Awesome

LMLT(Low-to-high Multi-Level Vision Transformer)

Jeongsoo Kim, Jongho Nang, Junsuk Choe^*

^* : Corresponding author

Requirements

# Install Packages
pip install -r requirements.txt
pip install matplotlib

# Install BasicSR
python3 setup.py develop

Dataset

We used DIV2K, Flickr2K as Training dataset. You can download two datasets at https://github.com/dslisleedh/Download_df2k/blob/main/download_df2k.sh and prepare other test datasets at https://github.com/XPixelGroup/BasicSR/blob/master/docs/DatasetPreparation.md#Common-Image-SR-Datasets

And also, you'd better extract subimages using

python3 scripts/data_preparation/extract_subimages.py

By running the code above, you may get subimages of training datasets.

Training

You can train LMLT following commands below

python3 basicsr/train.py -opt options/train/LMLT/train_tiny(base, large)_DF2K_X2(3, 4).yml

And also, we set torch.backends.cudnn.benchmark to True to accelerate training process so that results can be fluctuated a little. If you want to get fixed output, you should set it to False and set torch.backends.cudnn.deterministic to True.

Test

You can test LMLT following commands below

python3 basicsr/test.py -opt options/test/LMLT/test_tiny(base, large)_benchmark_X2(3, 4).yml

Result

Result table with #Param and #FLOPs

Result table with GPU Consumption and AVG Time

Results

We will provide visual results of LMLT_Base x4 scale soon. If you want to see only architecture, please refer to LMLT.py.