1.NNI-介绍
• NNI(NeuralNetwork Intelligence):An open source AutoMLtoolkit for neural architecture search and hyper-parameter tuning.
• https://github.com/microsoft/nni
• https://nni.readthedocs.io/en/latest/
2.NNI-安装
• Install& Verify
• 1.python -m pip install --upgrade nni
• 2.git clone -b v1.1 https://github.com/Microsoft/nni.git
下面3-4可选，也可以直接pip install nni
•3. cd nni
•4. powershell -ExecutionPolicyBypass -file install.ps1
3.NNI-mnist实例
• Run the MNIST example
• nnictl create --config
nni\examples\trials\mnist\config_windows.yml
the result as follows:
you can view the result though access to http://127.0.0.1:8080 on local system
4.NNI-WEBUI可以查看训练过程
5.NNI architecture & KeyConcepts
Key Concepts
Experiment : Anexperiment is one task of, for example, finding out the best hyperparameters of amodel, finding out the best neural network architecture. It consists of trialsand AutoMLalgorithms.
Search Space : Itmeans the feasible region for tuning the model. For example, the value range ofeach hyperparameters.
Configuration : Aconfiguration is an instance from the search space, that is, each hyperparameter hasa specific value.
Trial : Trial is an individual attempt atapplying a new configuration (e.g., a set of hyperparametervalues, a specific nerualarchitecture). Trial code should be able to run with the providedconfiguration.
Tuner : Tuner is an AutoMLalgorithm, which generates a new configuration for the next try. A new trialwill run with this configuration.
Assessor : Assessor analyzes trial’s intermediate results (e.g., periodically evaluatedaccuracy on test dataset) to tell whether this trial can be early stopped ornot.
Training Platform : Itmeans where trials are executed. Depending on your experiment’s configuration,it could be your local machine, or remote servers, or large-scale trainingplatform (e.g., OpenPAI,Kubernetes).
下面分别介绍NNI的两种使用方式API和anotation
6.NNI API
• Step 1 - Prepare a Search Space
parameters file
Reference:https://github.com/microsoft/nni/blob/master/docs/en\_US/Tutorial/SearchSpaceSpec.md
• Step 2 - Update model codes
• import nni
• Get configuration from Tuner
RECEIVED_PARAMS = nni.get_next_parameter()
RECEIVED\_PARAMS is an object, for example:
{"conv\_size": 2, "hidden\_size":124, "learning\_rate":0.0307, "dropout\_rate":0.2029}
• Reportmetric data periodically (optional)
nni.report_intermediate_result(metrics)
This metrics is reported to assessor .
Usually, metrics could be periodically evaluated loss or accuracy
• Report performance of the configuration
nni.report_final_result(metrics)
This metrics is reported to tuner .
Reference:https://nni.readthedocs.io/en/latest/sdk\_reference.html
• Step 3 - Enable NNI API
• useAnnotation:false
• searchSpacePath:/path/to/your/search_space.json
• Reference:https://github.com/microsoft/nni/blob/master/docs/en\_US/Tutorial/ExperimentConfig.md
7.NNI Python Annotation
• Step 1 - Update codes with annotations
• @nni.variable willtake effect on its following line, which is an assignment statement whose leftvalue mustbe specified by the keyword name in @nni.variable
• Step 2 - Enable NNI Annotation
• useAnnotation:true
• Reference:https://github.com/microsoft/nni/blob/master/docs/en\_US/Tutorial/AnnotationSpec.md