RecurrentMobileNet

RecurrentMobileNet copied to clipboard

Recurrent MobileNet for MegCup 2022

We (IIL2) propose Recurrent MobileNet for light parameters raw image denoising. Our code is based on Pytorch, we also convert model weights to MegEngine for inference.

Environment

conda env create -f requirements.yaml
conda activate recurrent_mobilenet

Check Params

Number of parameters of the model is restrict to 100k. To check the number of parameters, you could run each file in models directory.

cd models
python recurrent_mobilenet.py

Testing

cd test_models
python test.py --path path_to_dataset

it should generate a result.bin to current directory. The path_to_dataset is organized the same as downloaded:

path_to_dataset
|--burst_raw
   |--competition_train_input.0.2.bin
   |--competition_train_gt.0.2.bin
   |--competition_test_input.0.2.bin

(Validating)

We also divide the last 1024 pairs in the training set as our validation set, you can also validate on our dataset.

cd validate_models
python validate.py --path path_to_dataset

Training

cd train_models
python distributed_train.py --path path_to_dataset --worldsize 8