anjiecheng / mnasnet-pytorch

A PyTorch implementation of Mnasnet: MnasNet: Platform-Aware Neural Architecture Search for Mobile.

Overview

A PyTorch implementation of Mnasnet searched architecture: MnasNet: Platform-Aware Neural Architecture Search for Mobile. On the ImageNet classification task, the model achieves 74.0% top-1 accuracy with 76ms latency on a Pixel phone, which is 1.5× faster than MobileNetV2

MnasNet Architecture

Disclaimer

Codes modified from mobilenet-v2

mnasnet-pytorch's Issues

About stride's config

In code, the stride's config is like

        # t, c, n, s, k
        [3, 24,  3, 1, 3],  # -> 56x56
        [3, 40,  3, 2, 5],  # -> 28x28
        [6, 80,  3, 2, 5],  # -> 14x14
        [6, 96,  2, 2, 3],  # -> 14x14
        [6, 192, 4, 1, 5],  # -> 7x7
        [6, 320, 1, 2, 3],  # -> 7x7   

But in paper, it seems to be

        # t, c, n, s, k
        [3, 24,  3, 2, 3],  # -> 56x56
        [3, 40,  3, 2, 5],  # -> 28x28
        [6, 80,  3, 2, 5],  # -> 14x14
        [6, 96,  2, 1, 3],  # -> 14x14
        [6, 192, 4, 2, 5],  # -> 7x7
        [6, 320, 1, 1, 3],  # -> 7x7   

the two configs will make quite big differences in speed as well as feature map size
This made me confused, which is correct ?
Many thanks !

hyper params

Thank you for sharing your codes!

1. Could you also please share your hyper params (such as lr and weight_decay and batchsize)
2. Did you do data pre-processing as the same as here
3. And is there any other training skill during your training, since I stuck at Top1=65% when I tried to train it on imagenet.

Thank you again! Appreciate your help!

pytorch version?

please tell me the version of pytorch?

License?

Hello,

I noticed you didn't have a license, which makes it hard to do anything but look at this code without potentially get into trouble. If I may ask, would it be possible for you to add one?

Thanks,
Charlie West

how to re-implement your job?

hi anjieZheng
How ot re-implement your job by your repo?
Thanks!

BR

the pretrained model

@AnjieZheng Would you please release the pretrained model on ImageNet?