Validation Loss did not decrease in the HMDB51 notebook?

[neonampv]

I trained your HMDB51 notebook in 10 epochs but the validation loss did not decrease? Why did it happened?

[Atze00]

Can you provide more information? Maybe a log of the training.
It's probably due to the fact that it converges from the first epoch

[nguyenquibk1996]

I had the same problem.
When I trained Movinet with my own dataset. The training loss decreased but the validation loss increased from the first epoch.
( see the below image )
If I trained Movinet not using pretrained Kinetics with HMDB51 in the notebook sample and my own dataset (i did not save a log of the training), both losses had not decreased.
Can you explain it? Thank you

[Atze00]

The notebook functions correctly, also I use the networks daily. It's unlikely these are problem related to part of the code of this repository. If you provide a colab short script that reproduces the problem I will look at it.

[nguyenquibk1996]

I will run your notebook with HMDB51 for 10 epochs and show to you a log of the training. Because I used the same function for my own dataset and got the same problem. I don't think it can converges from the first epoch with many datasets. When you train Movinet with your dataset, the validation loss decreases or not?
Thank you

[poincarelee]

I trained HMDB51 dataset for 20 epochs with modelA0_stream_statedict_v3, the result is as follows:

[poincarelee]

@nguyenquibk1996
Hi， did you solve the problem? I met the same problem in my own dataset. Then I tried to train hmdb51 without pretrained, the evaluation accuracy is as follows:

Did I miss any key points during finetuning or could you give any clues about this?

[haowei2020]

@nguyenquibk1996 Hi， did you solve the problem? I met the same problem in my own dataset. Then I tried to train hmdb51 without pretrained, the evaluation accuracy is as follows:

Did I miss any key points during finetuning or could you give any clues about this?

I think the dataset is primary cause, and the data processing method(one clip or multiple clips sampled from one video) is second cause. I have train the X3D and SlowFast on the HMDB51 by mmaction2(default config samples one clip from one video), top1 acc is also about 30%, and validation loss can decrease.

[poincarelee]

Have you tried any other datasets? BTW, I see in mmaction2 there's no training codes of X3D, could you tell me how to get the codes?

[haowei2020]

Have you tried any other datasets? BTW, I see in mmaction2 there's no training codes of X3D, could you tell me how to get the codes?

I have tried UCF101, the top1 can over 45%. I write the codes of X3D by myself(of course refer to existing codes )