Hello,if I want to do compression blur, which model should I use? And how should I start training?

[linjing7]

Hi, thanks for your interest. You need to download the pretrained model and test your own image using the following command: python demo/restoration_video_demo.py configs/FGST_deblur_gopro_test.py ${CHKPT} ${IN_PATH} ${OUT_PATH} This model is trained on the motion blur dataset and if you want to do compression deblur, you may need to collect your own dataset as GOPRO and retrain your model by: bash tools/dist_train.sh configs/FGST_deblur_xxx.py GPU_Nums

[tlovem]

Thank you for your answer, but when I encounter some problems during training, “RuntimeError: cuDNN error: CUDN STATUS NOT_ INITIALIZED”, is it because my version doesn't correspond?

[tlovem]

I followed the steps you provided for training, except that the cuda version is 10.2

[linjing7]

Hi, this may be caused by the mismatch between Pytorch and CUDA versions. You need to specify the CUDA version when you install Pytorch.

[tlovem]

Thanks for your answer, I'll try again, looking forward to your S2SVR

[tlovem]

I reinstalled the pytorch using Conda, which solved the problem just now, but my graphics card doesn't have enough computational power. Can I reduce some parameters to make it work?

[linjing7]

Hi, you can try:

1. change the network parameters of FGST, e.g., dim, dim_emb, num_resblocks
2. change the training setting, e.g., reduce the video length by setting num_input_frames to a smaller number, reduce the training iters by setting total_iters in the config file.

[tlovem]

Thank you for your patient answer. I'll try it