MobileNetV4-PyTorch

Overview

This repository provides a PyTorch replication of the MobileNetV4 architecture as described in the paper "MobileNetV4: Universal Models for the Mobile Ecosystem". The implementation aims to mimic the architecture closely for all five variants:

• MobileNetV4ConvSmall
• MobileNetV4ConvMedium
• MobileNetV4ConvLarge
• MobileNetV4HybridMedium
• MobileNetV4HybridLarge

Repository Structure

• env/: Contains the environment YAML file to set up the necessary dependencies.
• logs/: Contains the architecture details of the different MobileNetV4 variants.
• paper/: Contains the original MobileNetV4 paper for reference.
• MobileNetV4.py: Contains the feature extractor for MobileNetV4 architectures.
• nn_blocks.py: Contains neural network block definitions used in the MobileNetV4 architecture.
• test.py: Contains the classifier and script for testing the implementations.

Installation

To create the environment with the necessary dependencies, use the provided YAML file:

conda env create -f env/MobileNetV4_env.yml
conda activate MobileNetV4-PyTorch

Usage

Training

To train a MobileNetV4 model on your dataset, modify the test.py script with your dataset and training configurations.

Pre-trained Weights

For pre-trained weights on ImageNet, you can use the weights provided by timm.

Example

import torch
from test import MobileNetV4WithClassifier

# Example usage
model = MobileNetV4WithClassifier(model_name='MobileNetV4ConvSmall', num_classes=1000)
input_tensor = torch.randn(1, 3, 224, 224)
output = model(input_tensor)
print(output)

Citations

If you find this work useful, please cite the original MobileNetV4 paper:

@article{MobileNetV4,
  title={MobileNetV4: Universal Models for the Mobile Ecosystem},
  author={Author Names},
  journal={arXiv preprint arXiv:2404.10518v1},
  year={2024}
}

If you use this work, please cite it as follows:

@misc{MobileNetV4-PyTorch,
  author = {Muhammad Junaid Ali Asif Raja},
  title = {MobileNetV4-PyTorch},
  year = {2024},
  url = {https://github.com/junaidaliop/MobileNetV4},
  note = {Version 1.0.0}
}

Contact

For research collaborations or any queries, please email me at muhammadjunaidaliasifraja@gmail.com

Contributions

Contributions are welcome! Please submit a pull request or open an issue to discuss your ideas.

TODO

• Train the model on ImageNet to attain weights
• Train the model on CIFAR-100
• Train the model on CIFAR-10

Star the Repository

If you find this repository useful, please consider giving it a star!

Acknowledgment

This is an unofficial implementation of MobileNetV4 in PyTorch. To the best of my ability, I believe this is the closest implementation to the original work found at TensorFlow MobileNetV4 Implementation.

For the official TensorFlow implementation, please visit: TensorFlow MobileNetV4 Implementation.