Restormer

A PyTorch implementation of Restormer based on CVPR 2022 paper Restormer: Efficient Transformer for High-Resolution Image Restoration.

Requirements

• Anaconda

• PyTorch

conda install pytorch=1.10.2 torchvision cudatoolkit -c pytorch

Dataset

Rain100L and Rain100H are used, download these datasets and make sure the directory like this:

|-- data     
    |-- rain100L
        |-- train
            |-- rain
                norain-1.png
                ...
            `-- norain
                norain-1.png
                ...
        `-- test                                                        
    |-- rain100H
        same as rain100L

Usage

You can easily train and test the model by running the script below. If you want to try other options, please refer to utils.py.

Train Model

python main.py --data_name rain100L --seed 0

Test Model

python main.py --data_name rain100H --model_file result/rain100H.pth

Benchmarks

The models are trained on one NVIDIA RTX A6000 GPU (48G). num_iter is 30,000, seed is 1 and milestone is [9200, 15600, 20400, 24000, 27600], the other hyper-parameters are the default values.

Due to the huge demand for GPU memory, we have to reduce the batch_size and patch_size:

Ours: batch_size is [64, 40, 32, 16, 8, 8] and patch_size is [32, 40, 48, 64, 80, 96];

Ours*: batch_size is [32, 20, 16, 8, 4, 4] and patch_size is [64, 80, 96, 128, 160, 192].

Results

More results could be downloaded from MEGA. Here we give some examples for Ours*.