FixRes is a technique and GitHub repository from Meta AI that fixes the train test resolution discrepancy in CNNs, allowing models trained at lower resolutions to perform much better at higher test resolutions.

Is FixRes still available and usable?

Yes, the code and pre trained weights are publicly available on GitHub, though the repository has been archived (read only) since May 2024 and is no longer actively updated.

How much accuracy improvement does FixRes provide?

It delivers notable gains, such as boosting ResNet 50 from 77.0% to 79.1% Top-1 on ImageNet at 384px, and even higher with combinations like CutMix.

What do I need to use FixRes?

You'll need Python 3.6+, PyTorch 1.0+, and the dependencies in requirements.txt. Scripts support fine tuning, evaluation, and feature extraction on datasets like ImageNet.

FixRes

Fix Train-Test Resolution Discrepancy for Better Image Classification Accuracy

From Meta AI

Founder: Hugo Touvron, Andrea Vedaldi, Matthijs Douze, Hervé Jégou

Image & Design

Jun 2019

10K+

Free

Pricing Model

Free

Starting Price

$0/Month

About This AI

FixRes is a research method and open-source codebase from Meta AI (formerly Facebook AI Research) that addresses the common performance drop in convolutional neural networks when test images use higher resolutions than training.

By fine-tuning models with adjusted augmentations and resolutions, it significantly boosts accuracy on benchmarks like ImageNet without major architectural changes.

AI Art Image Editor Poster Text To Image

Pricing

Pricing Model

Free

Starting Price

$0/Month

Key Features

Resolves train test resolution mismatch via simple fine-tuning
Supports major architectures like ResNet, ResNeXt, PNASNet, EfficientNet
Provides pre trained models at higher resolutions (e.g., 320px, 384px, 480px, 600px)
Includes scripts for evaluation, fine tuning, and feature extraction
Compatible with techniques like CutMix for further gains

Pros

Easy to apply on existing models with big accuracy lifts (e.g., +2% on ResNet-50)
No need for new architectures or heavy retraining
Reproduces strong results on ImageNet with open code and weights
Improves generalization across different input sizes
Influenced later SOTA models like FixEfficientNet

Cons

Repository is archived and no longer actively maintained since 2024
Requires PyTorch and specific setup for reproduction
Primarily research oriented, not a plug and play production tool
Limited to image classification tasks
License is CC BY-NC 4.0 (non commercial use)

FixRes is excellent for researchers and ML practitioners working on image classification who want simple, effective ways to squeeze more performance from pre-trained CNNs by fixing a common resolution pitfall.

FAQs

What is FixRes?
FixRes is a technique and GitHub repository from Meta AI that fixes the train test resolution discrepancy in CNNs, allowing models trained at lower resolutions to perform much better at higher test resolutions.
Is FixRes still available and usable?
Yes, the code and pre trained weights are publicly available on GitHub, though the repository has been archived (read only) since May 2024 and is no longer actively updated.
How much accuracy improvement does FixRes provide?
It delivers notable gains, such as boosting ResNet 50 from 77.0% to 79.1% Top-1 on ImageNet at 384px, and even higher with combinations like CutMix.
What do I need to use FixRes?
You’ll need Python 3.6+, PyTorch 1.0+, and the dependencies in requirements.txt. Scripts support fine tuning, evaluation, and feature extraction on datasets like ImageNet.