Definition

An AI challenge is a specification of a problem that should be solved employing AI applications. It defines the functional and non-functional specifications as well as the input/output inferfaces. The challenge artifact is stored in a file challenge.yml that contains all the information on the challenge.

This schema of the challenge.yml is described below.

Schema for Bonseyes Challenges

type	object
properties
metadata	Metadata for AI Artifacts
specification	Challenge specification
	A challenge specification is built from three objects. interface are dynamically validated against an external schema specifying the required parameters for this challenge. training_data and sample_data are required by all challenges.
	type	object
	properties
	interface	Defines a type and references and external schema that has to be dynamically validated the contents are then inside the value object that is generic and needs to be filled according to the referenced schema See types/image-classification.yml for an example of such an external schema Automated dynamic loading and validation of unknown external schemas is not supported by json-schema at this point.
		type	object
		properties
		id	ID that uniquely identify the AI app interface.
			type	string
		parameters	type	object
		additionalProperties	False
	training_data	Dataset used to train the ai application
		data.yml
	sample_data	Small dataset that can be use to test the AI app
		data.yml
	quantization_data	Dataset that can be use to quantize model after training
		data.yml
	additionalProperties	False
deployment	Deployment specification
	type	object
	properties
	platform	ID of the platform
		type	string
	tool	type	object
	additionalProperties	False
model	Path to a pre-trained model
	type	string
evaluation_protocol	Describes the evaluation procedure to execute and the criteria to consider the evaluation procedure successful.
	type	object
	additionalProperties	type	object
		properties
		tool	type	object
			properties
			id	ID of the evaluation tool to use for the evaluation
				type	string
			parameters	Parameters for the evaluation tool
				type	object
			additionalProperties	False
		data	type	object
			additionalProperties	data.yml
		evaluation_data	data.yml
		constraints	type	array
			items	type	object
				properties
				metric	Name of the metric on which the constraint apply (as generated by the evaluation protocol).
					type	string
				condition	The condition to be satisfied, currently the following are supported.
					type	string
					enum	smaller_than, larger_than, equal_to
				value	Value of the metric
					type	number
				additionalProperties	False
		additionalProperties	False
additionalProperties	False

A minimal algorithm configuration looks as follows:

metadata:
  title: <name> Challenge
  description: |
    Description

specification:
  interface:
    id: com_bonseyes/interfaces#<challenge_id>
    parameters:
      classes: classes.txt

Create an AI Challenge

This guide assumes you already have the done the following:

Setup the local environment as explained in Prerequisites

A template of the challenge YML file is available in the AI Asset Container Generator.

The workflow of creating an AI Challenge is as follows:

Creating the challenge metadata
Defining the challenge interface

The following steps can be skipped for the moment:

Defining the challenge training and sample data (Deprecated),
Defining the evaluation procedure (Deprecated)

Creating the challenge metadata

Once you have created a challenge.yml based on the template, you can start implementing the challenge itself. The first part of the challenge that needs to be completed is the metadata. The minimum metadata required for a challenge is the following:

metadata:
  title: Title of the Challenge
  description: |
    Description of the challenge

The description of the challenge can be a multiline string, just make sure the indentation is consistent. It is recommended to use a YAML editor (for instance the one included in PyCharm) to make sure you don’t make syntax mistakes.

Defining the challenge interface

Once the metadata for the challenge is defined you need to define the interface that is expected to be implemented by the resulting AI App. The interface of an AI app defines the inputs and outputs.

The Bonseyes eco-system provides a number of pre-defined parametrized interfaces that are supported by the deployment tool and address common use cases. These interfaces can be used without having to develop any custom embedded code. Other interfaces can be defined if necessary, but require to write a corresponding extension for the deployment tool.

An interface artifact is defined in a file interface.yml and available in the deployment packages under ${deploymentPackage}/tools/ai_app_generator/share/interfaces. The schema of this file is described below:

Bonseyes Interface Description

type	object
properties
metadata	Deployment metadata
	Metadata for AI Artifacts
parameters	Path to schema for the AI Class parameters in the challenge
	type	string
results	Path to schema for the JSON serialization of the results
	type	string
ground_truth	Path to schema for the JSON serialization of the ground truth used to train this class
	type	string
http_api	Path to Swagger file of the HTTP API for the AI class
	type	string

The challenge defines the interface of the AI app by adding the following section:

specification:
  interface:
    id: name of the interface
    parameters: parameters of the interface

Default Interfaces

The following interfaces are distributed by default with bonseyes:

Image Classification (id: com_bonseyes/interfaces#image_classification): given an image provide the probability of the image to belong to one of a set classes. Find more information here: Image Classification.
Object Detection (id: com_bonseyes/interfaces#object_detection): given an image find the bounding boxes of a set of classes of objects and optionally for each object provide the position of landmarks of the object. Find more information here: Object Detection.
Face Recognition (id: com_bonseyes/interfaces#face_recognition): given an image of a face create a fingerprint that can be used to identify the face. Find more information here: Face Recognition.
Image Segmentation: (id: com_bonseyes/interfaces#image_segmentation): given an image compute a segmentation mask and confidence of the detected objects classes. Find more information here: Image Segmentation.
Audio Classification (id: com_bonseyes/interfaces#audio_classification): given an audio snippet provide the probability of the snipped belong to one of a set classes of audio snippets (e.g. keywords). Find more information here: Audio Classification.
Signal classificaiton (id: com_bonseyes/interfaces#signal_classification): given a signal stream compute the confidence of each possible signal class. Find more information here: Signal Classification.

Depending on the type of AI-app that is selected different parameters need to be provided in the challenge file. In the next subsections we describe the details for each type.

Image classification

An image classification challenge aims at providing the probability of the input image to belong to one of a set classes.

A minimal challenge.yml file for an image classification challenge looks as follows:

metadata:
  title: My Sample Challenge
  description: |
    This challenge is a sample challenge

specification:
  interface:
    id: com_bonseyes/interfaces#image_classification
    parameters:
      classes: classes.txt

The specification.interface.parameters.classes parameter must name a path to a text file (relative to the position of the challenge.yml file) that contains on each line a label for a class. The text file is expected to be encoded in UTF-8 and use UNIX line endings.

Object detection

An object detection challenge aims at finding the bounding boxes of a set of classes of objects, and optionally for each object, provide the position of landmarks of the object.

A minimal challenge.yml file for an object detection challenge looks as follows:

metadata:
  title: My Sample Challenge
  description: |
    This challenge is a sample challenge

specification:
  interface:
    id: com_bonseyes/interfaces#object_detection
    parameters:
      classes: classes.txt
      landmarks: landmarks.txt

The specification.interface.parameters.classes parameter must name a path to a text file (relative to the position of the challenge.yml file) that contains on each line a label for a class. The text file is expected to be encoded in UTF-8 and use UNIX line endings.

The specification.interface.parameters.landmarks parameter is optional and if present must name a path to a text file (relative to the position of the challenge.yml file) that contains on each line a label for a class. The text file is expected to be encoded in UTF-8 and use UNIX line endings.

Face recognition

A face recognition challenge aims at creating a fingerprint that can be used to identify the face given an input image.

A minimal challenge.yml file for a face recognition challenge looks as follows:

metadata:
  title: My Sample Challenge
  description: |
    This challenge is a sample challenge

specification:
  interface:
    id: com_bonseyes/interfaces#face_recognition

Image segmentation

An image segmentation challenge aims at identifying a mask to segment each pixel of an input image based on the class they belong to.

A minimal challenge.yml file for a image segmentation challenge looks as follows:

metadata:
  title: My Sample Challenge
  description: |
    This challenge is a sample challenge

specification:
  interface:
    id: com_bonseyes/interfaces#image_segmentation
    parameters:
      classes: classes.txt

The specification.interface.parameters.classes parameter must name a path to a text file (relative to the position of the challenge.yml file) that contains on each line a label for a class. The text file is expected to be encoded in UTF-8 and use UNIX line endings.

Audio classification

An audio classification challenge aims at providing the probability of the audio snippet to belong to one of a set classes.

A minimal challenge.yml file for an audio classification challenge looks as follows:

metadata:
  title: My Sample Challenge
  description: |
    This challenge is a sample challenge

specification:
  interface:
    id: com_bonseyes/interfaces#audio_classification
    parameters:
      classes: classes.txt

The specification.interface.parameters.classes parameter must name a path to a text file (relative to the position of the challenge.yml file) that contains on each line a label for a class. The text file is expected to be encoded in UTF-8 and use UNIX line endings.

Signal classification

An signal classification challenge aims at providing the probability of the signal stream to belong to one of a set classes.

A minimal challenge.yml file for an signal classification challenge looks as follows:

metadata:
  title: My Sample Challenge
  description: |
    This challenge is a sample challenge

specification:
  interface:
    id: com_bonseyes/interfaces#signal_classification
    parameters:
      classes: classes.txt

The specification.interface.parameters.classes parameter must name a path to a text file (relative to the position of the challenge.yml file) that contains on each line a label for a class. The text file is expected to be encoded in UTF-8 and use UNIX line endings.

Defining the challenge training and sample data (Deprecated)

In its simples form training data can be attached to the challenge by adding the following sections to the challenge:

specification:
   interface: ...
   training_data:
     sources:
       - title: title of the first dataset
         url: url of the dataset
         description: |
           Multiline description of the first dataset
       - title: title of the another dataset
         url: url of the other dataset
         description: |
           Multiline description of the other dataset
   sample_data:
     sources:
       - title: title of the dataset
         url: url of the dataset
         description: |
           Multiline description of the dataset

Both training and sample data may have multiple sources. If the data is distributed with the challenge (that is in the same package), you can put a relative path to the challenge.yml file as the url of the dataset. Data described in this form needs to be manually downloaded by the challenge user.

For each data item it is possible to include data tools that allow the challenge user to automatically download and transform the data. Data tools can be created using the process described in /pages/dev_guides/data_tool/create_data_tool.

The data tools are reference as follows:

sample_data:
  tool:
    id: com_example/mydatatool#data_tool.yml
    parameters: optional parameters
    input_data: optional path to a directory with input data for the data tool
  sources:
    - title: title of the dataset
      url: url of the dataset
      description: |
        Multiline description of the dataset

The id property of the data tool is formed by the name of the package containing the data tool metadata and the path to the file name. The parameters is optional and is an object compliant with the parameters schema provided by the data tool. The input_data property is required only by data tools that require it in their metadata, it points to a directory containing the input data for the data tool.

Defining the evaluation procedure (Deprecated)

The last part of the challenge that can be defined is the evaluation procedure. This section defines the evaluation tool to be used to check the performance of the AI app deployed on the target hardware. The evaluation tool is capable of generating a number of metrics, this section allows to specify constraints on these metrics that must be met by the AI apps.

Most of the time the evaluation procedure requires some data to be carried out, the so called evaluation data. This section allows to define the evaluation data as well. Since the evaluation tool expects data in a pre-defined format the section allows also to define a data tool capable of transforming the evaluation data in this target format. This allows to support arbitrary input data formats. For more information about how to specify data you can check the section Signal classification, the same syntax used for training and sample data applies to the evaluation data.

The general format of the evaluation section of the challenge is the following:

evaluation_protocol:
  name_of_procedure:
    tool:
      id: id of the evaluation tool
      parameters: optional parameters of the evaluation tool
    data:
      name_of_the_data:
        tool:
          id: id of the data tool
          parameters: optional parameters of the data tool
          input_data: optional directory containing input data for the data tool
        sources:   # optional description of the dataset
          - title: title of the evaluation dataset
            url: original url of the dataset
            description: |
              Multiline description of the dataset
    constraints:
       - metric: accuracy
         condition: larger_than
         value: 0.99
       - metric: memory
         condition: smaller_than
         value: 1000
       - metric: latency
         value: smaller_than
         value: 100

The evaluation_protocol.{name_of_procedure}.data.{name_of_the_data}.sources section has the same structure as the corresponding sections in the training and sample data and it is optional. The name_of_procedure property can be chosen by the challenge writer, multiple evaluation procedures in the same challenge. The section evaluation_protocol.{name_of_procedure}.data.{name_of_the_data}.tool can be substituted with a single property evaluation_protocol.{name_of_procedure}.data.{name_of_the_data}.dataset that points to a directory containing the data in a already standardized format.

The Bonseyes platform provides some pre-defined evaluation tool for typical evaluations. The evaluation tools available are the following:

Image Classification Evaluation Tool (id: com_bonseyes/challenge_tools#evaluation/image_classification): measures accuracy and latency of a Image Classification AI app using a benchmark dataset.
Object Detection Evaluation Tool (id: com_bonseyes/challenge_tools#evaluation/object_detection): measures mAP and latency of a Object Classification AI app using a benchmark dataset.

The details of the tools are described in the next paragraphs. Custom evalution tools can be defined following the instructions in /pages/dev_guides/eval_tool/create_evaluation_tool

The image classification evaluation tool allows to measure accuracy and latency of a image classification ai app.

The minimal challenge file to use this evaluation tool is the following:

metadata:
  title: My Sample Challenge
  description: |
    This challenge is a sample challenge

specification:
  interface:
    id: com_bonseyes/interfaces#image_classification
    parameters:
      classes: classes.txt

evaluation_protocol:
  evaluation:
    tool:
      id: com_bonseyes/challenge_tools#evaluation/image_classification
    data:
      evaluation_data:
        dataset: benchmark_data

The sections metadata and specification are explained above. The evaluation protocol section describes a single evaluation procedure called evaluation. It uses the image classification evaluation tool and the data present in the directory benchmark_data relative to the challenge.yml file.

The data in the directory must have the following format:

The directory must contain a file dataset.json with the following format:

{
  "samples": [
    {
      "id": "id_of_image1",
      "class": "index of the class as defined in the classes.txt",
      "image": "path to the image relative to directory containing the dataset.json file"
    },
    {
      "id": "id_of_image2",
      ...
    }
  ]
}

The directory must contain all the images to be tested in either PNG or JPEG format.

The object detection evaluation tool allows to measure mAP and latency of an object detection ai app.

The minimal challenge file to use this evaluation tool is the following:

metadata:
  title: My Sample Challenge
  description: |
    This challenge is a sample challenge

specification:
  interface:
    id: com_bonseyes/interfaces#object_detection
    parameters:
      classes: classes.txt

evaluation_protocol:
  evaluation:
    tool:
      id: com_bonseyes/challenge_tools#evaluation/object_detection
    data:
      evaluation_data:
        dataset: benchmark_data

The sections metadata and specification are explained above. The evaluation protocol section describes a single evaluation procedure called evaluation. It uses the image classification evaluation tool and the data present in the directory benchmark_data relative to the challenge.yml file.

The data in the directory must have the following format:

The directory must contain a file dataset.json with the following format:

{
  "samples": [
    {
      "id": "id_of_image1",
      "image": "path to the image relative to directory containing the dataset.json file"
      "bb": [
        {
          "class_index": "index of the class of the first object as defined in classes.txt",
          "bounding_box": {
            "origin": {
              "x": minimum x of bounding box,
              "y": minimum y of bounding box
            },
            "size": {
              "x": width of bounding box,
              "y": height of boudning box
            }
          }
        },
        {
          "class_index": "index of the class of the second object as defined in classes.txt",
           ...
        }
      ],
    },
    {
      "id": "id_of_image2",
      ...
    }
  ]
}

The directory must contain all the images to be tested in either PNG or JPEG format.

Upload Challenge

THIS SECTION IS IN PROGRESS

This guide assumes you already have the done the following:

Setup the local environment as explained in Prerequisites
Created (see create_challenge) or obtained a challenge and stored it in the directory challenge
Created a git repository on Gitlab that will host the challenge
Enrolled your gitlab user in the marketplace

To upload a challenge you need to do the following:

Create your author keys:

$ bonseyes license create-author-keys --output-dir author_keys

Package your challenge for release with its dependencies and a redistribution licenses:

$ bonseyes marketplace package-challenge \
             --pull-images \
             --challenge challenge \
             --author-keys author_keys \
             --amount "100 CHF" \
             --recipients "test@dd.com,tost@dd.com" \
             --regions https://www.wikidata.org/wiki/Q39,https://www.wikidata.org/wiki/Q40 \
             --output-dir packaged_challenge \

The --recipients parameter can be used to restrict who can see the artifact on the marketplace, if omitted all users can see it.

The --regions parameter can be used to restrict the regions in which the artifact can be seen, if omitted users of all regions can see it.

Commit the packaged challenge in a gitlab repository that is accessible by the gitlab user you enabled in the marketplace

Upload the challenge to the marketplace:

$ bonseyes marketplace upload-challenge --url $HTTP_URL_OF_GITLAB_REPO

The challenge will be added to your profile.