Skip to content

Latest commit

 

History

History
310 lines (207 loc) · 11 KB

README.rst

File metadata and controls

310 lines (207 loc) · 11 KB
Raw

scikit-nni

Documentation Status

Hyper parameters search for scikit-learn components using Microsoft NNI

Introduction

Microsoft NNI (Neural Network Intelligence) is a toolkit to help users run automated machine learning (AutoML) experiments. The tool dispatches and runs trial jobs generated by tuning algorithms to search the best neural architecture and/or hyper-parameters in different environments like local machine, remote servers and cloud.

Read and explore more about Microsoft NNI here - https://github.com/microsoft/nni

scikit-nni is a helper tool (and a package) that :

  • generates the configuration (config.yaml & search-space.json) required for NNI
  • automatically builds the scikit-learn pipelines based on your specification and becomes an experiment/trial code for Microsoft NNI to run.

What value does this tool add to Microsoft NNI ?

First note that this tool is specifically written to only help with scikit-learn pipelines and to tune classification algorithms. In near future, I would add the support for regression algorithms as well.

Now when you use Microsoft NNI you need to specify (at minimum) 3 files :

  • A search space (json) file that contains the parameters that you want to search/tune.
  • Your code/experiment. In your experiment code, you perform these tasks in sequence :
    • Request for the parameters from NNI server
    • Create your model using these parameters
    • Fit your model
    • Score your model
    • Report the score to NNI server.
  • a configuration file where you specify the tuner, which mode to use to run, path to your code file and search space file.

scikit-nni eliminiates the second step i.e. it builds the scikit pipelines, request NNI server for parameters and also report back the score of your model. It also simplifies (in IMHO) the input specification by only requiring one file instead of 3.

Sounds interesting ? Then read the documentation below, install scikit-nni, and more importantly provide feedback if it does not work for you and/or you think it can be improved.

Features

  • Hyperparameters search for scikit-learn pipelines using Microsoft NNI
  • No code required to define the pipelines
  • Built-in datasource reader for reading npz files for classification
  • Support for using custom datasource reader
  • Single configuration file to define NNI configuration and search space

I plan to add more datasource readers (e.g. CSV, libSVM format files etc). Contributions are always welcome !

Usage

https://github.com/ksachdeva/scikit-nni/raw/master/images/demo.gif

Step 1 - Write a specification file

The specification file is essentially a YAML file but with extension .nni.yml

There are 4 parts (sections) in the configuration file.

Datasource Section

This is where you will specify the (python) callable that sknni would be invoking to get the training and test dataset.

The callable must return two values where each value is a tuple of two items. The first tuple consists of training data (X_train, y_train) and the second tuple consists of test data (X_test, y_test).

An example callable would look like this:

import numpy as np

from sklearn.datasets import load_digits
from sklearn.model_selection import train_test_split

class ACustomDataSource(object):
    def __init__(self):
        pass

    def __call__(self, test_size:float=0.25):
        digits = load_digits()
        X_train, X_test, y_train, y_test = train_test_split(digits.data, digits.target, random_state=99, test_size=test_size)

        return (X_train, y_train), (X_test, y_test)

In the above example, the callable generates the train and test dataset. The callable can even have paramaters for e.g. in this example you could optionally pass the fraction of data to be used for testing.

Now let's see how you would add it in the specification file.

# Datasource is how you specify which callable
# sknni will invoke to get the data
dataSource:
    reader: yourmodule.ACustomDataSource
    params:
        test_size: 0.30

Make sure that during the exeuction of the experiment your datasource (i.e. in this case yourmodule.ACustomDataSource) is available in the PYTHONPATH.

Here is an additional example showing the usage of a built-in datasource reader

dataSource:
    reader: sknni.datasource.NpzClassificationSource
    params:
        dir_path: /Users/ksachdeva/Desktop/Dev/myoss/scikit-nni/examples/data/multiclass-classification

NpzClassificationSource expects that at dir_path you have two folders - train and test. In each folder are the files named as 0.npz, 1.npz etc. Every file contains that features for that corresponding class.

The repository contains two such datasources to do binary and multiclass classifications.

Pipline definition Section

Below is an example of this type of section. You simply specify the list of steps of your scikit-learn Pipeline.

Note - The sequence of steps is very important.

What you MUST ensure is that the full qualified name of your scikit-learn preprocessors, transformers and estimators is correctly specified & spelled. sknni uses reflection and introspection to create the instances of these components so if you have a typo in the names and/or they are not available in your PYTHONPATH you will get an error at experiment execution time.

sklearnPipeline:
    name: normalizer_svc
    steps:
        normalizer:
            type: sklearn.preprocessing.Normalizer
            classArgs:
                norm: l2
        svc:
            type: sklearn.svm.SVC

In above example, there are 2 steps. The first step is to normalize the data and the second step is train a classifier using Support Vector Machine.

Search Space Section

This section corresponds to the search space for your hyperparameters. When you use `nnictrl` this is typically specified in search-space.json file.

See https://nni.readthedocs.io/en/latest/Tutorial/SearchSpaceSpec.html to learn more about the search space syntax.

Here are the important things to note about this section -

  • The syntax is the same (except we are using YAML here instead of JSON) for specifiying parameter types and ranges.
  • You MUST specifiy the parameters corresponding to the step in your scikit pipeline.
  • You MUST use the names of the parameters that are same as the ones accepted by the constructors of scikit-learn components (i.e. preprocessors, estimators etc).

Below is an example of this type of section.

nniConfigSearchSpace:
    - normalizer:
        norm:
            _type: choice
            _value: [l2, l1]
    - svc:
        C:
            _type: uniform
            _value: [0.1,0.0]
        kernel:
            _type: choice
            _value: [linear,rbf,poly,sigmoid]
        degree:
            _type: choice
            _value: [1,2,3,4]
        gamma:
            _type: uniform
            _value: [0.01,0.1]
        coef0:
            _type: uniform
            _value: [0.01,0.1]

Note that sklearn.svm.SVC takes C, kernel, degree, gamman and coef0 is the paramaters and hence we have used here the same names (keys) in the search space specification. You can add as many or as little parameters to search for.

NNI Config Section

This is the simplest of all sections as there is nothing new here from sknni perspective. You just copy-paste here your NNI's config.yaml here. You do not have to specify codedir and command field in the trial subsection as this is added by the sknni in the generated configuration files.

See https://nni.readthedocs.io/en/latest/Tutorial/ExperimentConfig.html

Here is an example of this type of section.

# This is exactly same as the one that of NNI
# except that you do not have to specify the command
# and code fields. They are automatically added by the sknni generator
nniConfig:
    authorName: default
    experimentName: example_sklearn-classification
    trialConcurrency: 1
    maxExecDuration: 1h
    maxTrialNum: 100
    trainingServicePlatform: local
    useAnnotation: false
    tuner:
        builtinTunerName: TPE
        classArgs:
            optimize_mode: maximize
    trial:
        gpuNum: 0

You can look at the various examples in the repository to learn how to define your own specification file.

Step 2 - Generate your experiment

sknni generate-experiment --spec example/basic_svc.nni.yml --output-dir experiments

Above command will create a directory experiments/svc-classification with the following files

  • The original specification file i.e. basic_svc.nni.yml (used during experiment run as well)
  • Generated Microsoft NNI's config.yml
  • Generated Microsoft NNI's search-space.json

Note - there is no python file as typically shown in the examples of Microsoft NNI as the command in ends up invoking sknni entry point when the experiment is run.

Step 3 - Run your experiment

This is same as running nnitctl

nnictl create --config experiments/svc-classification/config.yml

Troubleshooting

My trials are failing what is wrong ?

Your trial could fail for many reasons -

  • Bug in your DataSource code resulting the exception/error
  • Wrong inputs to your (or built-in) DataSources resulting in exception/error
  • Your DataSource (python callable) could not be found

Here is what I would recommend -

  • Test your DataSource code

  • The webui does not always display all the errors/logs so look at the log of your trials and more specifically stderr file

    cat $HOME/nni/experiments/<YOUR_EXPERIMENT_ID>/trials/<TRIAL_ID>/stderr

cat $HOME/nni/experiments/<YOUR_EXPERIMENT_ID>/trials/<TRIAL_ID>/trial.log

Credits

This package was created with Cookiecutter and the audreyr/cookiecutter-pypackage project template.