
  • cLR-GAN
  • Bicycle-GAN
• Dependencies
• Structure
• Setup
• Usage
  • Generate Graph
  • Training the network
  • Testing the network
• Visualizations
• TODO
• Licence

<a name='tensorboard'></a>

Model Architecture Visualization

• Network

Fig 1: Structure of BicycleGAN. (Image taken from the paper)

• Tensorboard visualization of the entire network

<a name='cvae'></a>

cVAE-GAN Network

<a name='depends'></a>

Dependencies

• tensorflow (1.4.0)
• numpy (1.13.3)
• scikit-image (0.13.1)
• scipy (1.0.0)

To install the above dependencies, run:

$ sudo pip install -r requirements.txt

<a name='struct'></a>

Structure

 -img2imgGAN/
            -nnet
            -utils
            -data/
                  -edges2handbags
                  -edges2shoes
                  -facades
                  -maps

<a name='setup'></a>

Setup

• Download the datasets from the following links

  • edges2handbags
  • edges2shoes
  • facades
  • maps

• To generate numpy files for the datasets,

  $ python main.py --create <dataset_name>
  

  This creates train.npy and val.npy in the corresponding dataset directory. This generates very huge files. As an alternate, the next step attempts to read images at run-time during training

• Alternate to the above step, you could read the images in real time during training. To do this, you should create files containing paths to the images. This can be done by running the following script in the root of this repo.

  $ bash setup_dataset.sh
  

<a name='usage'></a>

Usage

<a name='graph'></a>

• Generating graph:

  To visualize the connections between the graph nodes, we can generate the graph using the flag archi. This would be useful to assert the connections are correct. This generates the graph for bicycleGAN

  $ python main.py --archi
  

  To generate the model graph for cvae-gan,

  $ python main.py --model cvae-gan --archi
  

  Possible models are: cvae-gan, clr-gan, bicycle (default)

  To visualize the graph on tensorboard, run the following command:

  $ tensorboard --logdir=logs/summary/Run_1 --host=127.0.0.1
  

  Replace Run_1 with the latest directory name

• Complete list of options:

  $ python main.py --help
  

<a name='train'></a>

• Training the network

  To train model (say cvae-gan) on dataset (say facades) from scratch,

  $ python main.py --train --model cvae-gan --dataset facades
  

  The above command by default trains the model in which images from distribution of domain B are generated conditioned on the images from the distribution of domain A. To switch the direction,

  $ python main.py --train --model cvae-gan --dataset facades --direction b2a
  

  To resume the training from a checkpoint,

  $ python main.py --resume <path_to_checkpoint> --model cvae-gan
  

<a name='test'></a>

• Testing the network

  • Download the checkpoint file from here and place the checkpoint files in the ckpt directory

  To test the model from the given trained models, by default the model generates 5 different images (by sampling 5 different noise samples)

  $ ./test.sh <dataset_name> <test_image_path>
  

  To generate multiple output samples,

  $ ./test.sh <dataset_name> <test_image_path> < # of samples>
  

  Try it with some of the test samples present in the directory imgs/test

<a name='viz'></a>

Visualizations

Loss of discriminator and generator as function of iterations on edges2shoes dataset.

<a name='todo'></a>

TODO

• Residual Encoder
• Multiple discriminators for cVAE-GAN and cLR-GAN
• Inducing noise to all the layers of the generator
• Train the model on rest of the datasets

<a name='licence'></a>

License

Released under the MIT license