AdaDM

AdaDM: Enabling Normalization for Image Super-Resolution.

You can apply BN, LN or GN in SR networks with our AdaDM. Pretrained models (EDSR*/RDN*/NLSN*) can be downloaded from Google Drive or BaiduYun. The password for BaiduYun is kymj.

📢 If you use BasicSR framework, you need to turn off the Exponential Moving Average (EMA) option when applying BN in the generator network (e.g., RRDBNet). You can disable EMA by setting ema_decay=0 in corresponding .yml configuration file.

Preparation

Please refer to EDSR for instructions on dataset download and software installation, then clone our repository as follows:

git clone https://github.com/njulj/AdaDM.git

Training

cd AdaDM/src
bash train.sh

Example training command in train.sh looks like:

CUDA_VISIBLE_DEVICES=$GPU_ID python3 main.py --template EDSR_paper --scale 2\
        --n_GPUs 1 --batch_size 16 --patch_size 96 --rgb_range 255 --res_scale 0.1\
        --save EDSR_AdaDM_Test_DIV2K_X2 --dir_data ../dataset --data_test Urban100\
        --epochs 1000 --decay 200-400-600-800 --lr 1e-4 --save_models --save_results 

Here, $GPU_ID specifies the GPU id used for training. EDSR_AdaDM_Test_DIV2K_X2 is the directory where all files are saved during training. --dir_data specifies the root directory for all datasets, you should place the DIV2K and benchmark (e.g., Urban100) datasets under this directory.

Testing

cd AdaDM/src
bash test.sh

Example testing command in test.sh looks like:

CUDA_VISIBLE_DEVICES=$GPU_ID python3 main.py --template EDSR_paper --scale $SCALE\
        --pre_train ../experiment/test/model/EDSR_AdaDM_DIV2K_X$SCALE.pt\
        --dir_data ../dataset --n_GPUs 1 --test_only --data_test $TEST_DATASET

Here, $GPU_ID specifies the GPU id used for testing. $SCALE indicates the upscaling factor (e.g., 2, 3, 4). --pre_train specifies the path of saved checkpoints. $TEST_DATASET indicates the dataset to be tested.

Acknowledgement

This repository is built on EDSR and NLSN. We thank the authors for sharing their codes.