Implementing SRResnet/SRGAN Super Resolution with Tensorflow

The paper trained their networks by crops from the renowned ImageNet image recognition dataset. Although it is beneficial to train models in large amounts of data, the dataset found to be too heavy and I decided to use the tf_flowers dataset, consisting of 3,670 images which might seem too small but were just enough for a toy dataset to evaluate and compare the performance of each training method of the paper. We use the tensorflow_datasets module for loading the tf_flowers dataset and take the first 600 images as a validation dataset. We then define a function to map each image from the dataset to (128, 128) crops and a (32, 32) low-resolution copy of it. We can apply this function to our dataset by train_data.map(build_data,