[toc]
Team "Diggers" winner solution to Mobile AI 2021 Real-Time Video Super-Resolution Challenge
official report paper: https://arxiv.org/abs/2105.08826
pip install megengine -f https://megengine.org.cn/whl/mge.html
pip install -r requirements.txt
configs/restorers/BasicVSR/mai.py
cd xxx/MAI-VSR-Diggers
python tools/train.py configs/restorers/BasicVSR/mai.py --gpuids 0,1,2,3 -d
support multi gpus training, change to yours, e.g. --gpuids 0 --gpuids 0,2 etc...
you can find output information and checkpoints in .workdirs/...
use our trained model (generator_module.mge), already inside this repo: ./ckpt/epoch_62
which is only 92kb
it has been trained 62 epochs on 240 clips, it's PSNR on validation dataset(3000 frames) is 27.98
find the config file: configs/restorers/BasicVSR/mai_test_valid.py
change first lines for your situation, actually you only need to fix the dataroot
dataroot = "/path2yours/REDS/train/train_sharp_bicubic"
load_path = './ckpt/epoch_62'
exp_name = 'mai_test_for_validation'
eval_part = tuple(map(str, range(240, 270)))
and then , run it:
cd xxx/MAI-VSR-Diggers
python tools/test.py configs/restorers/BasicVSR/mai_test_valid.py --gpuids 0 -d
you can find the results in ./workdirs/...
find the config file: configs/restorers/BasicVSR/mai_test_test.py
change first lines for your situation, actually you only need to fix the dataroot
dataroot = "/path2yours/REDS/test/test_sharp_bicubic"
load_path = './ckpt/epoch_62'
exp_name = 'mai_test_for_test'
eval_part = None
and then , run it:
python tools/test.py configs/restorers/BasicVSR/mai_test_test.py --gpuids 0 -d
you can find the results in ./workdirs/...
notice: only support one gpu config for gpuids now
https://drive.google.com/file/d/1R0DDHmV8jZW_iYJQZksO2RWkZTrAYYPi/view?usp=sharing
cd xxx/tflite/
python main.py --mgepath /xxxxxx/ckpt/epoch_62/generator_module.mge
cd xxx/tflite/
python main.py --mgepath /xxxxxx/ckpt/epoch_62/generator_module.mge -n
notice that to use absolute path
you will get tflite files in the dir xxx/tflite/xxx
and we have supported our pre-built model.tflite
and model_none.tflite
in ckpt
dir
you can refer to https://github.com/Feynman1999/MAI-VSR-Diggers/issues/2.