- Linux or Windows
- Python 3
- CPU or NVIDIA GPU + CUDA CuDNN
We use Places365-Standard dataset at the resolution of 512*512 as the original non-distorted images to generate the 256*256 dataset.
data/distortion_model.py define the distortion model and the generation of distortion parameters and data/dataset_generation.py define how to generate distortion dataset.
train.py contains the main training function code, and some parameters and dataset loactions need to be specified.
trainDistill.py contains the main training function code of distilled model
trainPruned.py contains the main training function code of pruned model
resampling.py contains code for the resampling part relied upon during the inference process in infer.py.
predict.py is used for inference on distorted images after the network has been trained.
This directory contains the datasets used for training and testing the network. The data can be generated by running dataset/dataset_generator.py.