scorebook 0.0.9__py3-none-any.whl → 0.0.11__py3-none-any.whl

scorebook/evaluate/_sync/evaluate.py

@@ -0,0 +1,443 @@
+import logging
+from typing import Any, Callable, Dict, List, Literal, Optional, Union, cast
+
+from trismik import TrismikAsyncClient, TrismikClient
+from trismik.types import (
+    TrismikClassicEvalItem,
+    TrismikClassicEvalMetric,
+    TrismikClassicEvalRequest,
+    TrismikClassicEvalResponse,
+    TrismikRunMetadata,
+)
+
+from scorebook.eval_datasets import EvalDataset
+from scorebook.evaluate.evaluate_helpers import (
+    build_eval_run_specs,
+    create_trismik_sync_client,
+    format_results,
+    get_model_name,
+    make_trismik_inference,
+    prepare_datasets,
+    prepare_hyperparameter_configs,
+    resolve_show_progress,
+    resolve_upload_results,
+    score_metrics,
+    validate_parameters,
+)
+from scorebook.exceptions import InferenceError, ScoreBookError
+from scorebook.types import (
+    AdaptiveEvalRunResult,
+    AdaptiveEvalRunSpec,
+    ClassicEvalRunResult,
+    EvalResult,
+    EvalRunSpec,
+)
+from contextlib import nullcontext
+from scorebook.utils import evaluation_progress_context
+
+logger = logging.getLogger(__name__)
+
+
+def evaluate(
+    inference: Callable,
+    datasets: Union[str, EvalDataset, List[Union[str, EvalDataset]]],
+    hyperparameters: Optional[Union[Dict[str, Any], List[Dict[str, Any]]]] = None,
+    metadata: Optional[Dict[str, Any]] = None,
+    experiment_id: Optional[str] = None,
+    project_id: Optional[str] = None,
+    return_dict: bool = True,
+    return_aggregates: bool = True,
+    return_items: bool = False,
+    return_output: bool = False,
+    upload_results: Union[Literal["auto"], bool] = "auto",
+    sample_size: Optional[int] = None,
+    show_progress: Optional[bool] = None,
+) -> Union[Dict, List, EvalResult]:
+    """
+    Evaluate a model across a collection of hyperparameters and datasets.
+
+    Args:
+        inference: The inference callable to evaluate
+        datasets: Dataset(s) to evaluate on
+        hyperparameters: Hyperparameter configuration(s) to evaluate with
+        metadata: Optional metadata to attach to the evaluation
+        experiment_id: Optional experiment identifier
+        project_id: Optional project identifier
+        return_dict: If True, returns eval results as a dict
+        return_aggregates: If True, returns aggregate scores for each dataset
+        return_items: If True, returns individual items for each dataset
+        return_output: If True, returns model outputs for each dataset item
+        upload_results: If True, uploads results to Trismik's dashboard
+        sample_size: Optional number of items to sample from each dataset
+        show_progress: If None, uses SHOW_PROGRESS_BARS from settings.
+            If True/False, explicitly enables/disables progress bars for this evaluation.
+
+    Returns:
+        The evaluation results in the format specified by return parameters:
+            - return_dict=True: Returns the evaluation results as a dict
+            - return_dict=False: Returns an EvalResult object containing all run results
+    """
+    # Resolve and validate parameters
+    upload_results = cast(bool, resolve_upload_results(upload_results))
+    show_progress_bars = resolve_show_progress(show_progress)
+    validate_parameters(locals(), evaluate)
+
+    # Prepare datasets, hyperparameters, and eval run specs
+    datasets = prepare_datasets(datasets, sample_size)
+    hyperparameter_configs = prepare_hyperparameter_configs(hyperparameters)
+    eval_run_specs = sorted(
+        build_eval_run_specs(datasets, hyperparameter_configs, experiment_id, project_id, metadata),
+        key=lambda run: (run.dataset_index, run.hyperparameters_index),
+    )
+
+    # Create Trismik client if needed (for adaptive evals or uploads)
+    needs_client = upload_results or any(
+        isinstance(run, AdaptiveEvalRunSpec) for run in eval_run_specs
+    )
+
+    # Use context manager for automatic cleanup, or None if not needed
+    trismik_client = create_trismik_sync_client() if needs_client else None
+
+    with trismik_client or nullcontext():
+        # Execute evaluation runs
+        # Calculate total items across all runs
+        total_items = sum(len(run.dataset.items) for run in eval_run_specs)
+        model_display = get_model_name(inference)
+
+        with evaluation_progress_context(
+            total_eval_runs=len(eval_run_specs),
+            total_items=total_items,
+            dataset_count=len(datasets),
+            hyperparam_count=len(hyperparameter_configs),
+            model_display=model_display,
+            enabled=show_progress_bars,
+        ) as progress_bars:
+            eval_result = execute_runs(
+                inference,
+                eval_run_specs,
+                progress_bars,
+                experiment_id,
+                project_id,
+                metadata,
+                upload_results,
+                trismik_client,
+            )
+            logger.info("Synchronous evaluation complete")
+
+        return format_results(
+            eval_result, return_dict, return_aggregates, return_items, return_output
+        )
+
+
+def execute_runs(
+    inference: Callable,
+    runs: List[Union[EvalRunSpec, AdaptiveEvalRunSpec]],
+    progress_bars: Any,
+    experiment_id: Optional[str] = None,
+    project_id: Optional[str] = None,
+    metadata: Optional[Dict[str, Any]] = None,
+    upload_results: bool = False,
+    trismik_client: Optional[Union[TrismikClient, TrismikAsyncClient]] = None,
+) -> EvalResult:
+    """Run evaluation sequentially."""
+
+    # Worker function to execute individual runs and handle uploads
+    def worker(
+        run: Union[EvalRunSpec, AdaptiveEvalRunSpec]
+    ) -> Union[ClassicEvalRunResult, AdaptiveEvalRunResult]:
+        run_result = execute_run(
+            inference, run, experiment_id, project_id, metadata, trismik_client
+        )
+        # Update progress bars with items processed and success status
+        if progress_bars is not None:
+            items_processed = len(run.dataset.items)
+            progress_bars.on_run_completed(items_processed, run_result.run_completed)
+
+        if (
+            upload_results
+            and isinstance(run_result, ClassicEvalRunResult)
+            and experiment_id
+            and project_id
+            and run_result.run_completed
+            and trismik_client is not None
+        ):
+            try:
+                run_id = upload_classic_run_results(
+                    run_result, experiment_id, project_id, inference, metadata, trismik_client
+                )
+                run_result.run_id = run_id
+                if progress_bars is not None:
+                    progress_bars.on_upload_completed(succeeded=True)
+            except Exception as e:
+                logger.warning(f"Failed to upload run results: {e}")
+                if progress_bars is not None:
+                    progress_bars.on_upload_completed(succeeded=False)
+                # Continue evaluation even if upload fails
+
+        return run_result
+
+    # Execute all runs sequentially
+    run_results = [worker(run) for run in runs]
+
+    # Return in canonical (dataset_idx, hp_idx) order for stability
+    run_results.sort(
+        key=lambda result: (result.run_spec.dataset_index, result.run_spec.hyperparameters_index)
+    )
+
+    # Return EvalResult
+    return EvalResult(run_results)
+
+
+def execute_run(
+    inference: Callable,
+    run: Union[EvalRunSpec, AdaptiveEvalRunSpec],
+    experiment_id: Optional[str] = None,
+    project_id: Optional[str] = None,
+    metadata: Optional[Dict[str, Any]] = None,
+    trismik_client: Optional[Union[TrismikClient, TrismikAsyncClient]] = None,
+) -> Union[ClassicEvalRunResult, AdaptiveEvalRunResult]:
+    """Execute a single evaluation run."""
+
+    if isinstance(run, EvalRunSpec):
+        return execute_classic_eval_run(inference, run)
+
+    elif isinstance(run, AdaptiveEvalRunSpec):
+        resolved_experiment_id = experiment_id if experiment_id is not None else run.experiment_id
+        resolved_project_id = project_id if project_id is not None else run.project_id
+        return execute_adaptive_eval_run(
+            inference,
+            run,
+            resolved_experiment_id,
+            resolved_project_id,
+            metadata,
+            trismik_client,
+        )
+
+    else:
+        raise ScoreBookError(f"An internal error occurred: {type(run)} is not a valid run type")
+
+
+def execute_classic_eval_run(inference: Callable, run: EvalRunSpec) -> ClassicEvalRunResult:
+    """Execute a classic evaluation run."""
+    logger.debug("Executing classic eval run for %s", run)
+
+    inference_outputs = None
+    metric_scores = None
+
+    try:
+        inference_outputs = run_inference_callable(
+            inference, run.inputs, run.hyperparameter_config
+        )
+        metric_scores = score_metrics(run.dataset, inference_outputs, run.labels)
+        logger.debug("Classic evaluation completed for run %s", run)
+        return ClassicEvalRunResult(run, True, inference_outputs, metric_scores)
+
+    except Exception as e:
+        logger.warning("Failed to complete classic eval run for %s: %s", run, str(e))
+        return ClassicEvalRunResult(run, False, inference_outputs, metric_scores)
+
+
+def run_inference_callable(
+    inference: Callable,
+    inputs: List[Any],
+    hyperparameter_config: Dict[str, Any],
+) -> Any:
+    """Run inference on a given dataset and hyperparameter configuration."""
+
+    try:
+        predictions = inference(inputs, **hyperparameter_config)
+    except Exception as e:
+        logger.error(
+            "Inference callable raised an exception: %s",
+            str(e),
+        )
+        raise InferenceError(f"Inference failed: {str(e)}") from e
+
+    if not isinstance(predictions, list) or len(predictions) != len(inputs):
+        raise InferenceError(
+            "Inference callable must return a list of predictions "
+            "of shared length as the inputs. "
+            f"Inputs length: {len(inputs)}, predictions length: {len(predictions)}"
+        )
+
+    if all(prediction == "" for prediction in predictions):
+        logger.warning("Inference callable returned all empty strings for all items")
+
+    if all(prediction is None for prediction in predictions):
+        raise InferenceError("Inference callable returned all None for all items")
+
+    return predictions
+
+
+def execute_adaptive_eval_run(
+    inference: Callable,
+    run: AdaptiveEvalRunSpec,
+    experiment_id: str,
+    project_id: str,
+    metadata: Optional[Dict[str, Any]] = None,
+    trismik_client: Optional[Union[TrismikClient, TrismikAsyncClient]] = None,
+) -> AdaptiveEvalRunResult:
+    """Execute an adaptive evaluation run."""
+    logger.debug("Executing adaptive run for %s", run)
+
+    try:
+        if trismik_client is None:
+            raise ScoreBookError("Trismik client is required for adaptive evaluation")
+
+        adaptive_eval_run_result = run_adaptive_evaluation(
+            inference, run, experiment_id, project_id, metadata, trismik_client
+        )
+        logger.debug("Adaptive evaluation completed for run %s", adaptive_eval_run_result)
+
+        return adaptive_eval_run_result
+
+    except Exception as e:
+        logger.warning("Failed to complete adaptive eval run for %s: %s", run, str(e))
+        return AdaptiveEvalRunResult(run, False, {})
+
+
+def upload_classic_run_results(
+    run_result: ClassicEvalRunResult,
+    experiment_id: str,
+    project_id: str,
+    inference_callable: Optional[Callable],
+    metadata: Optional[Dict[str, Any]],
+    trismik_client: Union[TrismikClient, TrismikAsyncClient],
+) -> str:
+    """Upload a classic evaluation run result to Trismik platform.
+
+    Args:
+        run: The evaluation run result to upload
+        experiment_id: Trismik experiment identifier
+        project_id: Trismik project identifier
+        model: Model name used for evaluation
+        metadata: Optional metadata dictionary
+        trismik_client: Trismik client instance
+
+    Returns:
+        Run id
+    """
+    model = get_model_name(inference_callable)
+
+    # Create eval items from run_spec inputs, outputs, and labels
+    items: List[TrismikClassicEvalItem] = []
+    inputs_outputs = zip(run_result.run_spec.inputs, run_result.outputs)
+    for idx, (input_value, output) in enumerate(inputs_outputs):
+        labels = run_result.run_spec.labels
+        label = labels[idx] if idx < len(labels) else ""
+
+        # Calculate item-level metrics for this item
+        item_metrics: Dict[str, Any] = {}
+        if run_result.scores:
+            for metric_name, metric_data in run_result.scores.items():
+                if isinstance(metric_data, dict) and "item_scores" in metric_data:
+                    if idx < len(metric_data["item_scores"]):
+                        item_metrics[metric_name] = metric_data["item_scores"][idx]
+                else:
+                    # If scores is just a single value, use it for all items
+                    item_metrics[metric_name] = metric_data
+
+        eval_item = TrismikClassicEvalItem(
+            datasetItemId=str(idx),
+            modelInput=str(input_value),
+            modelOutput=str(output),
+            goldOutput=str(label),
+            metrics=item_metrics,
+        )
+        items.append(eval_item)
+
+    # Create eval metrics from run aggregate scores
+    metrics: List[TrismikClassicEvalMetric] = []
+    if run_result.scores:
+        for metric_name, metric_data in run_result.scores.items():
+            if isinstance(metric_data, dict) and "aggregate_scores" in metric_data:
+                # Handle structured metric data with aggregate scores
+                for agg_name, agg_value in metric_data["aggregate_scores"].items():
+                    metric_id = (
+                        f"{metric_name}_{agg_name}" if agg_name != metric_name else metric_name
+                    )
+                    metric = TrismikClassicEvalMetric(metricId=metric_id, value=agg_value)
+                    metrics.append(metric)
+            else:
+                # Handle simple metric data (single value)
+                metric = TrismikClassicEvalMetric(metricId=metric_name, value=metric_data)
+                metrics.append(metric)
+
+    classic_eval_request = TrismikClassicEvalRequest(
+        project_id,
+        experiment_id,
+        run_result.run_spec.dataset.name,
+        model,
+        run_result.run_spec.hyperparameter_config,
+        items,
+        metrics,
+    )
+
+    response: TrismikClassicEvalResponse = trismik_client.submit_classic_eval(
+        classic_eval_request
+    )
+
+    run_id: str = response.id
+    logger.info(f"Classic eval run uploaded successfully with run_id: {run_id}")
+
+    return run_id
+
+
+def run_adaptive_evaluation(
+    inference: Callable,
+    adaptive_run_spec: AdaptiveEvalRunSpec,
+    experiment_id: str,
+    project_id: str,
+    metadata: Any,
+    trismik_client: Union[TrismikClient, TrismikAsyncClient],
+) -> AdaptiveEvalRunResult:
+    """Run an adaptive evaluation using the Trismik API.
+
+    Args:
+        inference: Function to run inference
+        adaptive_run_spec: Specification for the adaptive evaluation
+        experiment_id: Experiment identifier
+        project_id: Trismik project ID
+        metadata: Additional metadata
+        trismik_client: Trismik client instance
+    Returns:
+        Results from the adaptive evaluation
+    """
+    trismik_results = trismik_client.run(
+        test_id=adaptive_run_spec.dataset,
+        project_id=project_id,
+        experiment=experiment_id,
+        run_metadata=TrismikRunMetadata(
+            model_metadata=TrismikRunMetadata.ModelMetadata(name="unknown"),
+            test_configuration={},
+            inference_setup={},
+        ),
+        item_processor=make_trismik_inference(inference),
+        return_dict=False,
+    )
+
+    # Convert TrismikRunResults to AdaptiveEvalRunResult
+    # Extract scores from the Trismik results
+    scores = {}
+    if hasattr(trismik_results, "scores") and trismik_results.scores:
+        scores = trismik_results.scores
+    elif hasattr(trismik_results, "__dict__"):
+        # If scores aren't directly available, include all attributes as scores
+        scores = {k: v for k, v in trismik_results.__dict__.items() if not k.startswith("_")}
+
+    # Convert AdaptiveTestScore objects to JSON-serializable dictionaries
+    def make_json_serializable(obj: Any) -> Any:
+        if hasattr(obj, "theta") and hasattr(obj, "std_error"):
+            # This is likely an AdaptiveTestScore object
+            return {"theta": obj.theta, "std_error": obj.std_error}
+        elif isinstance(obj, dict):
+            return {k: make_json_serializable(v) for k, v in obj.items()}
+        elif isinstance(obj, (list, tuple)):
+            return [make_json_serializable(item) for item in obj]
+        else:
+            return obj
+
+    # Make scores JSON serializable
+    scores = make_json_serializable(scores)
+
+    return AdaptiveEvalRunResult(run_spec=adaptive_run_spec, run_completed=True, scores=scores)