MobileOne: An Improved One millisecond Mobile Backbone

This software project accompanies the research paper, An Improved One millisecond Mobile Backbone.

Our model achieves Top-1 Accuracy of 75.9% under 1ms.

MobileOne Performance

Model Zoo


Model Top-1 Acc. Latency* Pytorch Checkpoint (url) CoreML Model
MobileOne-S0 71.4 0.79 S0(unfused) mlmodel
MobileOne-S1 75.9 0.89 S1(unfused) mlmodel
MobileOne-S2 77.4 1.18 S2(unfused) mlmodel
MobileOne-S3 78.1 1.53 S3(unfused) mlmodel
MobileOne-S4 79.4 1.86 S4(unfused) mlmodel

*Latency measured on iPhone 12 Pro.


To use our model, follow the code snippet below,

import torch
from mobileone import mobileone, reparameterize_model

# To Train from scratch/fine-tuning
model = mobileone(variant='s0')
# ... train ...

# Load Pre-trained checkpoint for fine-tuning
checkpoint = torch.load('/path/to/unfused_checkpoint.pth.tar')
# ... train ...

# For inference
model_eval = reparameterize_model(model)
# Use model_eval at test-time

To simply evaluate our model, use the fused checkpoint where branches are re-parameterized.

import torch
from mobileone import mobileone

model = mobileone(variant='s0', inference_mode=True)
checkpoint = torch.load('/path/to/checkpoint.pth.tar')
# ... evaluate/demo ...

ModelBench App

An iOS benchmark app for MobileOne CoreML models. See ModelBench for addition details on building and running the app.


If our code or models help your work, please cite our paper:

  title={An Improved One millisecond Mobile Backbone},
  author={Vasu, Pavan Kumar Anasosalu and Gabriel, James and Zhu, Jeff and Tuzel, Oncel and Ranjan, Anurag},
  journal={arXiv preprint arXiv:2206.04040},


View Github