Abstract:
While Generative Adversarial Networks (GANs) have seen huge successes in image synthesis tasks, they are notoriously difficult to use, in part due to instability during training. One commonly accepted reason for this instability is that gradients passing from the discriminator to the generator can quickly become uninformative, due to a learning imbalance during training. In this work, we propose the Multi-Scale Gradient Generative Adversarial Network (MSG-GAN), a simple but effective technique for addressing this problem which allows the flow of gradients from the discriminator to the generator at multiple scales. This technique provides a stable approach for generating synchronized multi-scale images. We present a very intuitive implementation of the mathematical MSG-GAN framework which uses the concatenation operation in the discriminator computations. We empirically validate the effect of our MSG-GAN approach through experiments on the CIFAR10 and Oxford102 flowers datasets and compare it with other relevant techniques which perform multi-scale image synthesis. In addition, we also provide details of our experiment on CelebA-HQ dataset for synthesizing 1024 x 1024 high resolution images.

The above figure describes the architecture of MSG-GAN for generating synchronized multi-scale images. Our method is based on the architecture proposed in proGAN, but instead of a progressively growing training scheme, includes connections from the intermediate layers of the generator to the intermediate layers of the discriminator. The multi-scale images input to the discriminator are converted into spatial volumes which are concatenated with the corresponding activation volumes obtained from the main path of convolutional layers.

For the discrimination process, appropriately downsampled versions of the real images are fed to corresponding layers of the discriminator as shown in the diagram (from above).

Above figure explains how, during training, all the layers in the MSG-GAN first synchronize colour-wise and subsequently improve the generated images at various scales. The brightness of the images across all layers (scales) synchronizes eventually

Please note to use value of learning_rate=0.003 for both G and D for all experiments for best results. The model is quite robust and converges to a very similar FID or IS very quickly even for different learning rate settings. Please use the relativistic-hinge as the loss function (set as default) for training.

Start the training by running the train.py script in the sourcecode/ directory. Refer to the following parameters for tweaking for your own use:

-h, --help            show this help message and exit
  --generator_file GENERATOR_FILE
                        pretrained weights file for generator
  --generator_optim_file GENERATOR_OPTIM_FILE
                        saved state for generator optimizer
  --shadow_generator_file SHADOW_GENERATOR_FILE
                        pretrained weights file for the shadow generator
  --discriminator_file DISCRIMINATOR_FILE
                        pretrained_weights file for discriminator
  --discriminator_optim_file DISCRIMINATOR_OPTIM_FILE
                        saved state for discriminator optimizer
  --images_dir IMAGES_DIR
                        path for the images directory
  --folder_distributed FOLDER_DISTRIBUTED
                        whether the images directory contains folders or not
  --flip_augment FLIP_AUGMENT
                        whether to randomly mirror the images during training
  --sample_dir SAMPLE_DIR
                        path for the generated samples directory
  --model_dir MODEL_DIR
                        path for saved models directory
  --loss_function LOSS_FUNCTION
                        loss function to be used: standard-gan, wgan-gp,
                        lsgan,lsgan-sigmoid,hinge, relativistic-hinge
  --depth DEPTH         Depth of the GAN
  --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
  --num_samples NUM_SAMPLES
                        number of samples to generate for creating the grid
                        should be a square number preferably
  --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
  --adam_beta1 ADAM_BETA1
                        value of beta_1 for adam optimizer
  --adam_beta2 ADAM_BETA2
                        value of beta_2 for adam optimizer
  --use_eql USE_EQL     Whether to use equalized learning rate or not
  --use_ema USE_EMA     Whether to use exponential moving averages or not
  --ema_decay EMA_DECAY
                        decay value for the ema
  --data_percentage DATA_PERCENTAGE
                        percentage of data to use
  --num_workers NUM_WORKERS
                        number of parallel workers for reading files

For training a network at resolution 256 x 256, use the following arguments:

$ python train.py --depth=7 \ 
                  --latent_size=512 \
                  --images_dir=<path to images> \
                  --sample_dir=samples/exp_1 \
                  --model_dir=models/exp_1

Set the batch_size, feedback_factor and checkpoint_factor accordingly. We used 2 Tesla V100 GPUs of the DGX-1 machine for our experimentation.

CelebA HQ (30K dataset)

LSUN Bedrooms (3M dataset)

CelebA (200K dataset)

Cifar-10 (50K dataset)

Oxford-102 Flowers (8K dataset)

@article{karnewar2019msg,
  title={MSG-GAN: Multi-Scale Gradient GAN for Stable Image Synthesis},
  author={Karnewar, Animesh and Wang, Oliver and Iyengar, Raghu Sesha},
  journal={arXiv preprint arXiv:1903.06048},
  year={2019}
}

Please feel free to open PRs here if you train on other datasets using this architecture.

Best regards,
@akanimax :)