craft-ai 2.4.3

craft ai API python 3 client

craft ai API python client

craft ai's Explainable AI API enables product & operational teams to quickly deploy and run explainable AIs. craft ai decodes your data streams to deliver self-learning services.

Get Started!

1 - Create a project

Once your account is setup, let's create your first project! Go in the 'Projects' tab in the craft ai control center at https://beta.craft.ai/inspector, and press Create a project.

Once it's done, you can click on your newly created project to retrieve its tokens. There are two types of tokens: read and write. You'll need the write token to create, update and delete your agent.

2 - Setup

Install

PIP / PyPI

Let's first install the package from pip.

pip install --upgrade craft-ai

Depending on your setup you may need to use pip3 or pipenv instead of pip.

Then import it in your code

import craft_ai

This client also provides helpers to use it in conjuction with pandas

Initialize

client = craft_ai.Client({
  "token": "{token}"
})

3 - Create an agent

craft ai is based on the concept of agents. In most use cases, one agent is created per user or per device.

An agent is an independent data set that stores the history of the context of its user or device's context, and learns which prediction to make based on the evolution of this context.

In this example, we will create an agent that learns the predictive model of a light bulb based on the time of the day and the number of people in the room. This dataset is treated as continuous context updates. If your data is more like events than context changes, please refer to the Advanced Configuration section to know how to configure operations_as_events for your agent. Here, the agent's context has 4 properties or features:

• peopleCount which is a continuous property,
• timeOfDay which is a time_of_day property,
• timezone, a property of type timezone needed to generate proper values for timeOfDay (cf. the context properties type section for further information),
• and finally lightbulbState which is an enum property that is also the output.

:information_source: timeOfDay is auto-generated, you will find more information below.

agent_id = "my_first_agent"
configuration = {
  "context": {
    "peopleCount": {
      "type": "continuous"
    },
    "timeOfDay": {
      "type": "time_of_day"
    },
    "timezone": {
      "type": "timezone"
    },
    "lightbulbState": {
      "type": "enum"
    }
  },
  "model_type": "decisionTree",
  "output": ["lightbulbState"]
}

agent = client.create_agent(configuration, agent_id)
print("Agent", agent["id"], "has successfully been created")

Pretty straightforward to test! Open https://beta.craft.ai/inspector, select you project and your agent is now listed.

Now, if you run that a second time, you'll get an error: the agent 'my_first_agent' was already created. Let's see how we can delete it before recreating it.

agent_id = "my_first_agent"
client.delete_agent(agent_id)
print("Agent", agent_id, "no longer exists")

configuration = ...
agent = client.create_agent(configuration, agent_id)
print("Agent", agent["id"], "has successfully been created")

For further information, check the 'create agent' reference documentation.

4 - Add context operations

We have now created our first agent but it is not able to do much, yet. To learn a model it needs to be provided with data, in craft ai these are called context operations.

In the following we add 8 operations:

1. The initial one sets the initial state of the agent, on July 25 2016 at 5:30, in Paris, nobody is there and the light is off;
2. At 7:02, someone enters the room the light is turned on;
3. At 7:15, someone else enters the room;
4. At 7:31, the light is turned off;
5. At 8:12, everyone leaves the room;
6. At 19:23, 2 persons enter the room;
7. At 22:35, the light is turned on;
8. At 23:06, everyone leaves the room and the light is turned off.

agent_id = "my_first_agent"
client.delete_agent(agent_id)
print("Agent", agent_id, "no longer exists")

configuration = ...
agent = client.create_agent(configuration, agent_id)
print("Agent", agent["id"], "has successfully been created")

context_list = [
  {
    "timestamp": 1469410200,
    "context": {
      "timezone": "+02:00",
      "peopleCount": 0,
      "lightbulbState": "OFF"
    }
  },
  {
    "timestamp": 1469415720,
    "context": {
      "timezone": "+02:00",
      "peopleCount": 1,
      "lightbulbState": "ON"
    }
  },
  {
    "timestamp": 1469416500,
    "context": {
      "timezone": "+02:00",
      "peopleCount": 2,
      "lightbulbState": "ON"
    }
  },
  {
    "timestamp": 1469417460,
    "context": {
      "timezone": "+02:00",
      "peopleCount": 2,
      "lightbulbState": "OFF"
    }
  },
  {
    "timestamp": 1469419920,
    "context": {
      "timezone": "+02:00",
      "peopleCount": 0,
      "lightbulbState": "OFF"
    }
  },
  {
    "timestamp": 1469460180,
    "context": {
      "timezone": "+02:00",
      "peopleCount": 2,
      "lightbulbState": "OFF"
    }
  },
  {
    "timestamp": 1469471700,
    "context": {
      "timezone": "+02:00",
      "peopleCount": 2,
      "lightbulbState": "ON"
    }
  },
  {
    "timestamp": 1469473560,
    "context": {
      "timezone": "+02:00",
      "peopleCount": 0,
      "lightbulbState": "ON"
    }
  }
]
client.add_agent_operations(agent_id, context_list)
print("Successfully added initial operations to agent", agent_id, "!")

In real-world applications you will probably do the same kind of thing when the agent is created, and then regularly throughout the lifetime of the agent with newer data.

For further information, check the 'add context operations' reference documentation.

5 - Compute the decision tree

The agent has acquired a context history, we can now compute a model (in this case a decision tree) from it! A decision tree models the output, allowing us to estimate what the output would be in a given context.

The decision tree is computed at a given timestamp, which means it will consider the data from the creation of this agent up to this moment. Let's first try to compute the decision tree at midnight on July 26, 2016.

agent_id = "my_first_agent"

client.delete_agent(agent_id)
print("Agent", agent_id, "no longer exists")

configuration = ...
agent = client.create_agent(configuration, agent_id)
print("Agent", agent["id"], "has successfully been created")

context_list = ...
client.add_agent_operations(agent_id, context_list)
print("Successfully added initial operations to agent", agent_id, "!")

dt_timestamp = 1469476800
decision_tree = client.get_agent_decision_tree(agent_id, dt_timestamp)
print("The full decision tree at timestamp", dt_timestamp, "is the following:")
import json
print(json.dumps(decision_tree,indent=2))
""" Outputted tree is the following
  {
    "_version":"2.0.0",
    "trees":{
      "lightbulbState":{
        "output_values" : ["OFF", "ON"],
        "children":[
          {
            "children":[
              {
                "prediction":{
                  "confidence":0.6774609088897705,
                  "distribution":[0.8, 0.2],
                  "value":"OFF",
                  "nb_samples": 5
                },
                "decision_rule":{
                  "operand":0.5,
                  "operator":"<",
                  "property":"peopleCount"
                }
              },
              {
                "prediction":{
                  "confidence":0.8630361557006836,
                  "distribution":[0.1, 0.9],
                  "value":"ON",
                  "nb_samples": 10
                },
                "decision_rule":{
                  "operand":0.5,
                  "operator":">=",
                  "property":"peopleCount"
                }
              }
            ],
            "decision_rule":{
              "operand":[
                5,
                5.6666665
              ],
              "operator":"[in[",
              "property":"timeOfDay"
            }
          },
          {
            "children":[
              {
                "prediction":{
                  "confidence":0.9947378635406494,
                  "distribution":[1.0, 0.0],
                  "value":"ON",
                  "nb_samples": 10
                },
                "decision_rule":{
                  "operand":[
                    5.6666665,
                    20.666666
                  ],
                  "operator":"[in[",
                  "property":"timeOfDay"
                }
              },
              {
                "children":[
                  {
                    "prediction":{
                      "confidence":0.969236433506012,
                      "distribution":[0.95, 0.05],
                      "value":"OFF",
                      "nb_samples": 10
                    },
                    "decision_rule":{
                      "operand":1,
                      "operator":"<",
                      "property":"peopleCount"
                    }
                  },
                  {
                    "prediction":{
                      "confidence":0.8630361557006836,
                      "distribution":[0.2, 0.8],
                      "value":"ON",
                      "nb_samples": 15
                    },
                    "decision_rule":{
                      "operand":1,
                      "operator":">=",
                      "property":"peopleCount"
                    }
                  }
                ],
                "decision_rule":{
                  "operand":[
                    20.666666,
                    5
                  ],
                  "operator":"[in[",
                  "property":"timeOfDay"
                }
              }
            ],
            "decision_rule":{
              "operand":[
                5.6666665,
                5
              ],
              "operator":"[in[",
              "property":"timeOfDay"
            }
          }
        ]
      }
    },
    "configuration":{
      "time_quantum":600,
      "learning_period":9000000,
      "context":{
        "peopleCount":{
          "type":"continuous"
        },
        "timeOfDay":{
          "type":"time_of_day",
          "is_generated":True
        },
        "timezone":{
          "type":"timezone"
        },
        "lightbulbState":{
          "type":"enum"
        }
      },
      "output":[
        "lightbulbState"
      ]
    }
  }
"""

Try to retrieve the tree at different timestamps to see how it gradually learns from the new operations. To visualize the trees, use the inspector!

For further information, check the 'compute decision tree' reference documentation.

6 - Make a decision

Once the decision tree is computed it can be used to make a decision or prediction. In our case it is basically answering this type of question: "What is the anticipated state of the lightbulb at 7:15 if there are 2 persons in the room ?".

agent_id = "my_first_agent"

client.delete_agent(agent_id)
print("Agent", agent_id, "no longer exists")

configuration = ...
agent = client.create_agent(configuration, agent_id)
print("Agent", agent["id"], "has successfully been created")

context_list = ...
client.add_agent_operations(agent_id, context_list)
print("Successfully added initial operations to agent", agent_id, "!")

dt_timestamp = 1469476800
decision_tree = client.get_agent_decision_tree(agent_id, dt_timestamp)
print("The decision tree at timestamp", dt_timestamp, "is the following:")
print(decision_tree)

context = {
  "timezone": "+02:00",
  "timeOfDay": 7.25,
  "peopleCount": 2
}
resp = client.decide(decision_tree, context)
print("The anticipated lightbulb state is:", resp["output"]["lightbulbState"]["predicted_value"])

For further information, check the 'make decision' reference documentation.

Python starter kit

If you prefer to get started from an existing code base, the official Python starter kit can get you there! Retrieve the sources locally and follow the "readme" to get a fully working Wellness Coach example using real-world data.

:package: Get the craft ai Python Starter Kit

API

Project

craft ai agents belong to projects. In the current version, each identified users defines a owner and can create projects for themselves, in the future we will introduce shared projects.

Configuration

Each agent has a configuration defining:

• the context schema, i.e. the list of property keys and their type (as defined in the following section),
• the output properties, i.e. the list of property keys on which the agent makes decisions,
• the model type, either decision tree or gradient boosting.

Context properties types

Base types: enum, continuous and boolean

enum, continuous and boolean are the three base craft ai types:

• an enum property is a string;
• a continuous property is a real number.
• a boolean property is a boolean value: true or false

:warning: the absolute value of a continuous property must be less than 10²⁰.

Here is a simple example of configuration for decision tree:

{
  "context": {
    "timezone": {
      "type": "enum"
    },
    "temperature": {
      "type": "continuous"
    },
    "lightbulbState": {
      "type": "enum"
    }
  },
  "model_type": "decisionTree",
  "output": ["lightbulbState"],
  "time_quantum": 100,
  "learning_period": 108000
}

And another simple example of configuration for gradient boosting:

{
  "context": {
    "timezone": {
      "type": "enum"
    },
    "temperature": {
      "type": "continuous"
    },
    "lightbulbState": {
      "type": "enum"
    }
  },
  "model_type": "boosting",
  "output": ["lightbulbState"],
  "learning_rate": 1,
  "num_iterations": 50,
  "time_quantum": 100,
  "learning_period": 108000
}

Time types: timezone, time_of_day, day_of_week, day_of_month and month_of_year

craft ai defines the following types related to time:

• a time_of_day property is a real number belonging to [0.0; 24.0[, each value represents the number of hours in the day since midnight (e.g. 13.5 means 13:30),
• a day_of_week property is an integer belonging to [0, 6], each value represents a day of the week starting from Monday (0 is Monday, 6 is Sunday).
• a day_of_month property is an integer belonging to [1, 31], each value represents a day of the month.
• a month_of_year property is an integer belonging to [1, 12], each value represents a month of the year.
• a timezone property can be:

  • a string value representing the timezone as an offset from UTC, supported formats are:

    • ±[hh]:[mm],
    • ±[hh][mm],
    • ±[hh],

    where hh represent the hour and mm the minutes from UTC (eg. +01:30)), between -12:00 and +14:00.

  • an integer belonging to [-720, 840] which represents the timezone as an offset from UTC:

    • in hours if the integer belongs to [-15, 15]
    • in minutes otherwise

  • an abbreviation among the following:

    • UTC or Z Universal Time Coordinated,
    • GMT Greenwich Mean Time, as UTC,
    • BST British Summer Time, as UTC+1 hour,
    • IST Irish Summer Time, as UTC+1,
    • WET Western Europe Time, as UTC,
    • WEST Western Europe Summer Time, as UTC+1,
    • CET Central Europe Time, as UTC+1,
    • CEST Central Europe Summer Time, as UTC+2,
    • EET Eastern Europe Time, as UTC+2,
    • EEST Eastern Europe Summer Time, as UTC+3,
    • MSK Moscow Time, as UTC+3,
    • MSD Moscow Summer Time, as UTC+4,
    • AST Atlantic Standard Time, as UTC-4,
    • ADT Atlantic Daylight Time, as UTC-3,
    • EST Eastern Standard Time, as UTC-5,
    • EDT Eastern Daylight Saving Time, as UTC-4,
    • CST Central Standard Time, as UTC-6,
    • CDT Central Daylight Saving Time, as UTC-5,
    • MST Mountain Standard Time, as UTC-7,
    • MDT Mountain Daylight Saving Time, as UTC-6,
    • PST Pacific Standard Time, as UTC-8,
    • PDT Pacific Daylight Saving Time, as UTC-7,
    • HST Hawaiian Standard Time, as UTC-10,
    • AKST Alaska Standard Time, as UTC-9,
    • AKDT Alaska Standard Daylight Saving Time, as UTC-8,
    • AEST Australian Eastern Standard Time, as UTC+10,
    • AEDT Australian Eastern Daylight Time, as UTC+11,
    • ACST Australian Central Standard Time, as UTC+9.5,
    • ACDT Australian Central Daylight Time, as UTC+10.5,
    • AWST Australian Western Standard Time, as UTC+8.

:information_source: By default, the values of the time_of_day and day_of_week properties are generated from the timestamp of an agent's state and the agent's current timezone. Therefore, whenever you use generated time_of_day and/or day_of_week in your configuration, you must provide a timezone value in the context. There can only be one timezone property.

If you wish to provide their values manually, add is_generated: false to the time types properties in your configuration. In this case, since you provide the values, the timezone property is not required, and you must update the context whenever one of these time values changes in a way that is significant for your system.

Examples

Let's take a look at the following configuration. It is designed to model the color of a lightbulb (the lightbulbColor property, defined as an output) depending on the outside light intensity (the lightIntensity property), the TV status (the TVactivated property) the time of the day (the time property) and the day of the week (the day property).

day and time values will be generated automatically, hence the need for timezone, the current Time Zone, to compute their value from given timestamps.

The time_quantum is set to 100 seconds, which means that if the lightbulb color is changed from red to blue then from blue to purple in less that 1 minutes and 40 seconds, only the change from red to purple will be taken into account.

The learning_period is set to 108 000 seconds (one month) , which means that the state of the lightbulb from more than a month ago can be ignored when learning the decision model.

{
  "context": {
    "lightIntensity": {
      "type": "continuous"
    },
    "TVactivated": {
      "type": "boolean"
    },
    "time": {
      "type": "time_of_day"
    },
    "day": {
      "type": "day_of_week"
    },
    "timezone": {
      "type": "timezone"
    },
    "lightbulbColor": {
      "type": "enum"
    }
  },
  "model_type": "decisionTree",
  "output": ["lightbulbColor"],
  "time_quantum": 100,
  "learning_period": 108000
}

In this second example, the time property is not generated, no property of type timezone is therefore needed. However values of time must be manually provided continuously.

{
  "context": {
    "time": {
      "type": "time_of_day",
      "is_generated": false
    },
    "lightIntensity": {
      "type": "continuous"
    },
      "TVactivated": {
      "type": "boolean"
    },
    "lightbulbColor": {
      "type": "enum"
    }
  },
  "model_type": "decisionTree",
  "output": ["lightbulbColor"],
  "time_quantum": 100,
  "learning_period": 108000
}

Timestamp

craft ai API heavily relies on timestamps. A timestamp is an instant represented as a Unix time, that is to say the amount of seconds elapsed since Thursday, 1 January 1970 at midnight UTC. Note that some programming languages use timestamps in milliseconds, but here we only refer to timestamps in seconds. In most programming languages this representation is easy to retrieve, you can refer to this page to find out how.

craft_ai.Time

The craft_ai.Time class facilitates the handling of time types in craft ai. It is able to extract the different craft ai formats from various datetime representations, thanks to datetime.

# From a unix timestamp and an explicit UTC offset
t1 = craft_ai.Time(1465496929, "+10:00")

# t1 == {
#   utc: "2016-06-09T18:28:49.000Z",
#   timestamp: 1465496929,
#   day_of_week: 4,
#   time_of_day: 4.480277777777778,
#   timezone: "+10:00"
# }

# From a unix timestamp and using the local UTC offset.
t2 = craft_ai.Time(1465496929)

# Value are valid if in Paris !
# t2 == {
#   utc: "2016-06-09T18:28:49.000Z",
#   timestamp: 1465496929,
#   day_of_week: 3,
#   time_of_day: 20.480277777777776,
#   timezone: "+02:00"
# }

# From a ISO 8601 string. Note that here it should not have any ":" in the timezone part
t3 = craft_ai.Time("1977-04-22T01:00:00-0500")

# t3 == {
#   utc: "1977-04-22T06:00:00.000Z",
#   timestamp: 230536800,
#   day_of_week: 4,
#   time_of_day: 1,
#   timezone: "-05:00"
# }

# Retrieve the current time with the local UTC offset
now = craft_ai.Time()

# Retrieve the current time with the given UTC offset
nowP5 = craft_ai.Time(timezone="+05:00")

Configuration parameters

The following configuration parameters can be set in specific cases.

Common parameters

• model_type, i.e. the selected model. Values can be decisionTree or boosting. If not set, the default value is decisionTree.
• time_quantum, i.e. the minimum amount of time, in seconds, that is meaningful for an agent; context updates occurring faster than this quantum won't be taken into account. As a rule of thumb, you should always choose the largest value that seems right and reduce it, if necessary, after some tests. Default value is 600. This parameter is ignored if operations_as_events is set to true.
• operations_as_events is a boolean, either true or false. The default value is false. If you are not sure what to do, set it to true. If it is set to false, context operations are treated as state changes, and models are based on the resulting continuous state including between data points, using time_quantum as the sampling step. If it is set to true, context operations are treated as observations or events, and models are based on these data points directly, as in most machine learning libraries. If operations_as_events is true, max_training_samples and learning_period for decision trees must be set, and time_quantum is ignored because events have no duration.
• max_training_samples is a positive integer. It can and must be set only if operations_as_events is true. It defines the maximum number of events on which a model can be based. It is complementary to learning_period for decision trees, which limits the maximum age of data on which a model is based.
• min_samples_per_leaf is a positive integer. It defines the minimum number of samples in a tree leaf. It is complementary to tree_max_depth in preventing the tree from overgrowing, hence limiting overfitting. By default, min_samples_per_leaf is set to 4.
• tree_max_depth is a positive integer. It defines the maximum depth of decision trees, which is the maximum distance between the root node and a leaf (terminal) node. A depth of 0 means that the tree is made of a single root node. By default, tree_max_depth is set to 6 if the output is categorical (e.g. enum), or to 4 if the output is numerical (e.g. continuous) or if it's a boosting configuration.

Decision tree parameters

• learning_period, i.e. the maximum amount of time, in seconds, that matters for an agent; the agent's decision model can ignore context that is older than this duration. You should generally choose the smallest value that fits this description. Default value is 15000 time quantums and the maximum learning_period value is 55000 * time_quantum.

Boosting parameters

• learning_rate is a positive float. It defines the step size shrinkage used between tree updates to prevent overfitting. Its value must be in ]0;1].
• num_iterations is a positive integer. It describes the number of trees that would be created for the forest.

Agent

Create

Create a new agent, and define its configuration.

The agent's identifier is a case sensitive string between 1 and 36 characters long. It only accepts letters, digits, hyphen-minuses and underscores (i.e. the regular expression /[a-zA-Z0-9_-]{1,36}/).

client.create_agent(
  { # The configuration
    "context": {
      "peopleCount": {
        "type": "continuous"
      },
      "timeOfDay": {
        "type": "time_of_day"
      },
      "timezone": {
        "type": "timezone"
      },
      "lightbulbState": {
        "type": "enum"
      }
    },
    "model_type": "decisionTree",
    "output": [ "lightbulbState" ],
    "time_quantum": 100,
    "learning_period": 108000
  },
  "my_new_agent" # id for the agent, if undefined a random id is generated
)

Delete

client.delete_agent(
  "my_new_agent" # The agent id
)

Retrieve

client.get_agent(
  "my_new_agent" # The agent id
)

List

client.list_agents()
# Return a list of agents' name
# Example: [ "my_new_agent", "joyful_octopus", ... ]

Create and retrieve shared url

Create and get a shareable url to view an agent tree. Only one url can be created at a time.

client.get_shared_agent_inspector_url(
  "my_new_agent", # The agent id.
  1464600256 # optional, the timestamp for which you want to inspect the tree.
)

Delete shared url

Delete a shareable url. The previous url cannot access the agent tree anymore.

client.delete_shared_agent_inspector_url(
  'my_new_agent' # The agent id.
)

Generator

The craft ai API lets you generate models built on data from one or several agents by creating a generator. It is useful to:

• test several hyper-parameters and features sets without reloading all the data for each try
• gather data from different agents to make new models based on several data sources, enhancing the possible data combinations and allowing you to inspect the global behavior across your agents

The data stream(s) used by a generator are defined by specifying a list of agents as a filter in its configuration. Other than the filter, the configuration of a generator is similar to an agent's configuration. But it has to verify some additional properties:

• Every feature defined in the context configuration of the generator must be present in all the agent that match the filter, with the same context types.
• The parameter operations_as_events must be set to true.
• It follows that the parameters max_training_samples, and learning_period in the case of decision trees, must be set.
• The agent names provided in the list must be valid agent identifiers.

Create

Create a new generator, and define its configuration.

The generator's identifier is a case sensitive string between 1 and 36 characters long. It only accepts letters, digits, hyphen-minuses and underscores (i.e. the regular expression /[a-zA-Z0-9_-]{1,36}/).

GENERATOR_NAME = "smarthome_gen"
GENERATOR_FILTER = ["smarthome"]
GENERATOR_CONFIGURATION = {
  "context": {
    "light": {
      "type": "enum"
    },
    "tz": {
      "type": "timezone"
    },
    "movement": {
      "type": "continuous"
    },
    "time": {
      "type": "time_of_day",
      "is_generated": True
    }
  },
  "model_type": "decisionTree",
  "output": [
    "light"
  ],
  "learning_period": 1500000,
  "max_training_samples": 15000,
  "operations_as_events": True,
  "filter": GENERATOR_FILTER
}

client.create_generator(
  GENERATOR_CONFIGURATION, # A valid configuration.
  GENERATOR_NAME # The generator id.
)

Delete

GENERATOR_NAME = 'smarthome_gen'
client.delete_generator(
  GENERATOR_NAME
)

Retrieve

GENERATOR_NAME = 'smarthome_gen'
client.get_generator(
  GENERATOR_NAME
)

### Ouputted info is the following
"""
{
  "_version": "2.0.0"
  "id": "smarthome_gen",
  "configuration": {
    "operations_as_events": True,
    "learning_period": 1500000,
    "max_training_samples": 15000,
    "context": {
      "light": {
        "type": "enum"
      },
      "tz": {
        "type": "timezone"
      },
      "movement": {
        "type": "continuous"
      },
      "time": {
        "type": "time_of_day",
        "is_generated": True
      }
    },
    "output": [
      "light"
    ],
    "filter": [
      "smarthome"
    ]
  },
  "firstTimestamp": 1254836352,
  "lastTimestamp": 1272721522,
  "agents": [
    "smarthome"
  ],
}
"""
###

Retrieve generators list

client.list_generators() # Return the list of generators in the project.

List operations in the generator

Retrieve the context operations of agents matching the generator's filter. Each operation also contains the identifier of the agent for which it was added, in the agent_id property.

GENERATOR_NAME = 'smarthome_gen'
START_TIMESTAMP = 1478894153
END_TIMESTAMP = 1478895266

client.get_generator_operations(
  GENERATOR_NAME,   # The generator id
  START_TIMESTAMP,  # Optional, the **start** timestamp from which the
                    # operations are retrieved (inclusive bound)
  END_TIMESTAMP     # Optional, the **end** timestamp up to which the
                    # operations are retrieved (inclusive bound)
)

This call can generate multiple requests to the craft ai API as results are paginated.

Context

Add operations

client.add_agent_operations(
  "my_new_agent", # The agent id
  [ # The list of context operations
    {
      "timestamp": 1469410200,
      "context": {
        "timezone": "+02:00",
        "peopleCount": 0,
        "lightbulbState": "OFF"
      }
    },
    {
      "timestamp": 1469415720,
      "context": {
        "timezone": "+02:00",
        "peopleCount": 1,
        "lightbulbState": "ON"
      }
    },
    {
      "timestamp": 1469416500,
      "context": {
        "timezone": "+02:00",
        "peopleCount": 2,
        "lightbulbState": "ON"
      }
    },
    {
      "timestamp": 1469417460,
      "context": {
        "timezone": "+02:00",
        "peopleCount": 2,
        "lightbulbState": "OFF"
      }
    },
    {
      "timestamp": 1469419920,
      "context": {
        "timezone": "+02:00",
        "peopleCount": 0,
        "lightbulbState": "OFF"
      }
    },
    {
      "timestamp": 1469460180,
      "context": {
        "timezone": "+02:00",
        "peopleCount": 2,
        "lightbulbState": "OFF"
      }
    },
    {
      "timestamp": 1469471700,
      "context": {
        "timezone": "+02:00",
        "peopleCount": 2,
        "lightbulbState": "ON"
      }
    },
    {
      "timestamp": 1469473560,
      "context": {
        "timezone": "+02:00",
        "peopleCount": 0,
        "lightbulbState": "ON"
      }
    }
  ]
)

Missing Values

If the value of a base type property is missing, you can send a null value. craft ai will take into account as much information as possible from this incomplete context.

A context operation with a missing value looks like:

[
  {
    "timestamp": 1469415720,
    "context": {
      "peopleCount": "OFF",
      "lightbulbState": null
    }
  },
  ...
]

Optional Values

If the value of an optional property is not filled at some point—as should be expected from an optional value—send the empty JSON Object {} to craft ai:

A context with an optional value looks like:

[
  {
    "timestamp": 1469415720,
    "context": {
      "timezone": "+02:00",
      "temperature": {},
      "lightbulbState": "OFF"
    }
  },
  ...
]

List operations

client.get_agent_operations(
  "my_new_agent", # The agent id
  1478894153, # Optional, the **start** timestamp from which the
              # operations are retrieved (inclusive bound)
  1478895266, # Optional, the **end** timestamp up to which the
              # operations are retrieved (inclusive bound)
)

This call can generate multiple requests to the craft ai API as results are paginated.

Retrieve state

client.get_context_state(
  "my_new_agent", # The agent id
  1469473600 # The timestamp at which the context state is retrieved
)

Retrieve state history

client.get_agent_states(
  "my_new_agent", # The agent id
  1478894153, # Optional, the **start** timestamp from which the
              # operations are retrieved (inclusive bound)
  1478895266, # Optional, the **end** timestamp up to which the
              # operations are retrieved (inclusive bound)
)

Gradient boosting

Models can be generated with gradient boosting by setting the configuration parameter model_type to boosting. Models are based on training data within a provided timestamp window among data that was added. You can only query predictions directly for gradient boosting models.

The implementation is based on LightGBM, but there are some parameters that differ from the ones used by default by LightGBM.

For classification:

• max_bin = 255. Max number of bins that feature values will be bucketed in (https://lightgbm.readthedocs.io/en/latest/Parameters.html#max_bin).

For regression:

• metric = L2 (alias mse). Metric(s) to be evaluated on the evaluation set(s) (https://lightgbm.readthedocs.io/en/latest/Parameters.html#metric).
• feature_fraction = 0.9. Randomly select a subset of features on each iteration (https://lightgbm.readthedocs.io/en/latest/Parameters.html#feature_fraction).
• bagging_freq = 5. Perform bagging at every k iteration. Every k-th iteration, LightGBM will randomly select bagging_fraction * 100% of the data to use for the next k iterations (https://lightgbm.readthedocs.io/en/latest/Parameters.html#bagging_freq).
• bagging_fraction = 0.8. It will randomly select part of data without resampling (https://lightgbm.readthedocs.io/en/latest/Parameters.html#bagging_fraction).
• min_sum_hessian_in_leaf = 5.0. It's the minimal sum hessian in one leaf (https://lightgbm.readthedocs.io/en/latest/Parameters.html#min_sum_hessian_in_leaf).

See the configuration section for parameters that you can set.

Get decision using boosting for agent

FROM_TIMESTAMP = 1469473600
TO_TIMESTAMP = 1529473600
PREDICTION_CONTEXT = {
  "tz": "+02:00",
  "movement": 2,
  "time": 7.5
}

client.get_agent_boosting_decision(
  'impervious_kraken', // The generator id
  FROM_TIMESTAMP,
  TO_TIMESTAMP,
  PREDICTION_CONTEXT
)
"""
{
  "context": {
    "tz": "+02:00",
    "movement": 2,
    "time": 7.5
  },
  "output": {
    "predicted_value": "OFF"
  }
}
"""

Get decision using boosting for generator

FROM_TIMESTAMP = 1469473600
TO_TIMESTAMP = 1529473600
PREDICTION_CONTEXT = {
  "tz": "+02:00",
  "movement": 2,
  "time": 7.5
}

client.compute_generator_boosting_decision(
  'impervious_kraken', // The generator id
  FROM_TIMESTAMP,
  TO_TIMESTAMP,
  PREDICTION_CONTEXT
)
"""
{
  "context": {
    "tz": "+02:00",
    "movement": 2,
    "time": 7.5
  },
  "output": {
    "predicted_value": "OFF"
  }
}
"""

Decision tree

Models can be generated as single decision trees by setting the configuration parameter model_type to decisionTree. Decision trees are computed based on data up to a specific timestamp and dating back to the learning_period configuration parameter among data that was added.

When you compute a decision tree, craft ai returns an object containing:

• the version of the model's format

• the agent's configuration as specified during the agent's creation

• the tree itself as a JSON object:

  • Internal nodes are represented by a "decision_rule" object and a "children" array. The first one, contains the "property, and the "property"'s value, to decide which child matches a context.
  • Leaves have a "predicted_value", "confidence" and "decision_rule" object for this value, instead of a "children" array. "predicted_value" is an estimation of the output in the contexts matching the node. "confidence" is a number between 0 and 1 that indicates how confident craft ai is that the output is a reliable prediction. When the output is a numerical type, leaves also have a "standard_deviation" that indicates a margin of error around the "predicted_value".
  • The root only contains a "children" array.

Get decision tree for an agent

client.get_agent_decision_tree(
  "my_new_agent", # The agent id
  1469473600 # Optional the timestamp at which we want the decision
             # tree, default behavior is to return the decision tree
             # from the latest timestamp in context operations
)

Get decision using a decision tree for an agent

:information_source: To make a decision (prediction) with decision tree, first compute the decision tree then use the offline interpreter.

Get decision tree for a generator

DECISION_TREE_TIMESTAMP = 1469473600
GENERATOR_NAME = 'smarthome_gen'
client.get_generator_decision_tree(
  GENERATOR_NAME, # The generator id
  DECISION_TREE_TIMESTAMP # The timestamp at which the decision tree is retrieved
)

""" Outputted tree is the following
{
  "_version": "2.0.0",
  "trees": {
    "light": {
      "children": [
        {
          "predicted_value": "OFF",
          "confidence": 0.9966583847999572,
          "decision_rule": {
            "operand": [
              7.25,
              22.65
            ],
            "operator": "[in[",
            "property": "time"
          }
        },
        {
          "children": [
            {
              "predicted_value": "ON",
              "confidence": 0.9618390202522278,
              "decision_rule": {
                "operand": [
                  22.65,
                  0.06666667
                ],
                "operator": "[in[",
                "property": "time"
              }
            },
            {
              "children": [
                {
                  "predicted_value": "OFF",
                  "confidence": 0.9797198176383972,
                  "decision_rule": {
                    "operand": [
                      0.06666667,
                      0.6
                    ],
                    "operator": "[in[",
                    "property": "time"
                  }
                },
                {
                  "children": [
                    {
                      "predicted_value": "ON",
                      "confidence": 0.9585137963294984,
                      "decision_rule": {
                        "operand": [
                          0.6,
                          2.25
                        ],
                        "operator": "[in[",
                        "property": "time"
                      }
                    },
                    {
                      "children": [
                        {
                          "predicted_value": "OFF",
                          "confidence": 0.8077218532562256,
                          "decision_rule": {
                            "operand": [
                              2.25,
                              2.4666667
                            ],
                            "operator": "[in[",
                            "property": "time"
                          }
                        }
                      ],
                      "decision_rule": {
                        "operand": [
                          2.25,
                          7.25
                        ],
                        "operator": "[in[",
                        "property": "time"
                      }
                    }
                  ],
                  "decision_rule": {
                    "operand": [
                      0.6,
                      7.25
                    ],
                    "operator": "[in[",
                    "property": "time"
                  }
                }
              ],
              "decision_rule": {
                "operand": [
                  0.06666667,
                  7.25
                ],
                "operator": "[in[",
                "property": "time"
              }
            }
          ],
          "decision_rule": {
            "operand": [
              22.65,
              7.25
            ],
            "operator": "[in[",
            "property": "time"
          }
        }
      ]
    }
  },
  "configuration": {
    "operations_as_events": True,
    "learning_period": 1500000,
    "max_training_samples": 15000,
    "context": {
      "light": {
        "type": "enum"
      },
      "tz": {
        "type": "timezone"
      },
      "movement": {
        "type": "continuous"
      },
      "time": {
        "type": "time_of_day",
        "is_generated": True
      }
    },
    "output": [
      "light"
    ],
    "filter": [
      "smarthome"
    ]
  }
}
"""

Get decision using a decision tree for a generator

const CONTEXT_OPS = {
  "tz": "+02:00",
  "movement": 2,
  "time": 7.5
};
const DECISION_TREE_TIMESTAMP = 1469473600;
const GENERATOR_NAME = 'smarthome_gen';

client.computeGeneratorDecision(
  GENERATOR_NAME, # The name of the generator
  DECISION_TREE_TIMESTAMP, # The timestamp at which the decision tree is retrieved
  CONTEXT_OPS # A valid context operation according to the generator configuration
)
"""
{
  "_version": "2.0.0",
  "context": {
    "tz": "+02:00",
    "movement": 2,
    "time": 7.5
  },
  "output": {
    "light": {
      "predicted_value": "OFF",
      "confidence": 0.8386044502258301,
      "decision_rules": [
        {
          "operand": [
            2.1166666,
            10.333333
          ],
          "operator": "[in[",
          "property": "time"
        },
        {
          "operand": [
            2.1166666,
            9.3
          ],
          "operator": "[in[",
          "property": "time"
        },
        {
          "operand": [
            2.1166666,
            8.883333
          ],
          "operator": "[in[",
          "property": "time"
        },
        {
          "operand": [
            3.5333333,
            8.883333
          ],
          "operator": "[in[",
          "property": "time"
        }
      ],
      "nb_samples": 442,
      "decision_path": "0-0-0-0-1",
      "distribution": [
        0.85067874,
        0.14932127
      ]
    }
  }
}
"""

Bulk

The craft ai API includes a bulk route which provides a programmatic option to perform multiple operations at once.

:warning: the bulk API comes on top of the basic routes described above, and requires an understanding of what they do. For more information, please refer to the basic routes that do the same operations one at a time.

Bulk - Create agents

To create several agents at once, use the method create_agents_bulk as the following:

agent_id_1 = 'my_first_agent'
agent_id_2 = 'my_second_agent'

configuration_1 = {
  "context": {
    "peopleCount": {
      "type": "continuous"
    },
    "timeOfDay": {
      "type": "time_of_day"
    },
    "timezone": {
      "type": "timezone"
    },
    "lightbulbState": {
      "type": "enum"
    }
  },
  "output": ["lightbulbState"]
}
configuration_2 = { ... }

creation_bulk_payload = [
  {'id': agent_id_1, 'configuration': configuration_1},
  {'id': agent_id_2, 'configuration': configuration_2}
]

created_agents = client.create_agents_bulk(creation_bulk_payload)
print(created_agents)

The variable created_agents is an array of responses. If an agent has been successfully created, the corresponding response is an object similar to the classic create_agent() response. When there are mixed results, created_agents should looks like:

[
  {'_version': '2.0.0',                                 # creation succeeded
   'configuration': {
      'context': {
        ...
      },
      'output': ...,
      'learning_period': 1500000,
      'time_quantum': 100
   },
   'id': 'my_first_agent'},
  {'error': CraftAiBadRequestError('error-message'),    # creation failed
   'id': 'my_second_agent'
  }
]

Bulk - Delete agents

To delete several agents at once, use the method delete_agents_bulk as the following:

agent_id_1 = 'my_first_agent'
agent_id_2 = 'my_second_agent'

deletion_bulk_payload = [
  {'id': agent_id_1},
  {'id': agent_id_2}
]

deleted_agents = client.delete_agents_bulk(deletion_bulk_payload)
print(agents_deleted)

The variable deleted_agents is an array of responses. If an agent has been successfully deleted, the corresponding response is an object similar to the classic delete_agent() response. When there are mixed results, deleted_agents should looks like:

[
  {'id': 'my_first_agent',                              # deletion succeeded
   'creationDate': 1557492944277,
   'lastContextUpdate': 1557492944277,
   'lastTreeUpdate': 1557492944277,
   'configuration': {
      'context': {
        ...
      },
      'output': ...,
      'learning_period': 1500000,
      'time_quantum': 100
   },
   '_version': '2.0.0'
  },
  {'error': CraftAiBadRequestError('error-message'),    # deletion failed
   'id': 'my_second_agent'
  },
  {'id': 'my_unknown_agent'}                            # deletion succeeded
]

Bulk - Add context operations

To add operations to several agents at once, use the method add_agents_operations_bulk as the following:

agent_id_1 = 'my_first_agent'
agent_id_2 = 'my_second_agent'

operations_agent_1 = [
  {
    'timestamp': 1469410200,
    'context': {
      'timezone': '+02:00',
      'peopleCount': 0,
      'lightbulbState': 'OFF'
    }
  },
  {
    'timestamp': 1469410200,
    'context': {
      'timezone': '+02:00',
      'peopleCount': 1,
      'lightbulbState': 'ON'
    }
  },
  {
    'timestamp': 1469410200,
    'context': {
      'timezone': '+02:00',
      'peopleCount': 2,
      'lightbulbState': 'ON'
    }
  },
  {
    'timestamp': 1469410200,
    'context': {
      'timezone': '+02:00',
      'peopleCount': 2,
      'lightbulbState': 'OFF'
    }
  }
]
operations_agent_2 = [ ... ]

addition_operations_bulk_payload = [
  {'id': agent_id_1, 'operations': operations_agent_1},
  {'id': agent_id_2, 'operations': operations_agent_2}
]

agents = client.addAgentContextOperationsBulk(addition_operations_bulk_payload)

The variable agents is an array of responses. If an agent has successfully received operations, the corresponding response is an object similar to the classic add_agent_operations() response. When there are mixed results, agents should looks like:

[
  {
    'status': 201,                                # Addition of operation succeeded
    'message': 'message',
    'id': 'my_first_agent'
  }
  {
    'status': 500,                                 # Addition of operation failed
    'error': CraftAiBadRequestError('error_message'),
    'id': 'my_second_agent'
  }
]

Bulk - Compute decision trees for agents

To get several decision trees of agents at once, use the method get_agents_decision_trees_bulk as the following:

agent_id_1 = 'my_first_agent'
agent_id_2 = 'my_second_agent'

decision_tree_bulk_payload =  [
  {'id': agent_id_1},
  {'id': agent_id_2}
]

trees = client.get_agents_decision_trees_bulk(decision_tree_bulk_payload)

The variable trees is an array of responses. If a decision trees has successfully been retrieved, the corresponding response is an object similar to the classic get_agent_decision_tree() response. When there are mixed results, trees should looks like:

[
  {'id': 'my_first_agent',                              # Getting of the tree succeeded
   'tree': {
     'trees': { ... }
     '_version': '1.1.0',
     'configuration': { ... }
   }
   'timestamp': 1458741735
   },
   {
   'error': CraftAiBadRequestError('error_message'),  # Getting of the tree failed
   'id': 'my_second_agent'
   }
   {
   'error': CraftAiNotFoundError('error_message'),    # Getting of the tree failed
   'id': 'my_unknown_agent'
   }
]

Bulk - Compute boosting decisions for agents

To fetch several boosting predictions at once for agents, use the method get_agent_bulk_boosting_decision as the following:

request_payload = [
  {
    "entityName": "my_first_agent",
    "timeWindow": [1469415600, 1679415800],
    "context": {
      "peopleCount": 19,
      "timeOfDay": 7.5,
      "timezone": "+02:00"
    }
  },
  {
    "entityName": "my_first_agent",
    "timeWindow": [1469415600, 1679415800],
    "context": {
      "peopleCount": 21,
      "timeOfDay": 5,
      "timezone": "+02:00"
    }
  },
  {
    "entityName": "my_second_agent",
    "timeWindow": [1469415600, 1679415800],
    "context": {
      "peopleCount": 33,
      "timeOfDay": 8,
      "timezone": "+01:00"
    }
  }
]

decisions = client.get_agent_bulk_boosting_decision(request_payload)

The variable decisions is an array of responses. If a decision was successfully received, the corresponding response is an object similar to the classic get_agent_boosting_decision() response. When there are mixed results, decisions should looks like:

[
  {
    "context": {
      "tz": "+02:00",
      "movement": 2,
      "time": 7.5
    },
    "output": {
      "predicted_value": "OFF"
    }
  },
  {
    "error": CraftAiBadRequestError('error_message')
  }
]

Bulk - Create generators

To create several generators at once, use the method create_generators_bulk as the following:

generator_id_1 = "my_first_generator"
generator_configuration_1 = {
  "context": {
    "light": {
      "type": "enum"
    },
    "tz": {
      "type": "timezone"
    },
    "movement": {
      "type": "continuous"
    },
    "time": {
      "type": "time_of_day",
      "is_generated": True
    }
  },
  "model_type": "decisionTree",
  "output": [
    "light"
  ],
  "learning_period": 1500000,
  "max_training_samples": 15000,
  "operations_as_events": True,
  "filter": ["smarthome"]
}
generator_id_2 = "my_second_generator"
generator_configuration_2 = {
  "context": {
    "light": {
      "type": "enum"
    },
    "tz": {
      "type": "timezone"
    },
    "movement": {
      "type": "continuous"
    },
    "time": {
      "type": "time_of_day",
      "is_generated": True
    }
  },
  "model_type": "decisionTree",
  "output": [
    "light"
  ],
  "learning_period": 1500000,
  "max_training_samples": 15000,
  "operations_as_events": True,
  "filter": ["smarthome", "cleverhouse", "cunningshed"]
}

creation_bulk_payload = [
  {"id": generator_id_1, "configuration": generator_configuration_1},
  {"id": generator_id_2, "configuration": generator_configuration_2}
]

created_generators = client.create_generators_bulk(creation_bulk_payload)
print(created_generators)

The variable created_generators is an array of responses. If a generator has been successfully created, the corresponding response is an object similar to the classic create_generator() response. When there are mixed results, created_generators should looks like:

[
  {
    "_version": "2.0.0",                                 # creation succeeded
    "configuration": {
      "context": {
        ...
      },
      "output": ...,
      "learning_period": 1500000,
      "time_quantum": 100
    },
    "id": "my_first_generator"
  },
  {
    "error": CraftAiBadRequestError("error-message"),    # creation failed
    "id": "my_second_generator"
  }
]

Bulk - Delete generators

To delete several generators at once, use the method delete_generators_bulk as the following:

generator_id_1 = "my_first_generator"
generator_id_2 = "my_second_generator"

deletion_bulk_payload = [
  {"id": generator_id_1},
  {"id": generator_id_2},
  {"id": "my_unknown_generator"}
]

deleted_generators = client.delete_generators_bulk(deletion_bulk_payload)
print(deleted_generators)

The variable deleted_generators is an array of responses. If a generator has been successfully deleted, the corresponding response is an object similar to the classic delete_generator() response. When there are mixed results, deleted_generators should looks like:

[
  {
    "id": "my_first_generator",                              # deletion succeeded
    "creationDate": 1557492944277,
    "lastContextUpdate": 1557492944277,
    "lastTreeUpdate": 1557492944277,
    "configuration": {
       "context": {
         ...
       },
       "output": ...,
       "learning_period": 1500000,
       "time_quantum": 100
    },
    "_version": "2.0.0"
  },
  {
    "error": CraftAiBadRequestError("error-message"),       # deletion failed
    "id": "my_second_generator"
  },
  {
    "id": "my_unknown_generator"                            # deletion succeeded
  }
]

Bulk - Compute decision trees for generators

To get several decision trees of generators at once, use the method get_generators_decision_trees_bulk as the following:

generator_id_1 = "my_first_generator"
generator_id_2 = "my_second_generator"

decision_tree_bulk_payload =  [
  {
    "id": generator_id_1,
    "timestamp": 1458741735
  },
  {
    "id": generator_id_2,
    "timestamp": 1458741737
  }
]

trees = client.get_generators_decision_trees_bulk(decision_tree_bulk_payload)

The variable trees is an array of responses. If a generator’s decision tree has successfully been retrieved, the corresponding response is an object similar to the classic get_generator_decision_tree() response. When there are mixed results, trees should looks like:

[
  {
    "id": "my_first_generator",                              # Getting of the tree succeeded
    "tree": {
      "trees": { ... }
      "_version": "1.1.0",
      "configuration": { ... }
    }
    "timestamp": 1458741735
   },
   {
     "error": CraftAiBadRequestError("error_message"),  # Getting of the tree failed
     "id": "my_second_generator"
   }
]

Bulk - Compute boosting decisions for generators

To fetch several boosting predictions at once for generators, use the method get_generator_bulk_boosting_decision as the following:

request_payload = [
  {
    "entityName": "my_firstgenerator",
    "timeWindow": [1469415600, 1679415800],
    "context": {
      "peopleCount": 19,
      "timeOfDay": 7.5,
      "timezone": "+02:00"
    }
  },
  {
    "entityName": "my_firstgenerator",
    "timeWindow": [1469415600, 1679415800],
    "context": {
      "peopleCount": 21,
      "timeOfDay": 5,
      "timezone": "+02:00"
    }
  },
  {
    "entityName": "my_secondgenerator",
    "timeWindow": [1469415600, 1679415800],
    "context": {
      "peopleCount": 33,
      "timeOfDay": 8,
      "timezone": "+01:00"
    }
  }
]

decisions = client.get_generator_bulk_boosting_decision(request_payload)

The variable decisions is an array of responses. If a decision was successfully received, the corresponding response is an object similar to the classic get_generator_boosting_decision() response. When there are mixed results, decisions should looks like:

[
  {
    "context": {
      "tz": "+02:00",
      "movement": 2,
      "time": 7.5
    },
    "output": {
      "predicted_value": "OFF"
    }
  },
  {
    "error": CraftAiBadRequestError("error_message")
  }
]

Advanced client configuration

The simple configuration to create the client is just the token. For special needs, additional advanced configuration can be provided.

Amount of operations sent in one chunk

client.add_agent_operations and client.decide_boosting_from_contexts_df split the provided operations into chunks in order to limit the size of the http requests to the craft ai API. In the client configuration, operationsChunksSize can be increased in order to reduce the number of requests, or decreased when large http requests cause errors.

client = craft_ai.Client({
    # Mandatory, the token
    "token": "{token}",
    # Optional, default value is 200
    "operationsChunksSize": {max_number_of_operations_sent_at_once}
})

Timeout duration for decision trees retrieval

It is possible to increase or decrease the timeout duration of client.get_agent_decision_tree, for exemple to account for especially long computations.

client = craft_ai.Client({
    # Mandatory, the token
    "token": "{token}",
    # Optional, default value is 600000 (10 minutes)
    "decisionTreeRetrievalTimeout": "{timeout_duration_for_decision_trees_retrieval}"
})

Proxy

It is possible to provide proxy configuration in the proxy property of the client configuration. It will be used to call the craft ai API (through HTTPS). The expected format is a host name or IP and port, optionally preceded by credentials such as http://user:pass@10.10.1.10:1080.

client = craft_ai.Client({
    # Mandatory, the token
    "token": "{token}",
    # Optional, no default value
    "proxy": "http://{user}:{password}@{host_or_ip}:{port}"
})

Advanced network configuration

For more advanced network configuration, it is possible to access the Requests Session used by the client to send requests to the craft ai API, through client._requests_session.

# Disable SSL certificate verification
client._requests_session.verify = False

Score

The following functions let you compute model scores.

:warning: At the moment, this is only available as bulk functions, for generators set to generate decision trees.

Score - Sliding window

client.get_sliding_window_scores_bulk(body)

Body

The body should be an array of dictionaries containing the following keys:

• id string (required)

  The identifier of the generator whose model is evaluated.

• test_from number

  The beginning timestamp of the first test window (inclusive). 3 parameters among test_from, test_to, step_size and nb_steps must be defined.

• test_to number

  The end timestamp of the last test window (inclusive). 3 parameters among test_from, test_to, step_size and nb_steps must be defined.

• step_size number

  The timestamp difference between the beginning of a test window and the next. 3 parameters among test_from, test_to, step and nb_steps must be defined.

• nb_steps number

  The number of test windows. 3 parameters among test_from, test_to, step_size and nb_steps must be defined.

• test_size number

  The actual size of the test set from the beginning of a test window. If the size of a window (step_size) is larger than this, only the beginning of a window will be used to compute scores. If the size of a window is smaller than this, data in a window can be used in several score computations.

• gap_size number

  The timestamp difference between the end of a training window (data used in the model) and the beginning of a test window. The end of the training window is defined by the formula testWindowStart - gap_size - 1. By default gap_size is 0, in which case the test set starts directly after the end of the training set, i.e. the end of the training set is just before the beginning of the test window. A negative gap means that there is an overlap between the training and test data.

• metrics array

  Array of objects containing a name property with the name of a valid metric. The metrics are used to evaluate the ML model. For classification models, the available metrics are accuracy and f1; for regression models, r2, mae and rmse. By default all available metrics are computed.

+ training data
* test data
                   gap_size test_size
                      <-><------------>
window 1  .+++++++++++...**************...................................
                         |
              window 1 start (test_from)
                         <--------->
                          step_size
window 2  ............+++++++++++...**************........................
                                    |
                             window 2 start
                                    <--------->
                                     step_size
window 3  .......................+++++++++++...**************.............
                                               |
                                         window 3 start

Example:

  const sliding_window_scores_request_payload = [
    {
      "id": "generator1",
      "test_from": 1461132001,
      "test_to": 1462106220,
      "step_size": 500000,
      "metrics": [{ "name": "accuracy" }, { "name": "f1" }]
    },
    {
      "id": "generator2",
      "test_from": 1477000801,
      "test_to": 1485385200,
      "step_size": 5000000,
      "metrics": [{ "name": "r2" }, { "name": "mae" }, { "name": "rmse" }]
    }
  ]

  scores = client.get_sliding_window_scores_bulk(sliding_window_scores_request_payload)

The variable scores is an array of responses. scores should look like:

[
  {
    "id": "generator1",
    "scores": [
      {
        "from": 1461132001,
        "to": 1461632000,
        "modelTimestamp": 1461132000,
        "nbSamples": 15,
        "type": "classification",
        "accuracy": 0.467,
        "f1": {
          "class_OPEN": {
            "nbSamples": 7,
            "score": 0.636
          }
        },
        "f1_weighted": 0.297
      },
      {
        "from": 1461632001,
        "to": 1462106220,
        "modelTimestamp": 1461632000,
        "nbSamples": 14,
        "type": "classification",
        "accuracy": 0.786,
        "f1": {
          "class_CLOSED": {
            "nbSamples": 7,
            "score": 0.8
          },
          "class_OPEN": {
            "nbSamples": 7,
            "score": 0.769
          }
        },
        "f1_weighted": 0.785
      }
    ]
  },
  {
    "id": "generator2",
    "scores": [
      {
        "from": 1477000801,
        "to": 1482000800,
        "modelTimestamp": 1477000800,
        "nbSamples": 57,
        "type": "regression",
        "r2": -0.864,
        "mae": 2.186,
        "rmse": 2.489
      },
      {
        "from": 1482000801,
        "to": 1485385200,
        "modelTimestamp": 1482000800,
        "nbSamples": 39,
        "type": "regression",
        "r2": 0.52,
        "mae": 0.95,
        "rmse": 1.164
      }
    ]
  }
]

Score - Single window

client.get_single_window_score_bulk(body)

Body

The body should be an array of dictionaries containing the following keys:

• id string (required)

  The identifier of the generator whose model is evaluated.

• test_from number (required)

  The beginning timestamp of the test window (inclusive).

• test_to number (required)

  The end timestamp of the test window (inclusive).

• model_timestamp number

  The last timestamp of the training data.

• metrics array

  Array of objects containing a name property with the name of a valid metric. The metrics are used to evaluate the ML model. For classification models, the available metrics are accuracy and f1; for regression models, r2, mae and rmse. By default all available metrics are computed.

+ training data
* test data
            model_timestamp    test_from            test_to
window  ++++++++++++|..............|*******************|..................

Example:

  const single_window_score_request_payload = [
    {
      "id": "generator1",
      "test_from": 1461132001,
      "test_to": 1462106220,
      "step_size": 500000,
      "metrics": [{ "name": "accuracy" }, { "name": "f1" }]
    },
    {
      "id": "generator2",
      "test_from": 1477000801,
      "test_to": 1485385200,
      "step_size": 5000000,
      "metrics": [{ "name": "r2" }, { "name": "mae" }, { "name": "rmse" }]
    }
  ]

  scores = client.get_single_window_score_bulk(single_window_score_request_payload)

The variable scores is an array of responses. scores should look like:

[
  {
    "id": "generator1",
    "score": {
      "from": 1461132001,
      "to": 1461632000,
      "modelTimestamp": 1461132000,
      "nbSamples": 15,
      "type": "classification",
      "accuracy": 0.467,
      "f1": {
        "class_OPEN": {
          "nbSamples": 7,
          "score": 0.636
        }
      },
      "f1_weighted": 0.297
    }
  },
  {
    "id": "generator2",
    "score": {
      "from": 1477000801,
      "to": 1485385200,
      "modelTimestamp": 1477000800,
      "nbSamples": 96,
      "type": "regression",
      "r2": -0.761,
      "mae": 2.06,
      "rmse": 2.356
    }
  }
]

Interpreter

The decision tree interpreter can be used offline from decisions tree computed through the API.

Make decision

Note that the python interpreter takes an array of contexts.

tree = { ... } # Decision tree as retrieved through the craft ai REST API

# Compute the decision on a fully described context
decision = craft_ai.Interpreter.decide(
  tree,
  [{ # The context on which the decision is made
    "timezone": "+02:00",
    "timeOfDay": 7.5,
    "peopleCount": 3
  }]
)

# Or Compute the decision on a context created from the given one and filling the
# `day_of_week`, `time_of_day` and `timezone` properties from the given `Time`

decision = craft_ai.Interpreter.decide(
  tree,
  [{
    "timezone": "+02:00",
    "peopleCount": 3
  },
  craft_ai.Time("2010-01-01T07:30:30+0200")
  ]
)

A computed decision on an enum output type would look like:

{
  "context": { # In which context the decision was made
    "timezone": "+02:00",
    "timeOfDay": 7.5,
    "peopleCount": 3
  },
  "output": { # The decision itself
    "lightbulbState": {
      "predicted_value": "ON"
      "confidence": 0.9937745256361138, # The confidence in the decision
      "decision_rules": [ # The ordered list of decision_rules that were validated to reach this decision
        {
          "property": "timeOfDay",
          "operator": ">=",
          "operand": 6
        },
        {
          "property": "peopleCount",
          "operator": ">=",
          "operand": 2
        }
      ],
      "nb_samples": 25,
      "distribution": [0.05, 0.95],
      "decision_path" "0-1-1"
    },
  }
}

A decision for a numerical output type would look like:

  "output": {
    "lightbulbIntensity": {
      "predicted_value": 10.5,
      "standard_deviation": 1.25, # For numerical types, this field is returned in decisions.
      "min": 8.0,
      "max": 11,
      "nb_samples": 25,
      "decision_rules": [ ... ],
      "confidence": ...,
      "decision_path" ...
    }
  }

A decision for a categorical output type would look like:

  "output": {
    "lightbulbState": {
      "predicted_value": "OFF",
      "nb_samples": 25,
      "distribution" : [ ... ], # Distribution of the output classes normalized by the number of samples in the reached node.
      "decision_rules": [ ... ],
      "confidence": ...,
      "decision_path" ...
    }
  }

A decision in a case where the tree cannot make a prediction:

  "output": {
    "lightbulbState": {
      "predicted_value": None,
      "distribution" : [ ... ], # Distribution of the output classes normalized by the number of samples in the reached node.
      "confidence": 0, # Zero confidence if the decision is null
      "decision_rules": [ ... ],
      "decision_path" ...
    }
  }

Reduce decision rules

From a list of decision rules, as retrieved when making a decision with a decision tree, compute an equivalent & minimal list of rules.

from craft_ai import reduce_decision_rules

# `decision` is the decision tree as retrieved from taking a decision
decision = craft_ai.Interpreter.decide( ... )

# `decision_rules` is the decision rules that led to decision for the `lightBulbState` value
decision_rules = decision["output"]["lightBulbState"]["decision_rules"]

# `minimal_decision_rules` has the mininum list of rules strictly equivalent to `decision_rules`
minimal_decision_rules = reduce_decision_rules(decisionRules)

Format decision rules

From a list of decision rules, compute a human readable version of these rules, in english.

from craft_ai import format_decision_rules

# `decision` is the decision tree as retrieved from taking a decision
decision = craft_ai.Interpreter.decide( ... )

# `decision_rules` is the decision rules that led to decision for the `lightBulbState` value
decision_rules = decision["output"]["lightbulbState"]["decision_rules"]

# `decision_rules_str` is a human readable string representation of the rules.
decision_rules_str = format_decision_rules(decision_rules)

Error Handling

When using this client, you should be careful wrapping calls to the API with try/except blocks, in accordance with the EAFP principle.

The craft ai python client has its specific exception types, all of them inheriting from the CraftAIError type.

All methods which have to send an http request (all of them except decide) may raise either of these exceptions: CraftAINotFoundError, CraftAIBadRequestError, CraftAICredentialsError or CraftAIUnknownError.

The decide method only raises CrafAIDecisionError of CraftAiNullDecisionError type of exceptions. The latter is raised when no the given context is valid but no decision can be made.

Pandas support

The craft ai python client optionally supports pandas a very popular library used for all things data.

You'll need to install craft-ai with its pandas extra

pip install --upgrade craft-ai[pandas]

Then, instead of importing the default module, do the following

import craft_ai.pandas

# Most of the time you'll need the following
import numpy as np
import pandas as pd

# Client must then be defined using craft_ai.pandas module
client = craft_ai.pandas.Client({
  "token": "{token}"
})

The craft ai pandas module is derived for the vanilla one, with the following methods are overriden to support pandas' DataFrame.

craft_ai.pandas.Client.get_agent_operations

Retrieves the desired operations as a DataFrame where:

• each operation is a row,
• each context property is a column,
• the index is time based, timezone-aware and matching the operations timestamps,
• np.NaN means no value were given at this property for this timestamp.

df = client.get_agent_operations("my_new_agent")

# `df` is a pd.DataFrame looking like
#
#                            peopleCount  lightbulbState   timezone
# 2013-01-01 00:00:00+00:00   0            OFF              +02:00
# 2013-01-02 00:00:00+00:00   1            ON               NaN
# 2013-01-03 00:00:00+00:00   2            NaN              NaN
# 2013-01-04 00:00:00+00:00   NaN          OFF              NaN
# 2013-01-05 00:00:00+00:00   0            NaN              NaN

craft_ai.pandas.Client.add_agent_operations

Add a DataFrame of operations to the desired agent. The format is the same as above.

df = pd.DataFrame(
  [
    [0, "OFF", "+02:00"],
    [1, "ON", np.nan], # timezone will be "+02:00"
    [2, np.nan, np.nan],
    [np.nan, "OFF", np.nan],
    [0, np.nan, np.nan]
  ],
  columns=['peopleCount', 'lightbulbState', 'timezone'],
  index=pd.date_range('20130101', periods=5, freq='D').tz_localize("UTC")
)
client.add_agent_operations("my_new_agent", df)

Given an object that is not a DataFrame this method behave like the vanilla craft_ai.Client.add_agent_operations.

Furthermore, missing values and optional values can be handled by the craft ai pandas client. To do so, we introduce two new types that are craft_ai.pandas.MISSING_VALUE for missing values and craft_ai.pandas.OPTIONAL_VALUE for optional values. To send your DataFrame with actual missing values or optional values, you must use one of these types:

from craft_ai.pandas import MISSING_VALUE, OPTIONAL_VALUE

df = pd.DataFrame(
  [
    [0, "+02:00"],
    [1, MISSING_VALUE],
    [2, MISSING_VALUE],
    [1, OPTIONAL_VALUE],
    [0, np.nan]
  ],
  columns=['peopleCount', 'timezone'],
  index=pd.date_range('20130101', periods=5, freq='D').tz_localize("UTC")
)
client.add_agent_operations("my_new_agent", df)

To ensure that all the missing values contained in your DataFrame are to the right format and can be handled by the craft ai pandas client, it is suggested to preprocess this latter by replacing all na values by the desired one:

contexts_df.fillna(MISSING_VALUE) # Or OPTIONAL_VALUE

craft_ai.pandas.Client.get_agent_states

Retrieves the desired state history as a DataFrame where:

• each state is a row,
• each context property is a column,
• the index is time based, timezone-aware and matching the operations timestamps.

df = client.get_agent_states("my_new_agent")

# `df` is a pd.DataFrame looking like
#
#                            peopleCount  lightbulbState   timezone
# 2013-01-01 00:00:00+00:00   0            OFF              +02:00
# 2013-01-02 00:00:00+00:00   1            ON               +02:00
# 2013-01-03 00:00:00+00:00   2            ON               +02:00
# 2013-01-04 00:00:00+00:00   2            OFF              +02:00
# 2013-01-05 00:00:00+00:00   0            OFF              +02:00

craft_ai.pandas.Client.decide_from_contexts_df

Make multiple decisions on a given DataFrame following the same format as above.

decisions_df = client.decide_from_contexts_df(tree, pd.DataFrame(
  [
    [0, "+02:00"],
    [1, "+02:00"],
    [2, "+02:00"],
    [1, "+02:00"],
    [0, "+02:00"]
  ],
  columns=['peopleCount', 'timezone'],
  index=pd.date_range('20130101', periods=5, freq='D').tz_localize("UTC")
))
# `decisions_df` is a pd.DataFrame looking like
#
#                            lightbulbState_predicted_value   lightbulbState_confidence  ...
# 2013-01-01 00:00:00+00:00   OFF                              0.999449                  ...
# 2013-01-02 00:00:00+00:00   ON                               0.970325                  ...
# 2013-01-03 00:00:00+00:00   ON                               0.970325                  ...
# 2013-01-04 00:00:00+00:00   ON                               0.970325                  ...
# 2013-01-05 00:00:00+00:00   OFF                              0.999449                  ...

This function also accepts craft ai missing values and optional values types, craft_ai.pandas.MISSING_VALUE and craft_ai.pandas.OPTIONAL_VALUE.

from craft_ai.pandas import MISSING_VALUE, OPTIONAL_VALUE

decisions_df = client.decide_from_contexts_df(tree, pd.DataFrame(
  [
    [0, "+02:00"],
    [MISSING_VALUE, "+02:00"],
    [2, "+02:00"],
    [MISSING_VALUE, "+02:00"],
    [0, "+02:00"]
  ],
  columns=['peopleCount', 'timezone'],
  index=pd.date_range('20130101', periods=5, freq='D').tz_localize("UTC")
))

This function never raises CraftAiNullDecisionError, instead it inserts these errors in the result Dataframe in a specific error column.

craft_ai.pandas.utils.create_tree_html

Returns a HTML version of the given decision tree. If this latter is saved in a .html file, it can be opened in a browser to be visualized.

from  craft_ai.pandas.utils import create_tree_html

tree = client.get_agent_decision_tree(
  "my_agent", # The agent id
  timestamp # The timestamp at which the decision tree is retrieved
)

html = create_tree_html(
  tree, # The decision tree
  decision_path, # (Optional) The path to a node. This will plot the tree with this node already selected. Default: ""
  edge_type, # (Optional) The way the decision tree edges are plotted - ("constant", "absolute" or "relative"). Default: "constant"
  folded_nodes, # (Optional) An array of nodes path to fold when the tree is plotted. Default: None
  height # (Optional) The height in pixel of the created plot. Default: 500.
)

# ...
# ... save the html string to visualize it in a browser
# ...

craft_ai.pandas.utils.display_tree

Display a decision tree in a Jupyter Notebook. This function can be useful for analyzing the induced decision trees.

from  craft_ai.pandas.utils import display_tree

tree = client.get_agent_decision_tree(
  "my_agent", # The agent id
  timestamp # The timestamp at which the decision tree is retrieved
)

display_tree(
  tree, # The decision tree
  decision_path, # (Optional) The path to a node. This will plot the tree with this node already selected. Default: ""
  edge_type, # (Optional) The way the decision tree edges are plotted - ("constant", "absolute" or "relative"). Default: "constant"
  folded_nodes, # (Optional) An array of nodes path to fold when the tree is plotted. Default: None
  height # (Optional) The height in pixel of the created plot. Default: 500.
)

craft_ai.pandas.client.add_agents_operations_bulk

Add operations to several agents at once.

agent_id_1 = 'my_first_agent'
agent_id_2 = 'my_second_agent'

operations_agent_1 = pd.DataFrame(
  [
    [0, "OFF", "+02:00"],
    [1, "ON", np.nan], # timezone will be "+02:00"
    [2, np.nan, np.nan],
    [np.nan, "OFF", np.nan],
    [0, np.nan, np.nan]
  ],
  columns=['peopleCount', 'lightbulbState', 'timezone'],
  index=pd.date_range('20130101', periods=5, freq='D').tz_localize("UTC")
)
operations_agent_2 = pd.DataFrame([...])

addition_operations_bulk_payload = [
    {'id': agent_id_1, 'operations': operations_agent_1},
    {'id': agent_id_2, 'operations': operations_agent_2}
]

client.add_agents_operations_bulk(addition_operations_bulk_payload)

Given an object that is not a DataFrame this method behave like the vanilla craft_ai.Client.add_agents_operations_bulk.

craft_ai.pandas.client.decide_boosting_from_contexts_df

Make multiple boosting decisions on a given DataFrame on an agent.

agent_id_1 = 'my_first_agent'
FROM_TIMESTAMP = 1469473600
TO_TIMESTAMP = 1529473600

context_df = pd.DataFrame(
  [
    [0, "+02:00"],
    [1, "+02:00"],
    [2, "+02:00"],
    [1, "+02:00"],
    [0, "+02:00"]
  ],
  columns=['peopleCount', 'timezone'],
  index=pd.date_range('20130101', periods=5, freq='D').tz_localize("UTC")
)

decisions_df = CLIENT.decide_boosting_from_contexts_df(
    agent_id_1,
    FROM_TIMESTAMP,
    TO_TIMESTAMP,
    context_df,
)

# `decisions_df` is a pd.DataFrame looking like
#
#                            lightbulbState_predicted_value
# 2013-01-01 00:00:00+00:00   OFF
# 2013-01-02 00:00:00+00:00   ON
# 2013-01-03 00:00:00+00:00   ON
# 2013-01-04 00:00:00+00:00   ON
# 2013-01-05 00:00:00+00:00   OFF

craft_ai.pandas.client.decide_generator_boosting_from_contexts_df

Make multiple boosting decisions on a given DataFrame on a generator.

generator_id = 'my_generator'

FROM_TIMESTAMP = 1469473600
TO_TIMESTAMP = 1529473600

context_df = pd.DataFrame(
  [
    [0, "+02:00"],
    [1, "+02:00"],
    [2, "+02:00"],
    [1, "+02:00"],
    [0, "+02:00"]
  ],
  columns=['peopleCount', 'timezone'],
  index=pd.date_range('20130101', periods=5, freq='D').tz_localize("UTC")
)

decisions = CLIENT.decide_generator_boosting_from_contexts_df(
    generator_id,
    FROM_TIMESTAMP,
    TO_TIMESTAMP,
    context_df,
)

# `decisions_df` is a pd.DataFrame looking like
#
#                            lightbulbState_predicted_value
# 2013-01-01 00:00:00+00:00   OFF
# 2013-01-02 00:00:00+00:00   ON
# 2013-01-03 00:00:00+00:00   ON
# 2013-01-04 00:00:00+00:00   ON
# 2013-01-05 00:00:00+00:00   OFF