Huang Ziyuan

> It seems that deleting the frame manually avoids the leak: > > ```python > vr = VideoReader(video_path) > for frame in vr: > print(frame.shape) > del frame > ```...

Hi, sorry for the delayed reply. Ideally, TAdaConv could be applied to anywhere that uses a convolutional backbone, as long as the training has multiple frames. We have applied TAdaConv...

Hi, thanks for the interest. I will look into this problem ASAP.

> 嗨，感谢您的关注。我会尽快调查这个问题. > Thank you for your attention. Later, I tried to use your fine-tuned checkpoint to load val files on hmdb51 using the R (2+1) D backbone for verification...

It is important to note that our model is trained on the first split, so it is only valid when it is evaluated on the test set of the first...

This is caused by the change of cropping function that we use. Set the cfg.DATA.TEST_SCALE to 112 and the problem would be solved.

I do not see a problem in your current config. Autoaugment only affects the training process. We have run on our end that we could achieve the exact 51.83 using...

On our side, we have reproduced a result similar to yours using decord==0.6.0. We used decord==0.4.0 for producing 51.83 on HMDB51.

Could you try using the code of [this version](https://github.com/alibaba-mmai-research/TAdaConv/commit/fa78f82bebd8b55215eaf516ac4b0aa2c9620d37 ) to reproduce the result and see whether it is normal?

Hi, sorry for the delayed reply. You can check out our latest work to see if it helps you :). The link is here: https://arxiv.org/pdf/2308.05787.pdf. We have released the code...