super_resolution
├── DIV2K
│ ├── DIV2K_train_HR
│ ├── DIV2K_train_LR_bicubic
│ │ ├── X2
│ │ ├── X3
│ │ ├── X4
│ ├── DIV2K_valid_HR
│ ├── DIV2K_valid_LR_bicubic
│ │ ├── X2
│ │ ├── X3
│ │ └── X4
│ └──
├── benchmark
│ ├── Set5
│ ├── Set14
│ ├── B100
│ ├── Urban100
│ └──
└──
-
Prepare image super-resolution dataset.
i. Download DIV2K training and validation images. Please download the low-resolution images in the NTIRE-2017 challenge.
ii. Download Set5, Set14, B100, and Urban100 benchmark from GoogleDrive or BaiduPan (extraction code: e5ae).
iii. Put the images in a folder called
super_resolution. So the folder structure should be like the one above. -
cd ./scriptsand run the following scripts to train the models.# EDSR-8-128: Basis-128-27 CUDA_VISIBLE_DEVICES=1 python ../main.py --model EDSR_Basis --save EDSR_Basis --scale 4 --patch_size 192 --batch_size 16 --basis_size 128 \ --n_basis 27 --n_resblocks 8 --n_feats 128 --res_scale 1 \ --epochs 300 --lr_decay 200 --lr 0.0001 --n_threads 8 --n_train 32208 --data_train DIV2KSUB --data_test Set5 --save_results # SRResNet: Basis-32-32 CUDA_VISIBLE_DEVICES=1 python ../main.py --model SRResNet_Basis --save SRResNet_Basis --scale 4 --patch_size 96 --batch_size 16 \ --basis_size 32 --n_basis 32 --n_resblocks 16 --n_feats 64 --bn_every \ --epochs 300 --lr_decay 200 --lr 0.0001 --n_threads 8 --n_train 32208 --data_train DIV2KSUB --data_test Set5 --save_results
For more information, please refer to the train scripts in
./scripts(./scripts)
-
Download the test model from GoogleDrive or BaiduPan (Extraction Code: vvfr).
-
Be sure to change the following directories in
./option.py.--dir_data: the directory where you put the dataset.--dir_save: the directory where you want to save your the results. -
Go to
./scripts/demo.shand changeDEVICESto your available CUDA devices andMODEL_PATHto the directory where you put the test model. -
Run the demo commands in
./scripts/demo.sh.DEVICES=1 MODEL_PATH="/home/thor/Downloads/model_super_resolution" ################################################################# # Table 2 ################################################################# # EDSR-8-128: Baseline CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR --save EDSR_X4_L8 --scale 4 --n_resblocks 8 --n_feats 128 --res_scale 1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_X4_L8.pt" --chop # EDSR-8-128: Basis-128-27 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR_Basis --save EDSR_Basis_X4_L8_B128+27 --scale 4 --basis_size 128 --n_basis 27 --n_resblocks 8 --n_feats 128 --res_scale 1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_Basis_X4_L8_B128+27.pt" --chop # EDSR-8-128: Basis-128-40 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR_Basis --save EDSR_Basis_X4_L8_B128+40 --scale 4 --basis_size 128 --n_basis 40 --n_resblocks 8 --n_feats 128 --res_scale 1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_Basis_X4_L8_B128+40.pt" --chop # EDSR-8-128: Factor-SIC2 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR_Factor --save EDSR_Factor_X4_L8_SIC2 --scale 4 --sic_layer 2 --n_resblocks 8 --n_feats 128 --res_scale 1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_Factor_X4_L8_SIC2.pt" --chop # EDSR-8-128: Factor-SIC3 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR_Factor --save EDSR_Factor_X4_L8_SIC3 --scale 4 --sic_layer 3 --n_resblocks 8 --n_feats 128 --res_scale 1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_Factor_X4_L8_SIC3.pt" --chop # SRResNet: Baseline CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model SRResNet --save SRResNet_X4_L16 --scale 4 --n_resblocks 16 --n_feats 64 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/SRResNet_X4_L16.pt" --chop # SRResNet: Basis-64-14 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model SRResNet_Basis --save SRResNet_Basis_X4_L16_B64+14 --scale 4 --basis_size 64 --n_basis 14 --bn_every --n_resblocks 16 --n_feats 64 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/SRResNet_Basis_X4_L16_B64+14.pt" --chop # SRResNet: Basis-32-32 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model SRResNet_Basis --save SRResNet_Basis_X4_L16_B32+32 --scale 4 --basis_size 32 --n_basis 32 --bn_every --n_resblocks 16 --n_feats 64 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/SRResNet_Basis_X4_L16_B32+32.pt" --chop # SRResNet: Factor-SIC2 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model SRResNet_Factor --save SRResNet_Factor_X4_L16_SIC2 --scale 4 --sic_layer 2 --n_resblocks 16 --n_feats 64 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/SRResNet_Factor_X4_L16_SIC2.pt" --chop # SRResNet: Factor-SIC3 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model SRResNet_Factor --save SRResNet_Factor_X4_L16_SIC3 --scale 4 --sic_layer 3 --n_resblocks 16 --n_feats 64 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/SRResNet_Factor_X4_L16_SIC3.pt" --chop ################################################################# # Table 3 ################################################################# # EDSR: Baseline X2 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR --save EDSR_X2_32 --scale 2 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_X2_L32.pt" --chop # EDSR: Baseline X3 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR --save EDSR_X3_32 --scale 3 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_X3_L32.pt" --chop # EDSR: Baseline X4 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR --save EDSR_X4_32 --scale 4 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_X4_L32.pt" --chop # EDSR: Basis X2 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR_Basis --save EDSR_Basis_X2_L32 --scale 2 --basis_size 256 --n_basis 32 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_Basis_X2_L32.pt" --chop # EDSR: Basis X3 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR_Basis --save EDSR_Basis_X3_L32 --scale 3 --basis_size 256 --n_basis 32 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_Basis_X3_L32.pt" --chop # EDSR: Basis X4 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR_Basis --save EDSR_Basis_X4_L32 --scale 4 --basis_size 256 --n_basis 32 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_Basis_X4_L32.pt" --chop # EDSR: Basis-S X2 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR_Basis --save EDSR_Basis_X2_L32U --scale 2 --share_basis --basis_size 256 --n_basis 32 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_Basis_X2_L32U.pt" --chop # EDSR: Basis-S X3 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR_Basis --save EDSR_Basis_X3_L32U --scale 3 --share_basis --basis_size 256 --n_basis 32 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_Basis_X3_L32U.pt" --chop # EDSR: Basis-S X4 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR_Basis --save EDSR_Basis_X4_L32U --scale 4 --share_basis --basis_size 256 --n_basis 32 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_Basis_X4_L32U.pt" --chop # EDSR: Factor X2 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR_Factor --save EDSR_Factor_X2_L32 --scale 2 --sic_layer 1 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_Factor_X2_L32.pt" --chop # EDSR: Factor X3 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR_Factor --save EDSR_Factor_X3_L32 --scale 3 --sic_layer 1 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_Factor_X3_L32.pt" --chop # EDSR: Factor X4 CUDA_VISIBLE_DEVICES=${DEVICES} python ../main.py --model EDSR_Factor --save EDSR_Factor_X4_L32 --scale 4 --sic_layer 1 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --data_test Set5 --save_results --n_GPUs 1 --test_only --pre_train "${MODEL_PATH}/EDSR_Factor_X4_L32.pt" --chop
If you find our work useful in your research of publication, please cite our work:
@inproceedings{li2019learning,
title = {Learning Filter Basis for Convolutional Neural Network Compression},
author = {Li, Yawei and Gu, Shuhang and Van Gool, Luc and Timofte, Radu},
booktitle = {Proceedings of the IEEE International Conference on Computer Vision},
year = {2019}
}