Implementation of Supervised Contrastive Learning with AMP, EMA, SWA, and many other tricks

Overview

SupCon-Framework

The repo is an implementation of Supervised Contrastive Learning. It's based on another implementation, but with several differencies:

  • Fixed bugs (incorrect ResNet implementations, which leads to a very small max batch size),
  • Offers a lot of additional functionality (first of all, rich validation).

To be more precise, in this implementations you will find:

  • Augmentations with albumentations
  • Hyperparameters are moved to .yml configs
  • t-SNE visualizations
  • 2-step validation (for features before and after the projection head) using metrics like AMI, NMI, mAP, precision_at_1, etc with PyTorch Metric Learning.
  • Exponential Moving Average for a more stable training, and Stochastic Moving Average for a better generalization and just overall performance.
  • Automatic Mixed Precision (torch version) training in order to be able to train with a bigger batch size (roughly by a factor of 2).
  • LabelSmoothing loss, and LRFinder for the second stage of the training (FC).
  • TensorBoard logs, checkpoints
  • Support of timm models, and pytorch-optimizer

Install

  1. Clone the repo:
git clone https://github.com/ivanpanshin/SupCon-Framework && cd SupCon-Framework/
  1. Create a clean virtual environment
python3 -m venv venv
source venv/bin/activate
  1. Install dependencies
python -m pip install --upgrade pip
pip install -r requirements.txt

Training

In order to execute Cifar10 training run:

python train.py --config_name configs/train/train_supcon_resnet18_cifar10_stage1.yml
python swa.py --config_name configs/train/swa_supcon_resnet18_cifar10_stage1.yml
python train.py --config_name configs/train/train_supcon_resnet18_cifar10_stage2.yml
python swa.py --config_name configs/train/swa_supcon_resnet18_cifar10_stage2.yml

In order to run LRFinder on the second stage of the training, run:

python learning_rate_finder.py --config_name configs/train/lr_finder_supcon_resnet18_cifar10_stage2.yml

The process of training Cifar100 is exactly the same, just change config names from cifar10 to cifar100.

After that you can check the results of the training either in logs or runs directory. For example, in order to check tensorboard logs for the first stage of Cifar10 training, run:

tensorboard --logdir runs/supcon_first_stage_cifar10

Visualizations

This repo is supplied with t-SNE visualizations so that you can check embeddings you get after the training. Check t-SNE.ipynb for details.

Those are t-SNE visualizations for Cifar10 for validation and train with SupCon (top), and validation and train with CE (bottom).

Those are t-SNE visualizations for Cifar100 for validation and train with SupCon (top), and validation and train with CE (bottom).

Results

Model Stage Dataset Accuracy
ResNet18 Frist CIFAR10 95.9
ResNet18 Second CIFAR10 94.9
ResNet18 Frist CIFAR100 79.0
ResNet18 Second CIFAR100 77.9

Note that even though the accuracy on the second stage is lower, it's not always the case. In my experience, the difference between stages is usually around 1 percent, including the difference that favors the second stage.

Training time for the whole pipeline (without any early stopping) on CIFAR10 or CIFAR100 is around 4 hours (single 2080Ti with AMP). However, with reasonable early stopping that value goes down to around 2.5-3 hours.

Custom datasets

It's fairly easy to adapt this pipeline to custom datasets. First, you need to check tools/datasets.py for that. Second, add a new class for your dataset. The only guideline here is to follow the same augmentation logic, that is

        if self.second_stage:
            image = self.transform(image=image)['image']
        else:
            image = self.transform(image)

Third, add your dataset to DATASETS dict still inside tools/datasets.py, and you're good to go.

FAQ

  • Q: What hyperparameters I should try to change?

    A: First of all, learning rate. Second of all, try to change the augmentation policy. SupCon is build around "cropping + color jittering" scheme, so you can try changing the cropping size or the intensity of jittering. Check tools.utils.build_transforms for that.

  • Q: What backbone and batch size should I use?

    A: This is quite simple. Take the biggest backbone you can, and after that take the highest batch size your GPU can offer. The reason for that: SupCon is more prone (than regular classification training with CE/LabelSmoothing/etc) to improving with stronger backbones. Moverover, it has a property of explicit hard positive and negative mining. It means that the higher the batch size - the more difficult and helpful samples you supply to your model.

  • Q: Do I need the second stage of the training?

    A: Not necessarily. You can do classification based only on embeddings. In order to do that compute embeddings for the train set, and at inference time do the following: take a sample, compute its embedding, take the closest one from the training, take its class. To make this fast and efficient, you something like faiss for similarity search. Note that this is actually how validation is done in this repo. Moveover, during training you will see a metric precision_at_1. This is actually just accuracy based solely on embeddings.

  • Q: Should I use AMP?

    A: If your GPU has tensor cores (like 2080Ti) - yes. If it doesn't (like 1080Ti) - check the speed with AMP and without. If the speed dropped slightly (or even increased by a bit) - use it, since SupCon works better with bigger batch sizes.

  • Q: How should I use EMA?

    A: You only need to choose the ema_decay_per_epoch parameter in the config. The heuristic is fairly simple. If your dataset is big, then something as small as 0.3 will do just fine. And as your dataset gets smaller, you can increase ema_decay_per_epoch. Thanks to bonlime for this idea. I advice you to check his great pytorch tools repo, it's a hidden gem.

  • Q: Is it better than training with Cross Entropy/Label Smoothing/etc?

    A: Unfortunately, in my experience, it's much easier to get better results with something like CE. It's more stable, faster to train, and simply produces better or the same results. For instance, in case on CIFAR10/100 it's trivial to train ResNet18 up tp 96/81 percent respectively. Of cource, I've seen cased where SupCon performs better, but it takes quite a bit of work to make it outperform CE.

  • Q: How long should I train with SupCon?

    A: The answer is tricky. On one hand, authors of the original paper claim that the longer you train with SupCon, the better it gets. However, I did not observe such a behavior in my tests. So the only recommendation I can give is the following: start with 100 epochs for easy datasets (like CIFAR10/100), and 1000 for more industrial ones. Then - monitor the training process. If the validaton metric (such as precision_at_1) doesn't impove for several dozens of epochs - you can stop the training. You might incorporate early stopping for this reason into the pipeline.

Owner
Ivan Panshin
Machine Learning Engineer: CV, NLP, tabular data. Kaggle (top 0.003% worldwide) and Open Source
Ivan Panshin
:couple: Multi-user accounts for Django projects

django-organizations Summary Groups and multi-user account management Author Ben Lopatin (http://benlopatin.com) Status Separate individual user ident

Ben Lopatin 1.1k Jan 09, 2023
Quick and simple security for Flask applications

Note This project is non maintained anymore. Consider the Flask-Security-Too project as an alternative. Flask-Security It quickly adds security featur

Matt Wright 1.6k Dec 19, 2022
Crie seus tokens de autenticação com o AScrypt.

AScrypt tokens O AScrypt é uma forma de gerar tokens de autenticação para sua aplicação de forma rápida e segura. Todos os tokens que foram, mesmo que

Jaedson Silva 0 Jun 24, 2022
Official implementation of the AAAI 2022 paper "Learning Token-based Representation for Image Retrieval"

Token: Token-based Representation for Image Retrieval PyTorch training code for Token-based Representation for Image Retrieval. We propose a joint loc

Hui Wu 42 Dec 06, 2022
Simple Login - Login Extension for Flask - maintainer @cuducos

Login Extension for Flask The simplest way to add login to flask! How it works First, install it from PyPI: $ pip install flask_simplelogin Then, use

Flask Extensions 181 Jan 01, 2023
A fully tested, abstract interface to creating OAuth clients and servers.

Note: This library implements OAuth 1.0 and not OAuth 2.0. Overview python-oauth2 is a python oauth library fully compatible with python versions: 2.6

Joe Stump 3k Jan 02, 2023
A JSON Web Token authentication plugin for the Django REST Framework.

Simple JWT Abstract Simple JWT is a JSON Web Token authentication plugin for the Django REST Framework. For full documentation, visit django-rest-fram

Simple JWT 3.3k Jan 01, 2023
Implementation of Supervised Contrastive Learning with AMP, EMA, SWA, and many other tricks

SupCon-Framework The repo is an implementation of Supervised Contrastive Learning. It's based on another implementation, but with several differencies

Ivan Panshin 132 Dec 14, 2022
🔐 Login & Register System

🔐 Login & Register System This is a developable login and register system. Enter your username and password to register or login to account. Automati

Firdevs Akbayır 10 Dec 12, 2022
FastAPI extension that provides JWT Auth support (secure, easy to use, and lightweight)

FastAPI JWT Auth Documentation: https://indominusbyte.github.io/fastapi-jwt-auth Source Code: https://github.com/IndominusByte/fastapi-jwt-auth Featur

Nyoman Pradipta Dewantara 468 Jan 01, 2023
Mock authentication API that acceccpts email and password and returns authentication result.

Mock authentication API that acceccpts email and password and returns authentication result.

Herman Shpryhau 1 Feb 11, 2022
Basic auth for Django.

Basic auth for Django.

bichanna 2 Mar 25, 2022
Django Auth Protection This package logout users from the system by changing the password in Simple JWT REST API.

Django Auth Protection Django Auth Protection This package logout users from the system by changing the password in REST API. Why Django Auth Protecti

Iman Karimi 5 Oct 26, 2022
This is a Token tool that gives you many options to harm the account.

Trabis-Token-Tool This is a Token tool that gives you many options to harm the account. Utilities With this tools you can do things as : ·Delete all t

Steven 2 Feb 13, 2022
Per object permissions for Django

django-guardian django-guardian is an implementation of per object permissions [1] on top of Django's authorization backend Documentation Online docum

3.3k Jan 01, 2023
A JSON Web Token authentication plugin for the Django REST Framework.

Simple JWT Abstract Simple JWT is a JSON Web Token authentication plugin for the Django REST Framework. For full documentation, visit django-rest-fram

Jazzband 3.2k Dec 28, 2022
Extending the Django authentication system with a phone verification step.

Extending the Django authentication system with a phone verification step.

Miguel Grinberg 50 Dec 04, 2022
Python's simple login system concept - Advanced level

Simple login system with Python - For beginners Creating a simple login system using python for beginners this repository aims to provide a simple ove

Low_Scarlet 1 Dec 13, 2021
Simple extension that provides Basic, Digest and Token HTTP authentication for Flask routes

Flask-HTTPAuth Simple extension that provides Basic and Digest HTTP authentication for Flask routes. Installation The easiest way to install this is t

Miguel Grinberg 1.1k Jan 05, 2023
OAuth2 goodies for the Djangonauts!

Django OAuth Toolkit OAuth2 goodies for the Djangonauts! If you are facing one or more of the following: Your Django app exposes a web API you want to

Jazzband 2.7k Jan 01, 2023