federicopozzi33 / mobileone-pytorch

A PyTorch implementation of MobileOne

MobileOne PyTorch

Unofficial PyTorch implementation of An Improved One millisecond Mobile Backbone paper.

Quickstart

Install with pip install mobileone_pytorch and create a MobileOne with:

from mobileone_pytorch import mobileone_s1
model = mobileone_s1()

Overview

This repository contains an implementation of MobileOne.

Features:

• Implementation of all MobileOne versions
• Reparametrization for model deployment

Upcomining features:

• Squeeze-and-Excitation block for MobileOne S4

Help wanted:

• Training models on ImageNet

Table of contents

1. About MobileOne
2. Installation
3. Usage
   • Create models
   • Deployment via reparametrization

About MobileOne

MobileOne is a novel architecture that with variants achieves an inference time under 1 ms on an iPhone12 with 75.9% top-1 accuracy on ImageNet.

• MobileOne achieves state-of-the-art performance within the efficient architectures while being many times faster on mobile.

• The best model (S4) obtains similar performance on ImageNet as Mobile-Former while being 38× faster. Moreover it obtains 2.3% better top-1 accuracy on ImageNet than EfficientNet at similar latency.

Installation

Install via pip:

pip install mobileone_pytorch

Or install from source:

git clone https://github.com/federicopozzi33/MobileOne-PyTorch.git
cd mobileone_pytorch
pip install -e .

Usage

Create models

Create MobileOne models:

from mobileone_pytorch import (
   mobileone_s0, 
   mobileone_s1, 
   mobileone_s2, 
   mobileone_s3, 
   mobileone_s4
)

model_s0 = mobileone_s0()
model_s1 = mobileone_s1()
model_s2 = mobileone_s2()
model_s3 = mobileone_s3()
model_s4 = mobileone_s4()

Deployment

Deploy a MobileOne through reparametrization:

import torch
from mobileone_pytorch import mobileone_s1

x = torch.rand(1, 3, 224, 224)

model = mobileone_s1()
deployed = model.reparametrize()

model.eval()
deployed.eval()

out1 = model(x)
out2 = deployed(x)

torch.testing.assert_close(out1, out2)

Contributing

If you find a bug, create a GitHub issue. Similarly, if you have questions, simply post them as GitHub issues.

mobileone-pytorch's Issues

Add pre-trained model

Can anyone add a pre-trained model for the MobileOne network?

Hello, where is the model's weights

Hello,How to train? Could you please provide more detailed steps?

I created a file named 'create_demo.py',and do just as you show,but nothing happened.
And I am confused with the training process.
Could you please provide more detailed steps? Thank you.

Questions about Multiple Inputs？

How do multiple inputs reparametrize, because the reparametrize function returns nn.Sequential, my multiple inputs are input between different layers, for example two inputs in the first layer and two inputs in the nth layer
The only thing I can think of at present is to build a model after reparametrize according to the trained model, and then read the parameters of the reparametrize model after training.
Is there a better way ？