MobileOne: An Improved One millisecond Mobile Backbone

This is an un-official implementation of Paper An Improved One millisecond Mobile Backbone, whose performance is close to the paper.

Actually I achieve the MobileOne' s0 architecture (the smallest one), and validate on ImageNet-1000k dataset. And a val accuracy is here.

Note that I only train a "s0" version, but you can easily modify the code to train other version, please see "mobileone.py" to refer to a configuration.

Different from original paper, We don't use:

• AutoAugment. In fact, S0 indeed abandons autoaugment.
• annealed weight decay. I set it constantly 4e-5.
• label smoothing regularization
• EMA update strategy
• progressive learning curriculum. I directly use 224px to train.
• Custom Weight decay Loss. I directly use WeightDecay in optimizer.

very Important

I thank grygielski for finding a bug of my implementation, which is a very easy-correct mistake. please refer to this issue

Because I don't plan to retrain the model, so I don't want to rewrite the code. For the mistake, you can just delete the "if" condition, like this:

self.dw_bn_layer = nn.BatchNorm2d(in_channels) if out_channels == in_channels and stride == 1 else None

to

self.dw_bn_layer = nn.BatchNorm2d(in_channels)

For validation

I release a pretrained model weight, click here to download. The test script validates the trained model, and also generates a converted deploy model.

python test.py {your imagenet-1000k dataset path} deploy mobileone_s0_hello_best.pth.tar

A converted deploy model is generated at "mobileone_deploy_model.pt" file.

For train

I train the mobileone-s0 on 8 32G-V100 GPUS, costing about 4 days.

python train.py -a mobileone_s0 --dist-url 'tcp://127.0.0.1:23333' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0 --workers 32 {your imagenet-1000k dataset path} --tag hello --wd 4e-5

For use

model = make_mobileone_s0(deploy=True)
model.load(torch.load('mobileone_depoly_model.pt'))
model.cuda()

acknowledgement

The entire code is based on RepVGG repository. Thanks for simply-using code.