human-analysis / muxconv

MUXConv: Information Multiplexing in Convolutional Neural Networks (CVPR '20), Pytorch Implementation

MUXConv

Code accompanying the paper.

MUXConv: Information Multiplexing in Convolutional Neural Networks

Zhichao Lu, Kalyanmoy Deb, and Vishnu Boddeti

CVPR 2020

Requirements

Python >= 3.7.x, PyTorch >= 1.4.0, torchvision >= 0.5.0, timm == 0.1.14, 
torchprofile >= 0.0.1 (optional for calculating FLOPs)

ImageNet Classification

Tranfer to CIFAR-10 and CIFAR-100

Pretrained models

The easiest way to get started is to evaluate our pretrained MUXNet models. Pretrained models are available from Google Drive.

python eval.py --dataset [imagenet/cifar10/cifar100] \
	       --data /path/to/dataset --batch-size 128 \
	       --model [muxnet_s/muxnet_m/muxnet_l] \ 
	       --pretrained /path/to/pretrained/weights

Train

To re-train from scratch on ImageNet, use distributed_train.sh from pytorch-image-models and follow the recommended training hyperparameter setting for EfficientNet-B0.

To re-train on CIFAR (transfer) from ImageNet, run

python transfer_cifar.py --dataset [cifar10/cifar100] \
			 --data /path/to/dataset \
			 --model [muxnet_s/muxnet_m/muxnet_l] \
			 --imagenet /path/to/pretrained/imagenet/weights

Citation

If you find the code useful for your research, please consider citing our works

@article{muxconv,
  title={MUXConv: Information Multiplexing in Convolutional Neural Networks},
  author={Lu, Zhichao and Deb, Kalyanmoy and Boddeti, Vishnu},
  booktitle={CVPR},
  year={2020}
}

Acknowledgement

Codes heavily modified from pytorch-image-models and pytorch-cifar10.

muxconv's Issues

Models for ChestX-Ray14

Hi authors, thanks for sharing this wonderful work! I am currently working on ChestX-ray14 as well. Is it convenient for you to share the re-training code and model weights for this dataset? I couldn't appreciate too much!

from torchprofile import profile_macs

hello, when I run this code, have a problem, how to install torchprofile? Thanks

Train on datasets other than imagenet and cifar

Hi there,

as I see in your guide, we can set the dataset to only Imagenet or Cifar, is it possible to train the model on other datasets such as Chexpert ?? if so, can you please provide a guide as to how we can do that?

Thanks in advance

about MUX conv parameter scales

hello,when I see this code,I have a problem. in the class of MuxInvertedResidual ,here is a parameter "scales", default 0,and if you want to use mux conv,this parameter must be a list. in MUXNet, this "scales" is always 0，that is to say, MUXNet is not used MUXcov?

pretrained models

Dear authors,
Thanks for sharing your wonderful work.
I am trying to experiment your work, but I found the pretrained model on the Google Doc contains only the m model.
It would be great if you can kindly share xs, s, and L models as well.

Best