umbcvision / compress Goto Github PK

Hi,

Thank you for sharing code and pretrained weight.

Can I ask you to share pretrained MobileNet v2 (MoCo v2) weight ?

I want to compare MobileNet v2 (MoCo v2 pretrained) with MobileNet v2 (Compress w/ MoCo ResNet-50).

Thank you.

First of all, thanks for releasing the code, but I can not find the resnet50x4 related code, do you forget to push it?

Thanks for sharing your code!
I wonder is train_student.py corresponds to ours-1q while train_student_without_momentum.py corresponds to ours-2q?
The second file is reasonable with the detached momentum encoder and the queue been updated with feats from this encoder.
However, I found in the first file, the student encoder will contrast to the previous samples' representation of itself rather than the teacher encoder, which is not clear to me.
Appreciate for you reply!