Meta Architecture Search

This repository contains the trained deep neural network architectures and weights, and training code for the BASE paper.

If you find this useful, or if you use it in your work, please cite:

@inproceedings{2019_SqueezeNAS,
    author = {Albert Shaw and Wei Wei and Weiyang Liu and Le Song and Bo Dai},
    title = {Meta Architecture Search},
    booktitle = {NeurIPS},
    year = {2019}
}

Requirements

Python >= 3.6.0
PyTorch >= 1.0.1
torchvision >= 0.2.2
numpy >= 1.15.4
Pillow

Instructions

1. Install the required packages.
2. Clone this repository.
3. Download and extract the Imagenet dataset to data/imagenet.

Evaluation

Use the train.py script to evaluate the models. Logs are saved into the logs folder.

Training the networks on cifar10 requires one 1080 TI and 2 1080 TI to train Imagenet.

To evaluate the trained networks run:
python3 train.py --model=get_cifar_tuned_model(True) --gpu 1 --eval 1
python3 train_imagenet.py --model=get_imagenet_tuned_model(True) --gpu 1 --eval 1

To train the found networks run:
python3 train.py --model=get_cifar_tuned_model(False) --gpu 1
python3 train_imagenet.py --model=get_imagenet_tuned_model(False) --gpu 1