DLSR-PyTorch

April 2, 2022 · View on GitHub

This repository is an implementation of the paper "Lightweight Image Super-Resolution with Hierarchical and Differentiable Neural Search".https://arxiv.org/pdf/2105.03939.pdf

The code is based on:

EDSR-PyTorch: https://github.com/sanghyun-son/EDSR-PyTorch
RFDN: https://github.com/njulj/RFDN
PAN: https://github.com/zhaohengyuan1/PAN

Dependencies

Python 3.7
PyTorch 1.2
numpy
skimage
imageio
matplotlib
logging

Search

The search script can be found at : script/RFDN_beta_NAS.sh

Usage :

NAME= **YOUR TASK NAME**
GPUS=0
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH: **YOUR ABSOLUTE PATH TO**/Anaconda3/lib/
**YOUR ABSOLUTE PATH TO**/python **YOUR ABSOLUTE PATH TO**/NAS_SR/search/RFDN6_para_loss.py --name $NAME \
      --save **YOUR ABSOLUTE PATH TO**/NAS_SR/checkpoint/ \
      --batch_size 64 \
      --gpu_ids $GPUS \
      --patch_size 64 \
      --seed 9 \
      --dir_data **YOUR ABSOLUTE PATH TO**/data/ \  
      --data_train DF2K \
      --data_test Set5 \
      --data_range 1-3000/3000-3450/3000-3450 \
      --rgb_range 255 \
      --ext sep > **YOUR ABSOLUTE PATH TO**/NAS_SR/training_logs/**file name**.log

retrain

Copy the genotype to models/genotypes_rfdn.py in the format listed in the file. Or you may directly use the genotypes in the file.

For network-level structure, you may slightly change the model in models/beta1_para_loss.py

The retrain script can be found at : script/train_beta_para.sh. The usage is basically the same as search, note to input the argument of genotypes

The model with the best performance is saved in checkpoint/**YOUR TASK NAME** during training.

The test script can be found at : script/test_beta_para.sh. The benchmark datasets can be downloaded in EDSR-PyTorch. The retrained models can be found in checkpoint/best. All files with 'tiny' in their name means the tiny model we build to compare with TPSR, ECCV2020.

We run the search and retrain process in single V100 GPU.