search

Slinene / SDD-FIQA

This repo is the official released code of SDD-FIQA (CVPR-2021)
33 stars 2 forks source link

Very time consuming #2

Open AaboutL opened 3 years ago

AaboutL commented 3 years ago

Hi, I reimplement the training code myself, but it seems that generating label is very time consuming. I am curious about how long does it take to train a model.

Slinene commented 3 years ago

It is a good question, we randomly sample the negative pairs to accelerate the generation of pseudo labels. The time consuming can be reduced from O(n^2) to O(n).

AaboutL commented 3 years ago

I have noticed the time complexity in your paper. What I am really care about is the real time you spend training the model from scratch, one week or two weeks. Can you give me a definite answer?

Slinene commented 3 years ago

For MSIM, we set m= 24 and K= 12，the generation of quality score takes about 3 days. The training of quality prediction model costs 1 day with P40*8.