d-li14/mobilenetv2.pytorch

how many epochs did you train?

Moran232 opened this issue · 3 comments

Moran232 commented

Hi, thanks for your nice work.
I am wandering how many epochs you trained? the default epochs=90 in your code and it may take around 60hrs. while the lr policy you set is more than 90 epochs .

nzmora commented

Hi,

I am also interested in understanding how you created the pre-trained models. I'm the author of NN Distiller, and I'd like to point our users to the pretrained mobilenet_v2 in your repository, but it is much better if we know the training hyper-parameters (this is because we need to use them, after pruning/compression for the fine-tuning phase).

If you can provide the full command-line, that would be best. If this is not possible, please provide the initial LR, LR decay policy, the number of epochs trained, the batch size, WD.

Thank you!
Neta

  • 👍2
d-li14 commented

Hi Neta,

This repo is actually also associated with some works during my internship at Intel Labs China. Please check your Intel e-mail for more information. If you have further questions, feel free to contact me anytime. @nzmora

Regards,
Duo

  • 👍1
d-li14 commented

@nzmora Please refer to https://github.com/d-li14/mobilenetv2.pytorch/blob/master/README.md#training for the full command line to reproduce.

  • 👍3