README.md

DARTS: Differentiable Architecture Search

Liu, Hanxiao, Karen Simonyan, and Yiming Yang. "Darts: Differentiable architecture search." arXiv preprint arXiv:1806.09055 (2018). [arxiv]

Requirements

• python 3
• pytorch >= 0.4
• graphviz
  • First install using apt install and then pip install.
  • or conda install may make it work.
• numpy
• tensorboardX

Results

Dataset	Final validation acc	Best validation acc
MNIST	99.75%	99.80%
Fashion-MNIST	99.20%	99.31%
CIFAR-10	97.17%	97.23%

97.17%, final validation accuracy in CIFAR-10, is the same number as the paper.

Architecture progress

CIFAR-10

MNIST

Fashion-MNIST

Plots

Search-training phase of Fashion-MNIST

Augment-validation phase of CIFAR-10 and Fashion-MNIST

Reference

https://github.com/quark0/darts (official implementation)

Main differences to reference code

• Supporting pytorch >= 0.4
• Code that is easy to read and commented.
• Implemenation of architect
  • Original implementation is very slow in pytorch >= 0.4.
• Various dataset
• Tensorboard
• No RNN

and so on.