training on ActivityNet v1.2

[sunsiyang2015]

hi, when I train on ActivityNet v1.2. the I got mIOU is nan.
how to solve the problem?

[sujoyp]

Not sure if this is resolved. Could you please try backtracking through the code to see what is causing the problem? I do not recall getting NaNs in mAP.

[yhZhai]

Hi, have you solved this issue? I encountered the same problem, particularly, the detection IoU and training loss became nan after about 1000 iterations. And I used the following parameter to run:
--dataset-name ActivityNet1.2 --num-class 100 --max-seqlen 3000 --model-name anet1.2 --batch-size 16

[yhZhai]

hi, when I train on ActivityNet v1.2. the I got mIOU is nan.
how to solve the problem?

Hi, I solved this problem. This is because thumos14 used validation set to train, while activitynet used training set to train. So you have to change the subset name in detectionMAP.py (maybe some other files).