A variant of the Self Attention GAN named: FAGAN (Full Attention GAN). The architecture of this gan contains the full attention layer as proposed in this project.

Celeba samples

celeba training samples

samples generated during training of the proposed architecture
on the celeba dataset.

Full attention layer

proposed full attention layer

The above figure describes the architecture of the proposed full attention layer. As you can see, on the upper path we compute traditional convolution output and the lower path, we have an attention layer which generalises to (k x k) convolution filters instead of just (1 x 1) filters. The alpha shown in the residual calculation is a trainable parameter.

Now why is the lower path not self attention? The reason for it is that while computing the attention maps, the input is first locally aggregated by the (k x k) convolutions, and therefore is no longer just self attention since it uses a small spatially neighbouring area into computations. Given enough depth and filter size, we could cover the entire input image as a receptive field for a subsequent attention calculation, hence the name: Full Attention.

Celeba Experiment

Hinge-Gan loss experiment

The following diagram is the plot of the loss (Hinge-GAN) generated from
the loss-logs obtained during training.

celeba loss plot

Relativistic Hinge-Gan loss experiment

Following are some of the samples obtained while training using the
relativistic hinge-gan loss function proposed in the paper

relativistic celeba training samples

The training of the relativistic version is percetually better (stabler)
as seen from the samples gif. Refer to the following loss_plot
for this experiment for more info.

relativistic celeba loss plot

Running the Code

Running the training is actually very simple.
Just install the attn_gan_pytorch package using the following command

$ workon [your virtual environment]
$ pip install attn-gan-pytorch

And then run the training by running the train.py script. Refer to
the following parameters for tweaking for your own use:

optional arguments:
  -h, --help            show this help message and exit
  --generator_config GENERATOR_CONFIG
                        default configuration for generator network
  --discriminator_config DISCRIMINATOR_CONFIG
                        default configuration for discriminator network
  --generator_file GENERATOR_FILE
                        pretrained weights file for generator
  --discriminator_file DISCRIMINATOR_FILE
                        pretrained_weights file for discriminator
  --images_dir IMAGES_DIR
                        path for the images directory
  --sample_dir SAMPLE_DIR
                        path for the generated samples directory
  --model_dir MODEL_DIR
                        path for saved models directory
  --latent_size LATENT_SIZE
                        latent size for the generator
  --batch_size BATCH_SIZE
                        batch_size for training
  --start START         starting epoch number
  --num_epochs NUM_EPOCHS
                        number of epochs for training
  --feedback_factor FEEDBACK_FACTOR
                        number of logs to generate per epoch
  --checkpoint_factor CHECKPOINT_FACTOR
                        save model per n epochs
  --g_lr G_LR           learning rate for generator
  --d_lr D_LR           learning rate for discriminator
  --data_percentage DATA_PERCENTAGE
                        percentage of data to use
  --num_workers NUM_WORKERS
                        number of parallel workers for reading files

Trained weights for generating cool faces :)

refer to the models/fagan_1/ directory to find the saved weights for
this model in pytorch format. For spawning the architectures,
refer to the configs/ folder for loading the generator
and discriminator configurations.