MobileNet-Caffe icon indicating copy to clipboard operation
MobileNet-Caffe copied to clipboard

Published 20 hours ago • verified publisher icon shicai

MobileNet accuracy

Open gudovskiy opened this issue 7 years ago • 4 comments

@shicai thanks for your work. I have two Qs:

  1. Did you convert weights from TF to Caffe? I mean, these .caffemodels are not trained in Caffe, right?
  2. Accuracies I get are slightly lower than stated: for V1 69.48 vs. 70.81 and for V2 71.2 vs. 71.90. Did you get your results in Caffe or TF? Paper results are slightly different as well.

gudovskiy avatar Feb 08 '18 00:02 gudovskiy

models are trained with my customized caffe, and convert to standard caffe. accuracy results are tested under standard caffe, please make sure that you use the original images from ImageNet Val dataset (2012).

shicai avatar Feb 08 '18 01:02 shicai

@shicai, I use original val dataset. Somebody also got the same 69.5 for V1 here. Anyway, the difference is not so big, just wonder why.

gudovskiy avatar Feb 08 '18 01:02 gudovskiy

It's so weird. there should be no problem with the pretrained weights. but there might be some problems with different image processing steps or different implementations. when it is converted to mxnet, the model accucary can get to 71.2 (see: https://github.com/KeyKy/mobilenet-mxnet).

shicai avatar Feb 08 '18 02:02 shicai

@shicai : Can you share your implementation of Caffe as well?

Coderx7 avatar Mar 26 '18 19:03 Coderx7