search

Sense-X / UniFormer

[ICLR2022] official implementation of UniFormer
Apache License 2.0
823 stars 111 forks source link

accuracy of k400 #57

Closed W4ngH4o closed 2 years ago

W4ngH4o commented 2 years ago

I have read the log file of 8x1x4. I find that the error rate of val dataset in the training process is 31.35 and the test accuracy rate of 8x1x1 is 74.41. As far as I am concerned, the val dataset and the test dataset are the same. So I want to know why the two results are different. Thanks for your reply.

Andy1621 commented 2 years ago

In fact, the error rate is equal to 1 - accuracy. Thus the accuracies are similar in both training and testing. As for the difference between training and testing, it is mainly because the random crop is used for validation. https://github.com/Sense-X/UniFormer/blob/1ce70bcccbe72962813aedf5eb1209f318f859b6/video_classification/slowfast/datasets/kinetics.py#L149

In my recent experiments, I have fixed the bug and the accuracies are the same. Just simply use the center crop for validation.

W4ngH4o commented 2 years ago

Thanks so much. I understood.

Andy1621 commented 2 years ago

As there is no more activity, I am closing the issue, don't hesitate to reopen it if necessary.