nni-hyperband

HyperParameter Tuning HyperBand with NNI

This example shows how to use NNI to perform hyperparameter tuning with HyperBand on Azure Machine Learning.

HPO with NNI

Neural Network Intelligence (NNI) is a library that provides a unified interface for hyperparameter optimization. Many tuning algorithm is included. See the following link for more details in reference section.

Prerequisites

• Azure Machine Learning Workspace
  • Compute Clusters for parallel training
  • Compute Instance with Azure ML CLI 2.0 and NNI library installed

Getting Started

1. Create conda environment

conda create -n nni python=3.6
conda init bash
source ~/.bashrc
conda activate nni

2. Install NNI library in your Compute Instance.

pip install nni==2.5 azureml-sdk==1.35.0

3. Create Compute Clusters in your Azure Machine Learning Workspace

4. Create a train script, a job configuration file and a search space configuration file.

• train script : train.py

• job configuration file : config_hyperband.yml

  configure TrainingService to Azure Machine Learning.

  TrainingService:
    platform: aml
    dockerImage: msranni/nni  # modify this if you bring your own docker image
    subscriptionId: <your subscription ID>
    resourceGroup: <azure machine learning workspace resource group>
    workspaceName: <azure machine learning workspace name>
    computeTarget: <compute cluster name>

• search space configuration file : search_space.json

  define your search space of hyperparameters in json file.

  {
      "dropout_rate":{"_type":"uniform","_value":[0.5,0.9]},
      "conv_size":{"_type":"choice","_value":[2,3,5,7]},
      "hidden_size":{"_type":"choice","_value":[124, 512, 1024]},
      "batch_size": {"_type":"choice","_value":[8, 16, 32, 64]},
      "learning_rate":{"_type":"choice","_value":[0.0001, 0.001, 0.01, 0.1]}
  }

5. Start trial job.

nnictl create --config config_hyperband.yml --port 8088

6. Access to dashboard.

• NNI Dashboard is running on Compute Instance. You can access to it from your client PC from URL like https://<your compute instance name>-8088.<region>instances.azureml.ms. For examples, if your Compute Instance name is "client" and region is "japaneast", you can access using https://client-8088.japaneast.instances.azureml.ms.

Reference

• Neural Network Intelligence (NNI)