search

naraysa / 3c-net

Weakly-supervised Action Localization
49 stars 9 forks source link

Can't achieve mAP as reported in the paper #3

Open JudeLiu opened 4 years ago

JudeLiu commented 4 years ago

I run the code for a several times, the mAP at IOU=0.5 is 25.34%. Then I use exponential lr scheduling, and it increases to 26.06%. But still lower than 26.6% as reported. Can you possibly explain what could be wrong? Thanks in advance.

naraysa commented 4 years ago

Kindly check your Pytorch version and run the code with the default settings. We obtained the reported results with version 0.4.1.

JudeLiu commented 4 years ago

yes, the pytorch version is 0.4.1. If the code is run with defaults, I get 25.34% detection mAP at 0.5 IOU and 86.06% classification mAP (on THUMOS14). I noticed that the classification mAP and detection mAP both decreased after 30k iterations, which is why I tried learning rate decay.

naraysa commented 4 years ago

Not sure what the problem is, since a few others could obtain the detection mAP around the reported numbers in the paper. We have not used any decay. It is constant lr. Kindly share the generated log file and the command used for the experiment.