Neural Architecture Search for Super Resolution UNet architecture

Based on article https://arxiv.org/pdf/1909.11409.pdf

Scripts description

• analysis.py - script used to create plots (distribution per generation) from log file
• blocks.py - module contain block class with the help of which the network is built (contain parameters responsible for all possible structures of block)
• evaluation.py - module responsible for evaluation of trained models
• evolution.py - main script runs the evolutionary training algorithm distributed between several GPUs
• loss_metric.py - module contain MSE loss and PSNR metric
• sr_data.py - module responsible for creating Super Resolution datasets from given folders
• train.py - training script which forming learning per each GPU, runs by subprocess module from evolution script
• UNet.py - module which create UNet model from given list of blocks

Example of usage

foo@bar:~/NAS_UNetSR$ python evolution.py --gpus 4

Example of obtained distribution

TODO

• Do weights sharing between models training
• Try different block structure (Leaky ReLU, ...)