Skip to content
MSG-GAN: Multi-Scale Gradients GAN (Architecture inspired from ProGAN but doesn't use layer-wise growing)
Branch: master
Clone or download
Latest commit 7cf7c8e Mar 25, 2019
Type Name Latest commit message Commit time
Failed to load latest commit information.
data removed merge-conflicts Dec 16, 2018
diagrams Add files via upload Sep 10, 2018
literature coding complete Aug 21, 2018
.gitignore code setup Jul 2, 2018
Gradients_explanation.jpg experiment 2 at (128 x 128) Celeba Aug 26, 2018
LICENSE.txt code setup Jul 2, 2018 Update Mar 25, 2019
architecture.jpeg Add files via upload Sep 10, 2018
celebA_samples.png Add files via upload Mar 25, 2019
detailed_architecture.jpg details of architecture added Aug 22, 2018

**Please note that this is not the repo for the MSG-GAN research paper. The code for the paper titled BMSG-GAN will be released soon.


MSG-GAN (Multi-Scale Gradients GAN): A Network architecture inspired from the ProGAN.

The architecture of this gan contains connections between the intermediate layers of the singular Generator and the Discriminator. The network is not trained by progressively growing the layers. All the layers get trained at the same time.

Implementation uses the PyTorch framework.

Celeba samples

celeba generated samples

Please note that all the samples at various scales are generated by the network simultaneously.

Multi-Scale Gradients architecture

proposed MSG-GAN architecture

The above figure describes the architecture of the proposed Multi-Scale gradients GAN. As you can notice, from every intermediate layer of the Generator, a particular resolution image is extracted through (1 x 1) convolutions. These extracted images are in turn fed to the appropriate layers of the Discriminator. This allows for gradients to flow from the Discriminator to the Generator at multiple scales.

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

The problem of occurence of random gradients for GANs at the higher resolutions is tackled by layerwise training in the ProGAN paper. I present another solution for it. I have run the following experiment that preliminarily validates the proposed approach.

gradients explanation

Above figure explains how the Meaningful Gradients penetrate the Generator from Bottoms-up. Initially, only the lower resolution gradients are menaingful and thus start generating good images at those resolutions, but eventually, all the scales synchronize and start producing images. This results in a stabler training for the higher resolution.

Celeba Experiment

I ran the experiment on a skimmed version of the architecture as described in the ProGAN paper. Following table summarize the details of the Networks:


For extracting images after every 3 layer block at that resolution, I used 1 x 1 convolutions. Similar operation is performed for feeding the images to discriminator intermediate layers.

The architecture for the discriminator is also the same (reverse mirror), with the distinction that half of the channels come from the (1 x 1 convolution) transformed downsampled real images and half from conventional top-to-bottom path.

All the 3 x 3 convolution weights have a forward hook that applies spectral normalization on them. Apart from that, in the discriminator for the 4 x 4 layer, there is a MinibatchStd layer for improving sample diversity. No other stablization techniques are applied.

64 x 64 experiment

Loss Plot

128 x 128 experiment

Loss Plot

The above diagrams are the loss plots obtained during training the Networks in an adversarial manner. The loss function used is Relativistic Hinge-GAN. Apart from some initial aberrations, the training has stayed smooth.

Running the Code

Please note to use value of learning_rate=0.0003 for both G and D for all experiments. TTUR doesn't work with this architecture (from experience). And, you can find other better learning rates, but the value 0.0003 always seems to work.

Running the training is actually very simple. Just start the training by running the 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
 --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
 --loss_function LOSS_FUNCTION
                    loss function to be used: 'hinge', 'relativistic-
 --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
 --gen_dilation GEN_DILATION
                    amount of dilation for the generator
 --dis_dilation DIS_DILATION
                    amount of dilation for the discriminator
 --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_spectral_norm USE_SPECTRAL_NORM
                    Whether to use spectral normalization or not
 --data_percentage DATA_PERCENTAGE
                    percentage of data to use
 --num_workers NUM_WORKERS
                    number of parallel workers for reading files

Running 1024 x 1024 architecture

For training a network as per the ProGAN CelebaHQ experiment, use the following arguments:

$ python --depth=9 \
                  --latent_size=512 \
                  --images_dir=<path to CelebaHQ images> \
                  --sample_dir=samples/CelebaHQ_experiment \

Set the batch_size, feedback_factor and checkpoint_factor accordingly. This experiment was carried out by me on a DGX-1 machine. The samples displayed in Figure 1. of this readme are the output of this experiment. You can use the models pretrained for 3 epochs at [1024 x 1024] for your training. These are available at ->

Trained weights for generating cool faces :)

Please refer to the models/Celeba/1/GAN_GEN_3.pth for the saved weights for this model in PyTorch format.

Other links

medium blog ->
Training video ->


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

Best regards,
@akanimax :)

You can’t perform that action at this time.