Skip to content

Latest commit

 

History

History
256 lines (193 loc) · 17.6 KB

README.md

File metadata and controls

256 lines (193 loc) · 17.6 KB

@awarns/ml-kit

npm (scoped) npm

This module allows to execute TensorFlow Lite machine learning models embedded in the mobile device, which is useful to make predictions based on some input data.

Supported models architectures are Convolutional Neural Networks (CNN) and Multilayer Perceptrons (MLP), both for classification and regression problems.

Install the plugin using the following command line instruction:

ns plugin add @awarns/ml-kit

Usage

After installing this plugin, you will have access to an API and two tasks to perform classification or regression on the provided input data. But before using the plugin, you must meet the following requirements regarding the machine learning model:

  • Have a TensorFlow Lite machine learning model (*.tflite) of the supported architectures (CNN or MLP). Classification models must have metadata (i.e., name, version, author, etc...) and an associated labels file with each label in a row of the file. While regression models don't have to include an associated labels file, it's recommended to add the metadata with the information of the model.
  • Place your TensorFlow Lite models in a folder named ml-models inside your app's src folder (i.e., same level as {app|main}.ts).
    • The model file name must follow the next format: {model_name}-{cnn|mlp}-[version].tflite. The file name must contain a name (model_name), the model's architecture (cnn or mlp) and, optionally, the model's version (version). The file name elements must be splitted by a dash (-).

API

In order to do a regression or a classification, first you have to load a machine learning model. You can do that using the getModel(...) method provided by the ModelManager. It also provides the listModels method, useful known which models are available in the device.

When you load a model using the getModel(...) method, you obtain a BaseModel or a ClassificationModel to perform a regression or a classification, respectively. To do so, you can create a Regressor using a BaseModel, and a Classifier using a ClassificationModel. Finally, to perform the prediction, you just have to call to the predict method, which will return a RegressionResult or a ClassificationResult, depending on which predictor has been used.

Here's a complete example:

import {
  BaseModel,
  ClassificationModel,
  ClassificationResult,
  Classifier,
  DelegateType,
  getModelManager,
  InputData,
  ModelType, RegressionResult, Regressor
} from '@awarns/ml-kit';

async function doClassification(inputData: InputData /* number[] */) {
  const model: ClassificationModel = await getModelManager().getModel(
    'activity_classifier-cnn',
    ModelType.CLASSIFICATION,
    { acceleration: DelegateType.GPU } // Use GPU, if available.
  );

  const classifier = new Classifier(model);
  const result: ClassificationResult = classifier.predict(inputData);
}

async function doRegression(inputData: InputData /* number[] */) {
  const model: BaseModel = await getModelManager().getModel(
    'stress_regressor-cnn',
    ModelType.REGRESSION,
    { acceleration: 4 } // Use 4 threads.
  );

  const regressor = new Regressor(model);
  const result: RegressionResult = regressor.predict(inputData);
}

Note: the RegressionResult and ClassificationResult are not framework records. If you want to introduce these results into the framework for example, to persist them using the persistence package, you have to manually create a Regression and Classification records.

ModelManager

You can obtain the singleton instance of the ModelManager calling the getModelManager() function.

Method Return type Description
listModels() Promise<Model[]> Returns a list of the models that are available for their use.
getModel(modelName: string, modelType: ModelType, modelOptions?: ModelOptions) Promise<BaseModel|ClassificationModel> Retrieves and loads the specified model, ready to be used.
Model
Property Type Description
modelInfo ModelInfo Contains model's metadata.
ModelInfo
Property Type Description
id string Identifier of the model, generally the name of its file.
name string Name info included in the model's metadata.
architecture ModelArchitecture Architecture of the model, i.e., CNN or MLP.
version string Version info included in the model's metadata.
author string Author info included in the model's metadata.
ModelType
Value Description
REGRESSION A model that performs a regression.
CLASSIFICATION A model that performs a classification.
ModelOptions
Property Type Description
acceleration DelegateType | number Which type of acceleration to use when running the model. It can take the values DelegateType.GPU (GPU acceleration), DelegateType.NNAPI (Android Neural Networks API acceleration) or a number indicating the quantity of threads to use.

Regressor

Method Return type Description
predict(inputData: InputData) RegressionResult Preforms a regression using the provided data.

Classifier

Method Return type Description
predict(inputData: InputData) ClassificationResult Preforms a classification using the provided data.

Tasks

Task name Description
{classificationAim}Classification{tag?} It performs a classification using the input data contained on the invocation event's payload. classificationAim is used to differentiate among classification tasks. An optional tag can be added to the task name.
{regressionAim}Regression{tag?} It performs a regression using the input data contained on the invocation event's payload. regressionAim is used to differentiate among regression tasks. An optional tag can be added to the task name.

Note: the input data provided through the invocation event's payload must be an array of numbers ready to be feed into the model. In other words, the main application is the one in charge of executing the required data preprocessing techniques (e.g., normalization, feature extraction, etc...) to prepare the data for the model, not this module.

To register these tasks for their use, you just need to import them and call their generator functions inside your application's task list:

import { Task } from '@awarns/core/tasks';
import {
  classificationTask,
  regressionTask,
  DelegateType,
} from '@awarns/ml-kit';
import { DelegateType } from './index';

export const demoTasks: Array<Task> = [
  classificationTask('human-activity', 'activity_classifier-mlp', '', { acceleration: 4 }),
  // humanActivityClassification

  classificationTask('human-activity', 'activity_classifier-cnn', 'UsingCNN', { acceleration: DelegateType.GPU }),
  // humanActivityClassificationUsingCNN
  
  regressionTask('stress-level', 'stress_regressor-cnn'),
  // stressLevelRegression
]

Task generator parameters:

Parameter name Type Description
{classification|regression}Aim string Objective of the classification/regression. Used to name the task.
modelName string | ModelNameResolver Name of the model (without tflite extension) stored in the ml-models folder to use for this task, or a function that returns the name of the model when called.
tag (Optional) string Adds a tag to the name of the task to differentiate it from other tasks with other configurations.
modelOptions (Optional) ModelOptions | ModelOptionsResolver Configuration to use with the model or a function that returns the configuration when called.

Note: It's highly recommended to provide the {classification|regression}Aim in snake-case format. This string will be used as the type of the output record of the task. All the records have their type in snake-case, so providing the {classification|regression}Aim in snake-case will keep the consistency of the framework.

ModelNameResolver: () => string

Useful to change the model used by a task at runtime. You can use the ModelManager to obtain a list with the models that are available in the device.

ModelOptionsResolver: () => ModelOptions

Useful to change the options used by the model of a task at runtime.

Tasks output events:

Example usage in the application task graph:

on('inputDataForHumanActivity', run('humanActivityClassificationUsingCNN'));
on('humanActivityPredicted', run('writeRecords'));

on('inputDataStressLevel', run('stressLevelRegression'));
on('stressLevelPredicted', run('writeRecords'));

Note: To use the writeRecords task, the persistence package must be installed and configured. See persistence package docs.

Events

Name Payload Description
{classificationAim}Predicted Classification Indicates that a classification has been completed.
{regressionAim}Predicted Regression Indicates that a regression has been completed.

Records

Classification

Property Type Description
id string Record's unique id.
type string Always {classificationAim}-prediction.
change Change Always NONE.
timestamp Date The local time when the model predicted the classification result.
classificationResult ClassificationResult Object containing the results of the classification.
ClassificationResult
Property Type Description
prediction ClassificationPrediction[] Array of the classification predictions generated by the model.
modelName string The name of the model used for the classification.
architecture string The architecture of the model used for the classification.
version string The version of the model used for the classification.
ClassificationPrediction
Property Type Description
label string Identifier of the class.
score number Score of the prediction.

Regression

Property Type Description
id string Record's unique id.
type string Always {regressionAim}-prediction.
change Change Always NONE.
timestamp Date The local time when the model predicted the regression result
regressionResult RegressionResult Object containing the results of the regression.
RegressionResult
Property Type Description
prediction number[] Array of numbers with the results generated by the model.
modelName string The name of the model used for the regression.
architecture string The architecture of the model used for the regression.
version string The version of the model used for the regression.

License

Apache License Version 2.0

Disclaimer

While we state that CNN models are supported, only 1D-CNN models have been tested. The code is general enough to support 2D and 3D CNN models with one input tensor, but they have not been tested. If you try 2D/3D-CNN models and something is not working as expected, contact us.