I am trying to replace Mobilenet-V1 with Mobilenet-V2. Any insights?
Please check the paper. In short, it did not give any speed/accuracy improvement.
I did check it. However, I was trying to implement MobileNetV2 on your algorithm. I am getting errors. Can you share the code please?
Thanks,
Or if you can share the pretrained model for MobileNet V2, that would be great too!
I don't have this code. You may post the error, may be it is possible to debug it.
Okay. Thank you!
I read your paper on this implementation and have one question. What does 'cut to conv4_1', 'cut to conv5_5', etc. mean in the MobileNet backbone selection?
This means how many layers from backbone take. MobileNet v1 has 6 blocks, which consist of depthwise convolutions and 1x1 pointwise convolutions. After each block stride is grow (there is convolution with stride == 2). First number conv5_5 is block index, the second conv5_5 is index of convolution (depthwise + pointwise) within block. We remove the stride in conv4_2 and add dilation in conv5_1, you can check it in the code.
Cool. This explains a lot.
Thanks once again!
You are welcome!
Is there any specific way you calculated GLOPS and Number of parameters? I am planning to use https://github.com/sovrasov/flops-counter.pytorch
This one is ok, looks like we have used it. You can also install it through pip.