search

xinntao / EDVR

Winning Solution in NTIRE19 Challenges on Video Restoration and Enhancement (CVPR19 Workshops) - Video Restoration with Enhanced Deformable Convolutional Networks. EDVR has been merged into BasicSR and this repo is a mirror of BasicSR.
https://github.com/xinntao/BasicSR
1.49k stars 318 forks source link

Aboat training #50

Open hx1024 opened 5 years ago

hx1024 commented 5 years ago

In the training .yml file，what does the "ft_tsa_only" mean? when I doesn't use it, prompted WARNING: Offset mean is XX, larger than 100. and the loss is larger than use the "ft_tsa_only" to train

xinntao commented 5 years ago

ft_tsa_only: only finetune TSA module We first train a model w/o TSA. Then we add the TSA module and only finetune TSA for some iterations. After that, we finetune the whole model.

TouqeerAhmad commented 5 years ago

I wanted to make sure I understood is correctly. Is this how the complete training goes:

  1. Train the model with only PCD and reconstruction modules i.e. woTSA,
  2. Train only TSA module.
  3. Train the full EDVR with everything in place.

If that is the case; can you please answer the following questions:

  1. For how long you train woTSA?
  2. How to train only TSA module? Is there a setting I can do in the yml file? And how long this should be trained/fine-tuned?
TouqeerAhmad commented 5 years ago

Hi, by looking at the provided log files; it feels you are first training for 600K iterations woTSA and then using this pre-trained model, further training for another 600K iterations with TSA module in place. Is this correct?

Why fine-tuning based on TSA required equal amount of training?

xinntao commented 5 years ago
  1. Yes, correct.
  2. I think we can shorten the training iteration without affecting the performance much. I set it the same as the first stage, for simplicity.