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.