Single-Path-One-Shot-NAS

This repo provides a Pytorch-based implementation of SPOS(Single Path One-Shot Neural Architecture Search with Uniform Sampling) by Zichao Guo, and et. al. However, this repo only contains 'Block Search' and it's very time consuming to train this network on ImageNet so I haven't got the final result yet. I will update the repo soon and greatly thanks to Zichao Guo for his advice on some details.
Yet, there are still some differences with the official version such as data preprocessing and some hyper parameters.

Progress

The problem that the loss doesn't decrease has been solved and it's easy to train on CIFAR-10 by only a single GPU.I am trying to make it to train on multi-GPUs on ImageNet.

Environments

Python == 3.6.8, Pytorch == 1.1.0, CUDA == 9.0.176, cuDNN == 7.3.0, GPU == Single GTX 1080Ti 

Dataset

SPOS directly can train on ImageNet.ImageNet needs to be manually downloaded and here are some instructions.

Usage

python train.py --train_dir YOUR_TRAINDATASET_PATH --val_dir YOUR_VALDATASET_PATH

To Do

• Block Search
• Train on CIFAR-10
• Channel Search
• Evolutionary Algorithm

Reference

[1]: Differentiable architecture search for convolutional and recurrent networks

Citation

@article{guo2019single,
        title={Single path one-shot neural architecture search with uniform sampling},
        author={Guo, Zichao and Zhang, Xiangyu and Mu, Haoyuan and Heng, Wen and Liu, Zechun and Wei, Yichen and Sun, Jian},
        journal={arXiv preprint arXiv:1904.00420},
        year={2019}
}