Regnet

Introduction

Here is our pytorch pytorch implementation of network design paradigm described in the paper "Designing Network Design Spaces"

Design space design

Comparison

P: Paper's. O: Our

Model	[P/O] gflops	[P/O] params	[P/O] top-1 error
RerNetY-200MF	0.2/0.22	3.2/3.27	29.6/updating...
RerNetY-400MF	0.4/0.42	4.3/4.45	25.9/updating...
RerNetY-600MF	0.6/0.60	6.1/5.66	24.5/updating...
RerNetY-800MF	0.8/0.82	6.3/6.26	23.7/updating...

Best models

Top RegNetX models

Top RegNetY models

Datasets

We use Imagenet (ILSVRC2012) for all experiments, as stated in the paper.

Create a data folder under this repository,

cd {repo_root}
mkdir data

• ImageNet: Download the ImageNet dataset and put the files as the following structure:

  data
  ├── train
  │   ├── n01440764
  │   └── n01443537
  │   └── ...
  │── val
  │   ├── n01440764
  │   └── n01443537
  │   └── ...
  

  Of course you could change this path to whatever you want based on your own preference, or mount it to a folder when using docker.

How to use our code

With our code, you can:

• Train your model with default arguments by running python train.py -d path/to/image/root/folder
• We also provide shell scripts which could be used to run training for first RegnetY models at ./scripts/. For example, if you want to train RegNetY 800MF, you could simply run ./scripts/RegnetY_800MF.sh

Requirements

• python 3.7
• pytorch 1.4
• opencv (cv2)
• pthflops
• torchsummary

Updating (21/04/2020)

Complete all networks and training script. We are training RegnetY models and will update result soon.

References

• facebookresearch/pycls
• pytorch/examples