pilhyeon / basnet-pytorch Goto Github PK

any colab version?

As the features extraction process for activitynet is very time-consuming, is it possible for you to provide the activitynet features you used for this project?

The reported performance cannot be reproduced unless I use the pre-trained model. If I train the model from scratch, the performance never reaches the reported ones.

Can you share random seeds/exact setting that you chose to produce the reported performance?

I find the test mAP before 150 is all zero.

I tweaked CAS module a little bit, and increased the learning rate. However after some iterations an "IndexError: list index out of range" occurred in test.py. I tried to locate the error, and it turned out that it's caused by this line:

Actually
assert proposals[i][0] is not None
failed.

Can you please help me solve this problem? Thanks in advance!

Dear friend

Which github repo do you use to extract the img and flow for a video?

Do you use tsn github repo to extract the img and flow as follow:
https://github.com/yjxiong/temporal-segment-networks.git

Thanks!

Do you have any plans to release the relevant code and features of ActivityNet 1.3 and 1.2?

Hi.
Why minmax_norm actually works scores / (max_scores per each class) (here is code)?

I found that there is a case that max_score is negative which means that sign is flipped due to minmax_norm.

Thank you!

Dear author:

What is your best random seed in the THUMOS 14 and ActivityNet?

Thanks!

you use following code to get predict probability
score_base = self.softmax(score_base)
score_supp = self.softmax(score_supp)

but your loss function is choose BCE loss.

I just cannot understand, this is a multiple classification problems, why you use softmax to get probability? mabey you should use sigmoid function, because every class may be occur.

Dear friend:

Could you share the visualization code of showing the Base and Suppression score curve as your paper?

Thanks a lot.

Dear author

Why the scale in config is 24? What is the meaning of scale?

Thanks a lot!

After train the BaSNet, I only get 26.3 map@tiou=0.5 after train without changing any code. Could you give me any hit to reproduce the result?

Maybe the seed? I don't know.

The following are eval curve and the final test result.