Home

Awesome

PyTorch-SRGAN

A modern PyTorch implementation of SRGAN

It is deeply based on Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network paper published by the Twitter team (https://arxiv.org/abs/1609.04802) but I replaced activations by Swish (https://arxiv.org/abs/1710.05941)

You can start training out-of-the-box with the CIFAR-10 or CIFAR-100 datasets, to emulate the paper results however, you will need to download and clean the ImageNet dataset yourself. Results and weights are provided for the ImageNet dataset.

Contributions are welcome!

Requirements

Training

usage: train [-h] [--dataset DATASET] [--dataroot DATAROOT]
             [--workers WORKERS] [--batchSize BATCHSIZE]
             [--imageSize IMAGESIZE] [--upSampling UPSAMPLING]
             [--nEpochs NEPOCHS] [--generatorLR GENERATORLR]
             [--discriminatorLR DISCRIMINATORLR] [--cuda] [--nGPU NGPU]
             [--generatorWeights GENERATORWEIGHTS]
             [--discriminatorWeights DISCRIMINATORWEIGHTS] [--out OUT]

Example: ./train --cuda

This will start a training session in the GPU. First it will pre-train the generator using MSE error for 2 epochs, then it will train the full GAN (generator + discriminator) for 100 epochs, using content (mse + vgg) and adversarial loss. Although weights are already provided in the repository, this script will also generate them in the checkpoints file.

Testing

usage: test [-h] [--dataset DATASET] [--dataroot DATAROOT] [--workers WORKERS]
            [--batchSize BATCHSIZE] [--imageSize IMAGESIZE]
            [--upSampling UPSAMPLING] [--cuda] [--nGPU NGPU]
            [--generatorWeights GENERATORWEIGHTS]
            [--discriminatorWeights DISCRIMINATORWEIGHTS]

Example: ./test --cuda

This will start a testing session in the GPU. It will display mean error values and save the generated images in the output directory, all three versions: low resolution, high resolution (original) and high resolution (generated).

Results

Training

The following results have been obtained with the current training setup:

Other training parameters are the default of train script

Tensorboard training graphs

Testing

Testing has been executed on 128 randomly selected ImageNet samples (disjoint from training set)

[7/8] Discriminator_Loss: 1.4123 Generator_Loss (Content/Advers/Total): 0.0901/0.6152/0.0908

Examples

See more under the output directory

High resolution / Low resolution / Recovered High Resolution

Original doggy <img src="https://raw.githubusercontent.com/ai-tor/PyTorchSRGAN/master/output/low_res/41.png" alt="Low res doggy" width="96" height="96"> Generated doggy

Original woman <img src="https://raw.githubusercontent.com/ai-tor/PyTorchSRGAN/master/output/low_res/38.png" alt="Low res woman" width="96" height="96"> Generated woman

Original hair <img src="https://raw.githubusercontent.com/ai-tor/PyTorchSRGAN/master/output/low_res/127.png" alt="Low res hair" width="96" height="96"> Generated hair

Original sand <img src="https://raw.githubusercontent.com/ai-tor/PyTorchSRGAN/master/output/low_res/72.png" alt="Low res sand" width="96" height="96"> Generated sand