guglielmocamporese / break_cifar10

Code for the Top-1 submission of contest of VCS AY 2020-2021, the Vision and Cognitive Service class, University of Padova, Italy.

How to Break the CIFAR-10

Code for the Top-1 submission of contest of VCS AY 2020-2021, the Vision and Cognitive Service class, University of Padova, Italy.

If you find this repo useful, please consider to give a star ⭐ !

The CIFAR-10 Dataset [link]

Results on CIFAR-10

	ResNet18 [1]	ResNet34 [1]	ResNet50 [1]	ResNet Ens [1]	MLP-Mixer [2]	ViT-S/16 [3]	ViT-B/16 [3]
Accuracy	95.01 %	96.92 %	95.46 %	97.53 %	94.67 %	96.05 %	98.67 %

Install

# Clone the repo 
$ git clone https://github.com/guglielmocamporese/break_cifar10.git break_cifar10

# Go the in project folder
$ cd break_cifar10

# Install conda env with all the needed packages
$ conda env create -f environment.yml

# Activate the conda env
$ conda activate torch

Train and Evaluate a model

# Run the main
$ python main.py --mode train --model vit

# Test
$ python main.py --mode test --model vit --model_checkpoint 'your/model/checkpoint.ckpt'

You can change the model in the config.py file. Supported models are ResNet-18, ResNet-34, ResNet50, ViT and MLP-Mixer.