Home

Awesome

SplitMixer: Fat Trimmed From MLP-like Models (Arxiv)

PyTorch implementation of the SplitMixer MLP model for visual recognition

by Ali Borji and Sikun Lin

Code overview

The most important code is in splitmixer.py. We trained SplitMixers (on ImageNet) using the timm framework, which we copied from here.

For CIFAR-{10,100} trainings or standalone model definitions, please refer to the cifar notebook.

Inside pytorch-image-models, we have made the following modifications:

Evaluation

CIFAR-10

Patch Size p=2, Kernel Size k=5

Model NameParams (M)FLOPS (M)Acc
ConvMixer-256/80.60152.694.17
SplitMixer-I 256/80.2871.893.91
SplitMixer-II 256/80.1746.292.25
SplitMixer-III 256/80.1779.892.52
SplitMixer-IV 256/80.3179.893.38

CIFAR-100

Patch Size p=2, Kernel Size k=5

Model NameParams (M)FLOPS (M)Acc
ConvMixer-256/80.62152.673.92
Splitixer-I 256/80.3071.972.88
SplitMixer-II 256/80.1946.270.44
SplitMixer-III 256/80.1979.870.89
SplitMixer-IV 256/80.3279.871.75

Flowers102

Patch Size p=7, Kernel Size k=7

Model NameParams (M)FLOPS (M)Acc
ConvMixer-256/80.7069660.47
Splitixer-I 256/80.3433162.03
SplitMixer-II 256/80.2422959.33
SplitMixer-III 256/80.2436359.00
SplitMixer-IV 256/80.3736361.51

Foods101

Patch Size p=7, Kernel Size k=7

Model NameParams (M)FLOPS (M)Acc
ConvMixer-256/80.7069674.59
Splitixer-I 256/80.3433173.56
SplitMixer-II 256/80.2422971.74
SplitMixer-III 256/80.2436372.78
SplitMixer-IV 256/80.3736372.92

ImageNet

Stay Tuned!

Citation

If you use this code in your research, please cite this project.

@inproceedings{borji2022SplitMixer,
title={SplitMixer: Fat Trimmed From MLP-like Models},
author={Ali Borji and Sikun Lin},
booktitle={Arxiv},
year={2022},
url={https://arxiv.org/pdf/2207.10255.pdf}
}