Training SRGAN to Generate High-Resolution Images

Of course, we need to have some data to work with.

⚠️ The dataset is intended only for non-commercial research and educational use.

The image collection we will be using is the VOC2012 datasethttps://www.kaggle.com/datasets/gopalbhattrai/pascal-voc-2012-dataset. With a batch size of 64, it takes several hours to train for 100 epochs on a GTX 1080Ti graphics card. GPU memory usage increases as we increase batch size.

Create a folder named data and place the training images into a folder called DIV2K_train_HR and the valid images into DIV2K_valid_HR. Next, create a folder named epochs to hold the epoch data. Finally, create a folder named training_results.

To train SRGAN, execute the following command in a Terminal: