search

lidq92 / MDTVSFA

[official] Unified Quality Assessment of In-the-Wild Videos with Mixed Datasets Training (IJCV 2021)
MIT License
83 stars 16 forks source link

end-to-end fine-tuning #7

Closed ciwei123 closed 3 years ago

ciwei123 commented 3 years ago

@lidq92 Thanks for your reply. When I train the network end to end ,that means I will train the Resnet-50,overfifting will occur,SROCC is 0.1.But if I just only train the regression net, the SROCC is 0.6. What should I do?Thank you very much!!

lidq92 commented 3 years ago

@ciwei123 Please read the paper and follow the README. I did not get what you said both in the title and the comment.

lidq92 commented 3 years ago

@ciwei123 Please change your issue's title to the correct one. I tried to understand your question, and if my understanding is correct, then one suggestion for end-to-end fine-tuning is using a smaller learning rate for the pretrained ResNet-50's weights.

ciwei123 commented 3 years ago

@lidq92 Thanks for your reply. I will try the method you said. I am sorry for my carelessness about the title.Thank you very much.

lidq92 commented 3 years ago

@ciwei123 Usually 1/10*lr for the ResNet-50 backbone's weights. See lidq92/LinearityIQA and the corresponding paper's Figure 5(c).