Ji Lin comments

Results 63 comments of


                                            Ji Lin

Pretrained model for sth-sth v1 and kinetics got lower accuracy

Hi, I think the possible reason could be data processing. How did you process your data?

Question about finetune on UCF101

Hi, firstly, UCF-101 does provide the official split for training and testing (but there are several splits to prevent overfitting). Could you please try the experiments on split 1 to...

Question about finetune on UCF101

@Jaychoukobbe No. You should split the dataset into train and eval according to split1.

Question about finetune on UCF101

Hi, for UCF-101, people usually perform experiments on the 3 splits and report the average accuracy to reduce over-fitting. You just need to convert the format of the 3 pairs...

how to process the dataset ?

You can reference https://github.com/metalbubble/TRN-pytorch/blob/master/test_video.py to test on your own video.

strided_slice get empty slice at axis 1

Hi, please git pull to get the recent update to the code and have a try!

[Bugs] lr_mult, decay_mult unavailable in optimizer

Hi, I agree that the code piece from TSN is a little hack to support different learning rate, but it should work fine with PyTorch. Do you find that it...

[Bugs] lr_mult, decay_mult unavailable in optimizer

The related code is here: https://github.com/mit-han-lab/temporal-shift-module/blob/master/main.py#L362

[Bugs] lr_mult, decay_mult unavailable in optimizer

Yes, I will refactor this part to make it more clear when time permits.

The accuary of TSM-NL is only 57.85%

Hi, could you please check if you can get a similar accuracy for other models without NL modules?