austingg / mobilenet-v2-caffe Goto Github PK

For example, the lr policy, thanks

Hi Austing,
You said there is not an optimized Depthwise conv layer in caffe.
I found there're 2 dw conv implementations:

https://github.com/BVLC/caffe/pull/5665/files
https://github.com/yonghenglh6/DepthwiseConvolution

Are they not good enough during your experiment? Or is there any other reasons for not training in caffe?

1. 请问你可以放出solver.prototxt吗？
2. 第一个batchnorm前的conv使用了bias_term: true，剩余所有的batchnorm前的conv都使用了bias_term: false，请问这样做的依据是什么？
3. 关于scale层的设置，你使用了默认mult参数，即1,1,1,1而mobilenet v1中scale层mult参数为1,0,1,0或者1, 0, 2, 0即不使用decay_mult，请问这样做的原因是什么？

Hi, austingg:

Thanks for your sharing of the mobilentv2 network. I want to repeat the experiment and hoping to restire the paper accuracy. There are three questions for me about the dataset.

1. Did you employ only the imagenet2012 dataset or also took the imagenet 2016 data when training?
2. When resizing the origin images, did you resize them to fixed 256x256 or 256xN. I hear that the later one might result in better accuracy.
3. How the data augmentation helps for the accrcury? Is there any comparison?

Thank you for your implementation. Have you released any pre-train model on imagenet database

there is an error in your MobileNetV2_deploy.prototxt about the use_global_stats.

Hi, I compute the the MAdds of your .prototxt, the MAdds is 313M, which is bigger than the MAdds stated in the paper(300M). I can not figure out where went wrong. I think my computation is right(I have used the script to compute other models), and I think your implementation is also right. Do you have any idea about this? Have you computed the MAdds of your implementation? Thanks in advance.

Is this data augmentation method mentioned in the “going deep with convolutions” paper?And in training or in testing to do data augmentation. I only use single-crop in picture's center now, so I have a lower accuracy in testing other's models. I often failed to achieve the accuracy mentioned in the paper, so I suspect that is when on the training only use mirror to augment data.

Using my TitanX(Maxwell), I got 20ms for a 224*224 image inference with batch=1.
How fast could you achieve on caffe?
Thanks,

Alex