Preparation

• Assuming you have three available computers to be ready for performing BenchENAS, and their configurations are shown below:

  ID	Type	IP address	Avaliable GPU	Environment
  1	Center Node	192.168.1.1	No avaliable GPUs	Ubuntu 18.04 python 3.6
  2	Compute Node	192.168.1.2	GPU#0, GPU#1	Ubuntu 18.04 python 3.6 torch 1.1 torchvision 0.3.0
  3	Compute Node	192.168.1.3	GPU#0, GPU#1, GPU#2

• The algorithm you are going to perform is AE-CNN from the below paper and other algorithms:

  Yanan Sun, Bing Xue, Mengjie Zhang, Gary G. Yen, Jiancheng Lv, "Automatically designing CNN architectures using genetic algorithm for image classification," IEEE Transactions on Cybernetics, vol. 50, no. 9, pp. 3840-3854, 2020. DOI:10.1109/TCYB.2020.2983860

• Make sure you have available SSH accounts to login the center node and computer nodes.

Step by step

• Copy the folder BenchENAS to your home dictionary on the center node computer.

• Environment Setup

  • On the center node:
    • Install sshpass

      sudo apt-get install sshpass

    • Install redis and start redis server

      sudo apt-get install redis-server
      redis-server --protected-mode on

    • Installation of python third-party libraries

      pip install redis
      pip install paramiko

  • On the compute node:
    • Install redis

    pip install redis

• Configuration of the algorithm and training parameters

  • Configure the parameters of the ENAS algorithm to be performed in the file global.ini:
    
    

    Section	Key	Type	Example	Description
    algorithm	name	string	aecnn_cifar10	The name of the output folder
    	run_algorithm	string	aecnn	The name of the algorithm
    	max_gen	positive interger	20	Max generation size
    	pop_size	positive interger	20	Population size
    redis	log_server	string	192.168.1.1	The ip of redis database
    	log_server_port	string	6379	The port of redis database

    
    
  • Configure optimizer and dataset in the file /train/train.ini:
    
    

    Section	Key	Type	Example	Description
    optimizer	_name	string	SGD	Class name of the optimizer used
    	_batch_size	positive interger	64	Batch size
    	_total_epoch	positive interger	500	Number of epochs
    LearningRate	lr	float	0.025	Learning rate
    	lr_strategy	string	CosineAnnealingLR	Learning rate strategy
    dataset	_name	string	CIFAR10	Name of the dataset used

    
    
  • Configure compute nodes and the available GPUs in file /compute/gpu.ini:
    
    

    Section	Key	Type	Example	Description
    192.168.1.2	worker_ip	string	192.168.1.2	The ip of the compute node
    	worker_name	string	cuda2	The name of the compute node
    	ssh_name	string	Bob	The name of user
    	ssh_password	string	123456	The password of user
    	gpu	string	0,1	List of available GPU ids
    192.168.1.3	worker_ip	string	192.168.1.3	The ip of the compute node
    	worker_name	string	cuda3	The name of the compute node
    	ssh_name	string	Bob	The name of user
    	ssh_password	string	123456	The password of user
    	gpu	string	0,1,2	List of available GPU ids

• Perform BenchENAS

  • Start the init_compute.py script to start the platform

    python init_compute.py

  • Start the algorithm you would like to perform

    python main.py

• Get the results

  The file result.txt under the folder runtime contains the fitness values for each individual.
  You can find the name and the fitness of the best individual from result.txt, which we assume is named indi_k_j (for the k-th individual of the j-th generation).
  And you can find the log information from the file indi_k_j.txt in the folder log, the architecture from the file indi_k_j.py from the folder scripts, and the encoding information from the file begin_k.txt in the folder populations.
  
  PS: All generated files are in the folder runtime as shown in Fig. 1. The meaning of each generated file is as follows:
  • aecnn_cifar10: The output folder of the algorithm whose name is set to example in the configuration.
    
    • log: Training log for each individual.
    • populations: Cache files, result files, and saved population files.
    • scripts: The python script file generated for each individual.
  • compute.log: The log file of BenchENAS.
  • PlatENAS.db: The SQL database generated by BenchENAS.
  
  

  Fig. 1. Example of the runtime folder.