Manual way

  • Step 1. Specify problem type, create the FEDOT model and load the datasets.

import pandas as pd

# specify additional automl training parameters
timeout, n_jobs, logging_level = ...  # tested with 3., 1, logging.FATAL

Hint

See the API documentation for descriptions of these and others parameters.

# build model for adjusting the pipeline
model = Fedot(
   problem='classification', timeout=timeout, n_jobs=n_jobs, logging_level=logging_level,
   seed=42
)

# add all datasets paths and load datasets
train_file_path: Union[str, os.Pathlike] = ...
validation_file_path: Union[str, os.Pathlike] = ...
# tested with default scoring classification from FEDOT's datasets

dataset_to_train = pd.read_csv(train_file_path)
dataset_to_validate = pd.read_csv(validation_file_path)

# specify target column and validation data
target_col: str = ...  # 'target' by default
validation_target = dataset_to_validate[target_col]

  • Step 2. Create Pipeline instance, i.e. create nodes with desired models.

node_first = PipelineNode('logit')
node_second = PipelineNode('xgboost')
node_final = PipelineNode('knn', nodes_from=[node_first, node_second])
pipeline = Pipeline(node_final)

The following pages provide examples of manually constructed pipelines for different tasks: classification, regression, time series forecasting.

  • Step 3. Fit the pipeline.

model.fit(features=dataset_to_train, target=target_col, predefined_model=pipeline)

Hint

If predefined_model is set to ‘auto’, FEDOT will choose and fit the default initial assumption for the task. Note that this differs from automated model creation, although it does not require manual selection of the model.

  • Step 4. Obtain a prediction and calculate the metrics.

# get the prediction
prediction = model.predict(features=dataset_to_validate)

# calculate the scores
metrics = model.get_metrics(validation_target)
print(f'metrics: {metrics}')
>>> metrics: {'roc_auc': 0.617, 'f1': 0.9205}

Eventually, we get a configured machine learning pipeline and its data-based predictions. Let’s say we’re looking for a baseline against which we can compare the result. We repeat the steps 3 and 4, using predefined_model=’auto’, and get the metric values:

'roc_auc': 0.785, 'f1': 0.934

In this case, our manually constructed pipeline outperforms the FEDOT’s first guess in terms of F1, but significantly looses in ROC AUC. Apparently, there is a better way to go.

Next, we can try to create another ML pipeline by hand and see if it gives higher scores. Or we can let FEDOT do it for us using evolutionary search. See Automated way for details.