search

ColumbiaDVMM / CDC

CDC: Convolutional-De-Convolutional Networks for Precise Temporal Action Localization in Untrimmed Videos
68 stars 18 forks source link

extract_image_feature generates "nan" probability #13

Closed hytseng0509 closed 6 years ago

hytseng0509 commented 7 years ago

Hi,

Thanks for sharing the codes. I followed the steps of reproducing results on THUMOS14. However, I found that some probabilities in .prob file generated by executing ./xfeat.sh were "nan". (not all probabilities in a single .prob file is "nan", only a few of them) Any idea what may be wrong?

Thanks!

sunnyxiaohu commented 6 years ago

How about you batch_size ? Just decrease your batch_size when testing.

hytseng0509 commented 6 years ago

My batch_size is 4, and the number of mini-batch is 9046, as suggested in the official sample code.

minhtriet commented 6 years ago

Why should decreasing batch_size help?

sunnyxiaohu commented 6 years ago

It may caused by the implementation of the caffe branch. Try it again, if you still get "nan" probability

minhtriet commented 6 years ago

That is actually right, thank you. In my case, I decreased the batch_size from 4 to 1. mini_batch is still 9046.

I think theoretically, the bigger the batch number is, the longer the fraction of video is, then hopefully the more accurate the network is able to classify the action. Will do some experiment to see which batch_size produce best result.

One more thing, could you elaborate more why do you think it may caused by the implementation of the caffe branch?

zhengshou commented 6 years ago

Hi,

Thank you for your interest. This issue seems due to a bug in C3D. I have fixed and updated it.

Best, Zheng