svishwa / crowdcount-mcnn

Single Image Crowd Counting via MCNN (Unofficial Implementation)
MIT License
499 stars 179 forks source link

BatchNorm momentum #64

Open Anzoteh96 opened 3 years ago

Anzoteh96 commented 3 years ago

Hello,

Not too sure if it's an appropriate to raise this, but I felt it strange that the momentum of batch normalization is set to 0 here:

https://github.com/svishwa/crowdcount-mcnn/blob/e04c67f799ff024dbea93b81a079ffdd1b6daa28/src/network.py#L11

According to the documentation it seems that we use the average value of the first batch ever seen.

I wonder if this is intended since many guides suggest 0.1, 0.01, etc (but not 0).