The loss cannot be reduced

Question

The loss cannot be reduced

Closed this issue 3 months ago · 1 comments

For the Multi-THUMOS dataset, I can obtain the expected results using the provided weights for testing. However, when training from scratch, after the third epoch, the loss fluctuates around 1.1 and does not decrease further. There have been no changes to the code or commands. Could you provide some possible reasons for this issue?

Answer 1 · 2024-09-14T18:24:48.000Z

Please make sure the following are correctly set.

THUMOS raw video data and Multi-THUMOS annotations: please use the provided data.
Pretrained VideoMAE checkpoint: AdaTAD is initialized with K400-pretrained VideoMAE; please make sure you download them and put them under the folder ./pretrained/ correctly, as illustrated here.

If you have any further questions, please email me.