MLP-Mixer-CIFAR10

MLP-Mixer architectures for classifying CIFAR10 images

This repo contains a re-implementation of the MLP-Mixer model and another MLP-Mixer-inspired model to classify CIFAR-10 images. Both pre-trained and trained models are provided.

Two main files are provided:

main_cifar10.py for training both models from scratch;
main_tiny_imagenet.py for pre-training both models on TinyImageNet (200 classes - 500 images/class) and then fine-tuning on CIFAR-10.

Steps:

Download the TinyImageNet dataset:

wget http://cs231n.stanford.edu/tiny-imagenet-200.zip
unzip tiny-imagenet-200.zip

Run one of the two main files:

python3 main_cifar10.py

python3 main_tiny_imagenet.py

Results:

	Mixer-MLP	Mixer-MLP-Inspired
From scratch	57.54 %	64.48 %
Pre-trained on TinyImageNet	68.36 %	56.87 %

Optional parameters can be provided as input (e.g., batch_size, depth, n_epochs, ...). Open the main files for the complete list of allowed input parameters.

To download both results and pre-trained models:

mkdir Results
cd Results
curl -L https://www.dropbox.com/sh/rycvh295g6975dd/AADCN1vYImM7gYdc5_omKjH7a?dl=0 >> Results.zip
unzip Results
rm Results.zip

TODO: parameters fine-tuning; solve over-fitting.

pasqualecoscia/MLP-Mixer-CIFAR10

MLP-Mixer-CIFAR10