aisquared 0.2.0.dev1

Utilities for interacting with the AI Squared Technology Stack

AISquared

This package contains utilities to interact with the AI Squared technology stack, particularly with developing and deploying models to the AI Squared Browser Extension or other applications developed through the AI Squared JavaScript SDK.

Installation

This package is available through Pypi and can be installed by running the following command:

pip install aisquared

Alternatively, the latest version of the software can be installed directly from GitHub using the following command

pip install git+https://github.com/AISquaredInc/aisquared

Capabilities

This package is currently in a state of constant development, so it is likely that breaking changes can be made at any time. We will work diligently to document changes and make stable releases in the future.

The aisquared package currently contains five subpackages, the aisquared.config package, the aisquared.base subpackage, the aisquared.logging subpackage, the aisquared.serving subpackage, and the aisquared.remote package. The config package holds objects for building the configuration files that need to be included with converted model files for use within the AI Squared Extension. The contents of the config subpackage contain both pre- and postprocessing steps as well as harvesting, analytic, rendering, and feedback objects to use with the model. The following will explain the functionality of the config package:

aisquared.config

The aisquared.config subpackage contains the following objects:

• ModelConfiguration
  • The ModelConfiguration object is the final object to be used to create the configuration file. It takes as input a list of harvesting steps, list of preprocessing steps, a list of analytics, a list of postprocessing steps, a list of rendering steps, an optional MLFlow URI, an optional MLFlow user, and an optional MLFlow token

aisquared.config.harvesting

The aisquared.config.harvesting subpackage contains the following objects:

• ImageHarvester
  • The ImageHarvester class indicates the harvesting of images within the DOM to perform prediction on
• TextHarvester
  • The TextHarvester class indicates the harvesting of text within the DOM to perform prediction on

aisquared.config.preprocessing

The aisquared.config.preprocessing subpackage contains the following objects:

• ImagePreprocessor
  • The ImagePreprocessor class takes in preprocessing steps (defined below) which define preprocessing steps for images.
• TabularPreprocessor
  • The TabularPreprocessor class takes in preprocessing steps (defined below) which define preprocessing steps for tabular data.
• TextPreprocessor
  • The TextPreprocessor class takes in preprocessing steps (defined below) which define preprocessing steps for text data.

aisquared.config.analytic

The aisquared.config.analytic subpackage contains the following objects:

• LocalAnalytic
  • The LocalAnalytic class indicates the use of an analytic or lookup table from a local file
• LocalModel
  • The LocalModel class indicates the use of a model from a local file
• DeployedAnalytic
  • The DeployedAnalytic class indicates the use of an analytic or lookup table from a remote resource
• DeployedModel
  • The DeployedModel class indicates the use of a model deployed to a remote resource

aisquared.config.postprocessing

The aisquared.config.postprocessing subpackage contains the following objects:

• Regression
  • The Regression object is a postprocessing class for models which perform regression. Since it is common to train regression models by scaling regression outputs to values between 0 and 1, this class is designed to convert output values between 0 and 1 to their original values, corresponding to min and max when the class is instantiated.
• BinaryClassification
  • The BinaryClassification object is a postprocessing class for models which perform binary classification. The class is instantiated with a label map and a cutoff value used to identify when the positive class (class 1) is identified.
• MulticlassClassification
  • The MulticlassClassification object is a postprocessing class for models which perform multiclass classification. The class is instantiated with a label map only.
• ObjectDetection
  • The ObjectDetection object is a postprocessing class for models which perform object detection. The class is instantiated with a label map and a cutoff value for identification.

aisquared.config.rendering

The aisquared.config.rendering subpackage contains the following objects:

• ImageRendering
  • The ImageRendering object is a rendering class for rendering single predictions on images.
• ObjectRendering
  • The ObjectRendering object is a rendering class for rendering object detection predictions on images.
• WordRendering
  • The WordRendering object is a rendering class for rendering highlights, underlines, or badges on individual words.
• DocumentRendering
  • The DocumentRendering object is a rendering class for rendering document predictions.

aisquared.config.feedback

The aisquared.config.feedback subpackage contains the following objects:

• SimpleFeedback
  • The SimpleFeedback object is a feedback object for simple thumbs up/thumbs down for predictions
• BinaryFeedback
  • The BinaryFeedback object is a feedback object for binary classification use cases
• MulticlassFeedback
  • The MulticlassFeedback object is a feedback object for multiclass classification use cases
• RegressionFeedback
  • The RegressionFeedback object is a feedback object for regression use cases
• ModelFeedback
  • The ModelFeedback object is a feedback object for configuring feedback for the model directly, rather than its predictions
• QualitativeFeedback
  • The QualitativeFeedback object is a feedback object for configuring questions asked about each individual prediction the model makes

Preprocessing Steps

The aisquared.config.preprocessing subpackage contains PreProcStep objects, which are then fed into the ImagePreprocessor, TabularPreprocessor, and TextPreprocessor classes. The PreProcStep classes are:

• ZScore
  • This class configures standard normalization procedures for tabular data
• MinMax
  • This class configures Min-Max scaling procedures for tabular data
• OneHot
  • This class configures One Hot encoding for columns of tabular data
• AddValue
  • This class configures adding values to pixels in image data
• SubtractValue
  • This class configures subtracting values to pixels in image data
• MultiplyValue
  • This class configures multiplying pixel values by a value in image data
• DivideValue
  • This class configures dividing pixel values by a value in image data
• ConvertToColor
  • This class configures converting images to the specified color scheme
• Resize
  • This class configures image resize procedures
• Tokenize
  • This class configures how text will be tokenized
• RemoveCharacters
  • This class configures which characters should be removed from text
• ConvertToCase
  • This class configures which case - upper or lower - text should be converted to
• ConvertToVocabulary
  • This class configures how text tokens should be converted to vocabulary integers
• PadSequences
  • This class configures how padding should occur given a sequence of text tokens converted to a sequence of integers

These step objects can then be placed within the TabularPreprocessor, ImagePreprocessor, or TextPreprocessor objects. For the TabularPreprocessor, the ZScore, MinMax, and OneHot Steps are supported. For the ImagePreprocessor, the AddValue, SubtractValue, MultiplyValue, DivideValue, ConvertToColor, and Resize Steps are supported. For the TextPreprocessor, the Tokenize, RemoveCharacters, ConvertToCase, ConvertToVocabulary, and PadSequences Steps are supported

Final Configuration and Model Creation

Once harvesting, preprocessing, analytic, postprocessing, and rendering objects have been created, these objects can then be passed to the aisquared.config.ModelConfiguration class. This class utilizes the objects passed to it to build the entire model configuration automatically.

Once the ModelConfiguration object has been created with the required parameters, the .compile() method can be used to create a file with the .air extension that can be loaded into an application which utilizes the AI Squared JavaScript SDK.

aisquared.base

The aisquared.base subpackage contains two classes, the DocumentPredictor and the ImagePredictor classes, which streamline document prediction and image prediction using locally-saved models. These classes abstract away the steps required in the ModelConfiguration class. However, just like the ModelConfiguration class, objects of these classes support the .compile() method; using this method creates the .air file as well.

aisquared.remote

The aisquared.remote subpackage contains utilities and classes for interacting with cloud-based resources for deploying and managing models and results. Currently, we have the following client objects:

• AWSClient
  • This client facilitates the interaction with AWS cloud storage
• AzureClient
  • This client facilitates the interaction with Azure cloud storage

aisquared.serving

The aisquared.serving subpackage contains utilities for serving models locally or remotely using MLflow or locally using Flask.

aisquared.logging

The aisquared.logging subpackage is powered by MLflow, a powerful open-source platform for the machine learning lifecycle. The logging subpackage inherits nearly all functionality from mlflow, so we highly recommend users refer to the MLflow documentation site for additional information.

In this subpackage, we have additionally added implementations of individual functions to save TensorFlow, Keras, Scikit-Learn, and PyTorch models in a format that can be deployed quickly using MLflow.

Changes

Below are a list of additional features, bug fixes, and other changes made for each version.

Version 0.1.3

• Added flags parameter to TextHarvester using regular expression harvesting
• Deleted model_feedback parameter in ModelConfiguration object and included functionality in feedback_steps parameter
• Changed format parameter to header for both deployed analytics
• Added feedback and stages to DocumentPredictor and ImagePredictor objects
• Non-API changes for ALLOWED_STAGES
• Fixed bugs preventing Windows users from importing the package
• Updated ModelConfiguration to include url parameter
• Changed default tokenization string

Version 0.2.0

• Moved preprocessing steps under subpackages for specific kinds of preprocessing steps
• Cleaned up documentation to render within programmatic access environments
• Added aisquared.logging subpackage
• Created InputHarvester
• Created the aisquared.serving subpackage