MLP-Mixer

Unoffical Implementation of MLP-Mixer, easy to use with terminal. Train and test easly.

https://arxiv.org/abs/2105.01601

MLP-Mixer is an architecture based exclusively on multi-layer perceptrons (MLPs).

According to paper, Model offers:

• Better accuracy than CNNs and Transformers
• Lower time complexity than CNNs and Transformers
• Lower parameters than CNNs and Transformers

Quick Start

Clone the repo and install the requirements.txt in a Python>=3.8 environment.

git clone https://github.com/Oguzhanercan/MLP-Mixer
cd MLP-Mixer
pip install -r requirements.txt

Dataset

There are 2 options for dataset. You can use pre-defined datasets listed below

• CIFAR10
• Mnist
• Fashion Mnist

or you can use your own dataset. Organize your folder structure as:

      data---
            |
            --0
               |
                --img0.png
                .
                .
                --img9999.png
            |
            -- 1
                |
                --img0.png
                .
                .
                --img9999.png
            .
            .

0 and 1 represents folders that contains images belongs only one particular class. There is no limit for classes or images.

Train

Open a terminal at the same directory of clone. Then run the code below.

python main.py --mode train --dataset CIFAR10 --save True --device cuda --epochs 20 --valid_per 0.2 

You can customize the model hyperparameters, all arguments listed below "Arguments:

• dataset : Categorical Option --- Choose the dataset, Options: CIFAR10, Mnist, Fashion Mnist, Custom
• train_path : Path --- Enter the train path, if you are using custom dataset mode
• test_path : Path --- Enter the test path, if you are using custom dataset mode
• batch_size : integer number ---
• im_size : integer number --- Enter the biggest dimension of image, Example : for 48x32x3 enter 48
• valid_per : float number between 0,1 --- Validation percantage, train dataset will be splitted as train and validation
• epochs : integer number --- Number of epochs to train
• learning_rate : float number --- Learning rate for optimizer
• beta1 : float number between 0,1 --- Beta1 value for adam optimizer
• beta2 : float number between 0,1 --- Beta2 value for adam optimizer
• n_classes : integer number --- Number of classes that dataset has
• cuda : True or false --- if you have cuda compute capability enviroment True suggested
• -eveluate_per_epoch : integer number --- Prints the state information per epoch
• save_model : True or False --- If true model parameters will be saved.
• model_params : Path --- If you have pretrained parameters, enter the path

Custom dataset mode should include following arguments: mode,dataset,train_path,n_classes,im_size