ACEncoding

Code accompanying the paper

Action Command Encoding for Surrogate Assisted Neural Architecture Search

ACEncoding is an encoding strategy of NAS. The proposed ACEncoding is inspired by the process of gene expression, where each architecture is constructed by a series of action commands and each action command is encoded by three integers.

Requirements

pip install -r requirements.txt

# Surrogate dependence

git clone https://github.com/anonymone/pytorch-seq2seq.git ./Res/seq2seq

cd ./Res/Seq2Seq/

python3 setup.py install

Neural Architecture Search

# Evolutionary Algorithm
sh search_EA.sh

# Reinforcement Learning
sh search_RL.sh

# Gradient Descent
sh search_GD.sh

# Random (Baseline)
sh search_RD.sh

Neural Architecture Validation

Your need to specify the paramenter $ENCODE = "Your ACEncoding String."

# CIFAR-10, CIFAR-100, Fashion-MNIST
# You need to specify the parameter $DATASET = ["CIFAR10", "CIFAR100", "Fashion-MNIST"]
sh valid_CIFAR.sh

# ImageNet
sh valid_ImageNet.sh

If your GPU server has more than one GPU. You can choose the GPU by changing $DEVICE = cuda:$ID. $ID is the ID of GPUs.

Visualization

In our framework, each neural architecture is saved as a visual DOT file.

Install Graphviz

# CentOS
sudo yum install graphviz

# Ubuntu
sudo apt install graphviz

# MacOS
brew install graphviz

Usage

# Command Line
dot -Tpng YOUR_DOT_FILE.dot -o OUTPUT.png

# We highly recommand to use VScode extension Graphviz Preview.
# And then, you can check your DOT file in a convenience way.

Acknowledgement

Code inspired and modified from NSGA-Net, NAO, MetaQNN, ENAS and IBM/Seq2Seq-torch.