Repository files navigation

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

Design space design

P: Paper's. O: Our

Top RegNetX models

Top RegNetY models

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.

With our code, you can:

• Train your model with default arguments by runningpython 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

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

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

• facebookresearch/pycls
• pytorch/examples