Component Libraries¶

Component libraries let you use pre-built pipeline steps without writing custom @step code. Instead of implementing training, splitting, or charting logic yourself, you declare a component: key in your process config and the platform loads the implementation from an installed plugin package.

How It Works¶

Component libraries register themselves under the expops.component_runners Python entry-point group. When ExpOps executes a pipeline, it inspects all component: references in the config, then loads only the matching plugin packages before running the pipeline.

For example, a process with component: "sklearn.LogisticRegression.fit_predict" causes the platform to load the sklearn entry point, which registers the LogisticRegression fit+predict runner under that name.

Installing Component Packages¶

Component packages are installed into the same environment as the expops CLI:

pip install expops-sklearn      # sklearn estimators, data loading, splitting
pip install expops-matplotlib   # bar charts, histograms

The packages must be installed in the environment where you invoke expops run, not inside the project's isolated virtual environment.

Using Components in Config¶

Add a component: key to any process in your project_config.yaml:

experiment:
  pipeline:
    processes:
      - name: "split_data"
        component: "sklearn.train_test_split"
        parameters:
          test_size: 0.2

The full component string format is <library>.<ComponentClass>.<operation> (or <library>.<operation> for stateless steps like sklearn.csv_to_xy).

`input_transform`¶

Maps canonical component input names to the output names produced by upstream processes:

- name: "train_model"
  component: "sklearn.LogisticRegression.fit_predict"
  input_transform:
    X: "X_train"   # canonical input "X" ← upstream output "X_train"
    y: "y_train"   # canonical input "y" ← upstream output "y_train"

Without input_transform, the component expects its canonical input names to match upstream output names exactly.

`output_mapping`¶

Renames component outputs before they are stored for downstream processes:

- name: "predict_model"
  component: "sklearn.LogisticRegression.predict"
  input_transform:
    X: "X_test"
  output_mapping:
    predictions: "y_pred"   # canonical output "predictions" → stored as "y_pred"

`parameters`¶

Passes keyword arguments to the component (hyperparameters, chart settings, etc.):

- name: "train_model"
  component: "sklearn.LogisticRegression.fit_predict"
  parameters:
    max_iter: 200
    C: 0.5

`metrics`¶

Logs evaluation metrics at the end of a process step. Each metric calls a sklearn-compatible scoring function:

metrics:
  - name: "accuracy"
    fn: "accuracy_score"
    inputs:
      y_true: "y_train"
      y_pred: "predictions"

Logged metrics are stored in the KV backend and can be read back via probe paths.

Components vs. Custom Code¶

	Component	Custom `@step` code
Setup	`pip install expops-sklearn`	Write Python in `src/`
Config	`component: "sklearn.LogisticRegression.fit_predict"`	`code: "model.train"`
Flexibility	Fixed interface, configurable via `parameters`	Full control
Best for	Standard ML steps, quick prototyping	Custom models, novel preprocessing

Both approaches can coexist in the same pipeline.

Available Libraries¶

sklearn — Data loading, splitting, and sklearn estimators
matplotlib — Bar charts and histograms from probe metrics