Need help with ISONet?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

138 Stars 19 Forks MIT License 2 Commits 0 Opened issues


Deep Isometric Learning for Visual Recognition (ICML 2020)

Services available


Need anything else?

Contributors list

# 53,964
2 commits

Deep Isometric Learning for Visual Recognition

This repository is an official PyTorch implementation of the ICML paper:

Deep Isometric Learning for Visual Recognition
Haozhi Qi, Chong You, Xiaolong Wang, Yi Ma, Jitendra Malik
International Conference on Machine Learning (ICML), 2020
[Project Webpage], [arXiv]


In this project, we argue the notion of isometry is a central guiding principle for training deep ConvNet. In particular, we take a minimalist approach and show that a vanilla deep Isometric Network (ISONet) (i.e., without BN and shortcut) can be trained and achieve suprisingly good accuracy. We also show that if combined with skip connections, such near isometric networks (i.e. R-ISONet) can achieve performances on par with the standard ResNet, even without normalization at all.

Main results

Here we show the Top-1 Classification Accuracy on ImageNet Validation dataset:

| Methods | depth18 | depth 34 | depth 50 | depth 101 | :---: | :---: | :---: | :---: | :---:
| ISONet | 68.10 | 70.90 | 71.20 | 71.01 | R-ISONet | 69.17 | 73.43 | 76.18 | 77.08

For more results and pretrained models, see Model Zoo.

Using ISONet

Data Preparation

The ImageNet folder train/val folder should locate at

, respectively.


This codebase is developed and tested with python 3.6, PyTorch 1.4, and cuda 10.1. But any version newer than that should work.

Here we gave an example of installing ISONet using conda virtual environment: ``` git clone cd ISONet conda create -y -n isonet conda activate isonet

install pytorch according to

conda install -y pytorch==1.4 torchvision cudatoolkit=10.1 -c pytorch pip install yacs ```


You can download the pre-trained models from the links in Model Zoo.

For example, if we want to test the performance of R-ISONet 18, download it from model zoo, and use the following command:

# change config files if you are going to test other pre-trained models
python --cfg configs/IN1k-RISO18.yaml --gpus {GPU_ID} --ckpt


To train our model from scratch, use the following command:

python --cfg {CONFIG_FILE} --gpus {GPU_ID} --output {OUTPUT_NAME}

Citing ISONet

If you find ISONet or this codebase helpful in your research, please consider citing:

  author={Qi, Haozhi and You, Chong and Wang, Xiaolong and Ma, Yi and Malik, Jitendra},
  title={Deep Isometric Learning for Visual Recognition},

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.