torch-evolvenet

This is implementation of architectures in the paper entitled
‘Exploring Huge-space Global Architecture Search via Network Evolving Based on Fine-grained Convolutional Neural Network Modeling’

Environment

Hardware: 1 GPU (the memory should exceed 8G)
Packages: Python 3.6; Pytorch 0.4.1

Steps to run

Prepare the Dataset: Put the batch files in the certain path, torch-evolvenet/lib/ CIFAR10_dataset/.
Run any pleasant python script provided.

Explanations

There are two key points implemented by the codes

Generate fine-grained Convolutional Neural Networks (fgCNNs) with Directed Acyclic Graphs (DAGs)
Train the evolving fgCNNs which have dynamic architecture

Following figure shows the overview of the architecture.

Four scripts are provided

File name	Function
train_fgCNN_rand.py	Train and test a CNN which is generated by a random DAG. Topology of the DAG is fixed.
train_EvolveNet.py	Train and test the CNN from a randomly generated DAG. Topology of the DAG is dynamic. It changes with the training.
train_fgCNN_searched.py	Train and test a CNN which is generated by a searched DAG.
test_fgCNN_searched.py	Test the best model obtained in experiments of the noted paper on CIFAR-10 benchmark.

Interesting findings: the growing connections

Following figure shows the visualized dynamic topology of the DAG in two generations (correspond to run 1 of experiments in the paper)

As a nascent project, we are eager for interchanges and comments.
Watch this branch for more information about future possible progress!