search

ryanxingql / stdf-pytorch

Implementation of "Spatio-Temporal Deformable Convolution for Compressed Video Quality Enhancement" (AAAI'20).
https://www.aiide.org/ojs/index.php/AAAI/article/view/6697
Apache License 2.0
155 stars 20 forks source link

Question about training #7

Closed linjing7 closed 3 years ago

linjing7 commented 3 years ago

Hi, I have met some problems during training.

  1. As metioned in the last issue, I have made YUV lmdb instead of Y lmdb. However, when I use it as training set, it takes much longer time than origin method, about 2 times. I think it may be caused by the cv2.imdecode() function. I fould that when making lmdb, we use the cv2.imencode() with compression level as 1, therefore, we need to decode while reading data from lmdb. I would like to skip the cv2.imencode() while making lmdb, that is, save the data into lmdb without compression, in this way I don't need to use cv2.imdecode() function during training, which may save some time. I wonder that if this will make influence on the final result?
  2. BTY, you have mentioned in the Q&A that you enlarge the dataset by 'set sampling index = target index % dataset len'. I don't understand it, could you please tell me where it is in the code?
ryanxingql commented 3 years ago
  1. I don't know. You can have a try and set the compression level being 0, which means no compression.
  2. https://github.com/RyanXingQL/STDF-PyTorch/blob/36a6adfe0c33f030a68155eb706326d4ee896027/utils/deep_learning.py#L86
linjing7 commented 3 years ago

Okay, thank you very much, I will have a try.