Note
Go to the end to download the full example code
Creating and Using a Custom Flow¶
The most convenient way to create a flow for your machine learning workflow is to generate it automatically as described in the Obtaining Flow IDs tutorial. However, there are scenarios where this is not possible, such as when the flow uses a framework without an extension or when the flow is described by a script.
In those cases you can still create a custom flow by following the steps of this tutorial. As an example we will use the flows generated for the AutoML Benchmark, and also show how to link runs to the custom flow.
# License: BSD 3-Clause
from collections import OrderedDict
import numpy as np
import openml
from openml import OpenMLClassificationTask
from openml.runs.functions import format_prediction
Warning
This example uploads data. For that reason, this example connects to the test server at test.openml.org. This prevents the main server from crowding with example datasets, tasks, runs, and so on. The use of this test server can affect behaviour and performance of the OpenML-Python API.
openml.config.start_using_configuration_for_example()
1. Defining the flow¶
The first step is to define all the hyperparameters of your flow.
The API pages feature a descriptions of each variable of the openml.flows.OpenMLFlow.
Note that external version and name together uniquely identify a flow.
The AutoML Benchmark runs AutoML systems across a range of tasks. OpenML stores Flows for each AutoML system. However, the AutoML benchmark adds preprocessing to the flow, so should be described in a new flow.
We will break down the flow arguments into several groups, for the tutorial. First we will define the name and version information. Make sure to leave enough information so others can determine exactly which version of the package/script is used. Use tags so users can find your flow easily.
general = dict(
name="automlbenchmark_autosklearn",
description=(
"Auto-sklearn as set up by the AutoML Benchmark"
"Source: https://github.com/openml/automlbenchmark/releases/tag/v0.9"
),
external_version="amlb==0.9",
language="English",
tags=["amlb", "benchmark", "study_218"],
dependencies="amlb==0.9",
)
Next we define the flow hyperparameters. We define their name and default value in parameters, and provide meta-data for each hyperparameter through parameters_meta_info. Note that even though the argument name is parameters they describe the hyperparameters. The use of ordered dicts is required.
flow_hyperparameters = dict(
parameters=OrderedDict(time="240", memory="32", cores="8"),
parameters_meta_info=OrderedDict(
cores=OrderedDict(description="number of available cores", data_type="int"),
memory=OrderedDict(description="memory in gigabytes", data_type="int"),
time=OrderedDict(description="time in minutes", data_type="int"),
),
)
It is possible to build a flow which uses other flows. For example, the Random Forest Classifier is a flow, but you could also construct a flow which uses a Random Forest Classifier in a ML pipeline. When constructing the pipeline flow, you can use the Random Forest Classifier flow as a subflow. It allows for all hyperparameters of the Random Classifier Flow to also be specified in your pipeline flow.
Note: you can currently only specific one subflow as part of the components.
In this example, the auto-sklearn flow is a subflow: the auto-sklearn flow is entirely executed as part of this flow. This allows people to specify auto-sklearn hyperparameters used in this flow. In general, using a subflow is not required.
Note: flow 9313 is not actually the right flow on the test server, but that does not matter for this demonstration.
autosklearn_flow = openml.flows.get_flow(9313) # auto-sklearn 0.5.1
subflow = dict(
components=OrderedDict(automl_tool=autosklearn_flow),
# If you do not want to reference a subflow, you can use the following:
# components=OrderedDict(),
)
With all parameters of the flow defined, we can now initialize the OpenMLFlow and publish. Because we provided all the details already, we do not need to provide a model to the flow.
In our case, we don’t even have a model. It is possible to have a model but still require to follow these steps when the model (python object) does not have an extensions from which to automatically extract the hyperparameters. So whether you have a model with no extension or no model at all, explicitly set the model of the flow to None.
autosklearn_amlb_flow = openml.flows.OpenMLFlow(
**general,
**flow_hyperparameters,
**subflow,
model=None,
)
autosklearn_amlb_flow.publish()
print(f"autosklearn flow created: {autosklearn_amlb_flow.flow_id}")
2. Using the flow¶
This Section will show how to upload run data for your custom flow. Take care to change the values of parameters as well as the task id, to reflect the actual run. Task and parameter values in the example are fictional.
flow_id = autosklearn_amlb_flow.flow_id
parameters = [
OrderedDict([("oml:name", "cores"), ("oml:value", 4), ("oml:component", flow_id)]),
OrderedDict([("oml:name", "memory"), ("oml:value", 16), ("oml:component", flow_id)]),
OrderedDict([("oml:name", "time"), ("oml:value", 120), ("oml:component", flow_id)]),
]
task_id = 1200 # Iris Task
task = openml.tasks.get_task(task_id)
dataset_id = task.get_dataset().dataset_id
The last bit of information for the run we need are the predicted values. The exact format of the predictions will depend on the task.
The predictions should always be a list of lists, each list should contain:
the repeat number: for repeated evaluation strategies. (e.g. repeated cross-validation)
the fold number: for cross-validation. (what should this be for holdout?)
0: this field is for backward compatibility.
index: the row (of the original dataset) for which the prediction was made.
p_1, …, p_c: for each class the predicted probability of the sample belonging to that class. (no elements for regression tasks) Make sure the order of these elements follows the order of task.class_labels.
the predicted class/value for the sample
the true class/value for the sample
When using openml-python extensions (such as through run_model_on_task), all of this formatting is automatic. Unfortunately we can not automate this procedure for custom flows, which means a little additional effort is required.
Here we generated some random predictions in place. You can ignore this code, or use it to better understand the formatting of the predictions.
Find the repeats/folds for this task:
n_repeats, n_folds, _ = task.get_split_dimensions()
all_test_indices = [
(repeat, fold, index)
for repeat in range(n_repeats)
for fold in range(n_folds)
for index in task.get_train_test_split_indices(fold, repeat)[1]
]
# random class probabilities (Iris has 150 samples and 3 classes):
r = np.random.rand(150 * n_repeats, 3)
# scale the random values so that the probabilities of each sample sum to 1:
y_proba = r / r.sum(axis=1).reshape(-1, 1)
y_pred = y_proba.argmax(axis=1)
class_map = dict(zip(range(3), task.class_labels))
_, y_true = task.get_X_and_y()
y_true = [class_map[y] for y in y_true]
# We format the predictions with the utility function `format_prediction`.
# It will organize the relevant data in the expected format/order.
predictions = []
for where, y, yp, proba in zip(all_test_indices, y_true, y_pred, y_proba):
repeat, fold, index = where
prediction = format_prediction(
task=task,
repeat=repeat,
fold=fold,
index=index,
prediction=class_map[yp],
truth=y,
proba={c: pb for (c, pb) in zip(task.class_labels, proba)},
)
predictions.append(prediction)
Finally we can create the OpenMLRun object and upload. We use the argument setup_string because the used flow was a script.
benchmark_command = f"python3 runbenchmark.py auto-sklearn medium -m aws -t 119"
my_run = openml.runs.OpenMLRun(
task_id=task_id,
flow_id=flow_id,
dataset_id=dataset_id,
parameter_settings=parameters,
setup_string=benchmark_command,
data_content=predictions,
tags=["study_218"],
description_text="Run generated by the Custom Flow tutorial.",
)
my_run.publish()
print("run created:", my_run.run_id)
openml.config.stop_using_configuration_for_example()