PyTorch Image Model Framework

Overview

Nowadays, developers use PyTorch or Tensorflow frameworks to train image classification models. This course focuses on PyTorch implementation via an open-source Python library called the PyTorch Image Model.

The PyTorch Image Model (timm) is a deep-learning framework created by Ross Wightman. It contains models, utilities, data loaders, and scripts related to computer vision. Most of the base implementation relies on the state-of-the-art algorithms adapted for PyTorch.

The main objective of this framework is to provide an out-of-the-box library for researchers and developers to reproduce training results for ImageNet.

It’s in active development and is suitable for testing various state-of-the-art architectures for image classification.

The PyTorch Image Model framework offers a lot of possible architectures with the option to change it with a single line of code. Some of the most popular architectures include:

• MixNet-XL
• SE-ResNeXt-26-D
• EfficientNet-B0
• EfficientNet-B2
• EfficientNet-B3
• EfficientNet-ES
• ResNet50
• MobileNetV3-Large-100

In this course, we’ll focus on the usage of the following architectures:

• ResNet50
• EfficientNet-B0

---

Validation

The PyTorch Image Model also provides its own validation results. Developers can cross-check the output and benchmark the performance of their desired models.

Based on its latest benchmark, ResNet50 achieved 79.038 in top-1 accuracy. The top-5 accuracy was 94.390.

---

License

Code

The code provided by this framework is based on the Apache 2.0 license. We can train custom classification models with it for commercial purposes.

Pre-trained weights (ImageNet)

Most of the pre-trained weights provided by this framework were trained using ImageNet datasets, ImageNet was released for noncommercial purposes. If we intend to use the pre-trained weights directly, we have to seek legal advice on it.

Pre-trained weights (Others)

Pre-trained weights such as Facebook’s WSL, SSL, SWSL ResNe(Xt), and the Google Noisy Student EfficientNet were trained using proprietary datasets with an explicit noncommercial license (CC-BY-NC 4.0).

---

Documentation

We can find the official documentation at Pytorch Image Model. There’s also alternative comprehensive documentation called timmdocs.