search

OpenGVLab / UniFormerV2

[ICCV2023] UniFormerV2: Spatiotemporal Learning by Arming Image ViTs with Video UniFormer
https://arxiv.org/abs/2211.09552
Apache License 2.0
294 stars 19 forks source link

Requesting for training logs #21

Closed vateye closed 1 year ago

vateye commented 1 year ago

Hi, is it convenient for you to share the training logs for all of the experiments? I would be appericiated if you share those precious training logs as the reference for reproducing the results. Thanks!

Andy1621 commented 1 year ago

Since the model code is recleaned before publishing, the output is different from the current codebase's, which may cause misunderstanding. But the loss and accuracy are the same. However, the logs are too many to be released. I can provide part of them. Could you please tell me the experiments you are conducting? Then I can find the corresponding log and upload them to the Google Drive.

vateye commented 1 year ago

Currently, I am reproducing the results for k400_b16_8frames, k710_b16_8frames, and ssv2_b16_16frames. Would you mind sharing those logs as the reference for my reproducing? Thanks.

Andy1621 commented 1 year ago

Thanks for your question, I have added the corresponding log here.

I tried to add EMA in the original experiments, but it only works sometimes. So I remove EMA in the later experiments. For the logs of k400_b16_f8 and ssv2_b16_f16, they show the EMA testing, and the first one is testing without EMA. For the logs of k710_b16_f8, I have tested the model on the k400. The first one is k710, and the second one is k400. Note that without extra fine-tuning on the k400, the top-1 accuracy is similar to training on only the k400.

vateye commented 1 year ago

Thanks!