langfun 0.1.2.dev202411110804__py3-none-any.whl → 0.1.2.dev202411120804__py3-none-any.whl

langfun/core/eval/v2/evaluation.py

@@ -0,0 +1,615 @@
+# Copyright 2024 The Langfun Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Base class for Langfun evaluation tasks."""
+
+import abc
+import functools
+import time
+
+from typing import Annotated, Any, Callable, Iterable
+import langfun.core as lf
+import langfun.core.coding as lf_coding
+
+from langfun.core.eval.v2 import example as example_lib
+from langfun.core.eval.v2 import experiment as experiment_lib
+from langfun.core.eval.v2 import metric_values as metric_values_lib
+from langfun.core.eval.v2 import metrics as metrics_lib
+
+import pyglove as pg
+
+
+class Evaluation(experiment_lib.Experiment):
+  """Evaluation.
+
+  An evaluation can be a leaf node or a container of other evaluations,
+  depending on whether the current evaluation object is configured with
+  any `pg.oneof`.
+
+  For example, `MyEval(lm=pg.oneof([lf.llms.Gpt4(), lf.llms.Gemini1_5Pro()]))`
+  is a container of two sub-experiments, one for each LLM. In such case, the
+  evaluation object with `pg.oneof` is called a hyper evaluation, which
+  represents a search space of evaluations, and each sub-evaluation is called
+  a leaf evaluation, which will perform the actual evaluation.
+  """
+
+  inputs: Annotated[
+      pg.Functor,
+      'A functor that returns a list of inputs.'
+  ]
+
+  metrics: Annotated[
+      list[metrics_lib.Metric],
+      'The metrics to be evaluated.'
+  ]
+
+  max_workers: Annotated[
+      int,
+      'The maximum number of workers to use for the evaluation.'
+  ] = 32
+
+  def _on_bound(self):
+    # Invalidate cached properties.
+    self.__dict__.pop('is_leaf', None)
+    self.__dict__.pop('children', None)
+    super()._on_bound()
+
+  #
+  # Handling evaluation hierarchy (materialized vs. hyper evaluations).
+  #
+
+  @functools.cached_property
+  def is_leaf(self) -> bool:
+    """Returns whether the task is a leaf."""
+    return self.is_deterministic
+
+  @functools.cached_property
+  def children(self) -> list['Evaluation']:
+    """Returns the children tasks."""
+    if self.is_leaf:
+      return []
+    children = []
+    for i, child in enumerate(pg.iter(self)):
+      child.sym_setparent(self)
+      child.sym_setpath(self.sym_path + 'children' + i)
+      children.append(child)
+    return children
+
+  #
+  # Handling evaluation inputs.
+  #
+
+  @functools.cached_property
+  def example_inputs(self) -> Iterable[Any]:
+    """Returns the examples from the inputs."""
+    return self.inputs()
+
+  def example_input_by_id(self, example_id: int) -> Any:
+    """Returns the example from the inputs by ID."""
+    assert example_id <= len(self.example_inputs), example_id
+    return self._example_input_by_id[example_id]
+
+  @functools.cached_property
+  def _example_input_by_id(self) -> dict[int, Any]:
+    """Returns the examples from the inputs by ID."""
+    return {i + 1: v for i, v in enumerate(self.example_inputs)}
+
+  @property
+  def num_examples(self) -> int:
+    """Returns the number of examples from the inputs."""
+    # NOTE(daiyip): setting `num_examples` of the input functor allows fast
+    # retrieval of number of examples without interating the whole dataset.
+    return getattr(self.inputs, 'num_examples', len(self.example_inputs))
+
+  #
+  # Evaluation logics.
+  #
+
+  @abc.abstractmethod
+  def process(self, example_input: Any) -> Any | tuple[Any, dict[str, Any]]:
+    """Processes a single example from the evaluation set.
+
+    Users should override this method to implement the evaluation logic.
+
+    Args:
+      example_input: An object returned from `Evaluable.inputs`.
+
+    Returns:
+      A processed output. Or a tuple of (output, metadata).
+      The output will be used for computing the metrics, and the metadata will
+      be included in the evaluation HTML view.
+    """
+
+  def evaluate(
+      self,
+      example: example_lib.Example | int,
+      raise_if_has_error: bool = False,
+  ) -> example_lib.Example:
+    """Evaluates a single example input.
+
+    Args:
+      example: An example ID or an example object with ID.
+      raise_if_has_error: Whether to raise an error if the example has error.
+
+    Returns:
+      The evaluated example with the output and metric metadata populated.
+    """
+    if isinstance(example, int):
+      example = example_lib.Example(id=example)
+    assert isinstance(example, example_lib.Example), example
+
+    if pg.MISSING_VALUE == example.input:
+      example.input = self.example_input_by_id(example.id)
+
+    cached = self._state.get(example.id)
+
+    with pg.timeit('evaluate') as timeit, lf.track_usages() as usage_summary:
+      if cached is None or cached.has_error:
+        example.start_time = time.time()
+        self._process(example, raise_if_has_error=raise_if_has_error)
+      else:
+        example.start_time = cached.start_time
+
+        # Use cached output and metadata obtained from the previous processing.
+        example.output = cached.output
+        example.metadata = cached.metadata
+        example.newly_processed = False
+
+        # For previously processed examples, we merge previous usages as
+        # cached, so the usage summary will account previous usages, but as
+        # cached.
+        assert cached.usage_summary is not None
+        usage_summary.merge(cached.usage_summary, as_cached=True)
+
+      # Recompute the metrics and metadata for the example even its processed
+      # output and metadata were from the cache.
+      # NOTE(daiyip): It's possible that metrics could use LLMs, so we need to
+      # track the usage of the metrics separately.
+      with pg.timeit('metric'):
+        metric_metadata = {}
+        for metric in self.metrics:
+          metric_metadata.update(metric.audit(example))
+        example.metric_metadata = metric_metadata
+
+    # For previously processed examples, we keep the execution status for the
+    # processing step.
+    execution_status = dict(example.execution_status or {})
+    execution_status.update(timeit.status())
+
+    example.execution_status = execution_status
+    example.usage_summary = usage_summary
+    if example.newly_processed:
+      example.end_time = time.time()
+
+    self._state.update(example)
+    return example
+
+  def _process(
+      self,
+      example: example_lib.Example,
+      raise_if_has_error: bool = False
+  ) -> None:
+    """Processes a single example."""
+    with (
+        pg.notify_on_change(False),
+        pg.allow_writable_accessors(True),
+        # NOTE(daiyip): set the `input` symbol of the globals to None, so
+        # LLM generated code with calls to `input` will raise an error, thus
+        # not blocking the evaluation.
+        lf_coding.context(input=None),
+    ):
+      try:
+        with pg.timeit('process'):
+          output = self.process(example.input)
+        if (isinstance(output, tuple)
+            and len(output) == 2
+            and isinstance(output[1], dict)):
+          output, metadata = output
+        else:
+          metadata = {}
+        example.output = output
+        example.metadata = metadata
+      except BaseException as e:  # pylint: disable=broad-except
+        if raise_if_has_error:
+          raise
+        example.error = pg.object_utils.ErrorInfo.from_exception(e)
+
+  #
+  # Handling evaluation scheduling.
+  #
+
+  def resource_ids(self) -> set[str]:
+    """Returns a set of resource IDs required by this evaluation.
+
+    Resource IDs are used to by the runner to determine which evaluations can
+    be run in parallel. Evaluations using the same resource key will be run
+    sequentially.
+
+    Returns:
+      A unique string representing the resource required.
+    """
+    return {
+        v.resource_id for _, v in self.sym_init_args.items()
+        if isinstance(v, lf.LanguageModel)
+    }
+
+  #
+  # Handling evaluation state.
+  #
+
+  def load_state(
+      self, state_file: str, raise_if_not_exist: bool = False
+  ) -> None:
+    """Loads saved state from a sequence IO file."""
+    if pg.io.path_exists(state_file):
+      self._state.load(state_file, self.example_input_by_id)
+    elif raise_if_not_exist:
+      raise ValueError(f'State file {state_file} does not exist.')
+
+  def _reset(self) -> None:
+    """Resets the state of the evaluation."""
+    super()._reset()
+    if self.is_leaf:
+      # Create a new state for the leaf evaluation.
+      self._state = EvaluationState()
+      for metric in self.metrics:
+        metric.reset()
+
+  #
+  # HTML views.
+  #
+
+  def _html_tree_view_content(
+      self, *, view, extra_flags: dict[str, Any] | None, **kwargs
+  ):
+    if not self.is_leaf:
+      return super()._html_tree_view_content(
+          view=view, extra_flags=extra_flags, **kwargs
+      )
+
+    extra_flags = extra_flags or {}
+    run = extra_flags.pop('current_run', None)
+    if extra_flags.pop('card_view', True):
+      return self._summary_card_view(
+          extra_flags.get('interactive', True), run
+      )
+    assert run is not None
+    return self._details_view(run)
+
+  def _parameter_badge(self, key, value) -> pg.Html.WritableTypes:
+    """Renders a badge for a parameter."""
+    face_value = pg.format(
+        value,
+        compact=True,
+        python_format=True,
+        hide_default_values=True,
+        use_inferred=True
+    )
+    short_text = face_value
+    if len(face_value) > 40:
+      short_text = f'{type(value).__name__}(...)'
+    label = f'{key.split(".")[-1]}: {short_text}'
+    tooltip = f'{key}: {face_value}'
+    return pg.views.html.controls.Badge(
+        text=label,
+        tooltip=tooltip,
+        css_classes=['parameter'],
+        interactive=False,
+    )
+
+  def _summary_card_view(
+      self,
+      interactive: bool = True,
+      run: experiment_lib.Run | None = None,
+  ) -> pg.Html.WritableTypes:
+    """Renders the summary card view of the evaluation."""
+    del run
+    return pg.Html(
+        pg.Html.element(
+            'div',
+            [
+                pg.views.html.controls.LabelGroup([
+                    self._parameter_badge(k, v)
+                    for k, v in self.non_default_values(
+                        flatten=True
+                    ).items()
+                ], css_classes=['parameter-group']),
+                pg.Html.element(
+                    'div',
+                    [
+                        m.to_html(
+                            extra_flags=dict(
+                                interactive=interactive,
+                            )
+                        )
+                        for m in self.metrics
+                    ],
+                    css_classes=['metric-group'],
+                ),
+            ],
+            css_classes=['badge-groups'],
+        )
+    )
+
+  def _details_view(
+      self, run: experiment_lib.Run
+  ) -> pg.Html:
+    """Renders the details view of the evaluation."""
+
+    def _title():
+      return pg.Html.element(
+          'div',
+          [
+              pg.views.html.controls.LabelGroup(
+                  [
+                      pg.views.html.controls.Label(
+                          'Summary',
+                          link=run.experiment.output_link(run, 'summary.html'),
+                          css_classes=['summary-link'],
+                      ),
+                      '|',
+                      pg.views.html.controls.Label(
+                          'Directory',
+                          link=self.output_link(run, ''),
+                          css_classes=['dir-link'],
+                      ),
+                  ],
+                  css_classes=['experiment-links'],
+              ),
+              pg.views.html.controls.Label(
+                  self.id,
+                  css_classes=['experiment-id'],
+              ),
+              self.progress.to_html(
+                  extra_flags=dict(interactive=False),
+              ),
+              self.usage_summary.to_html(
+                  extra_flags=dict(as_badge=True, interactive=False),
+              ),
+          ]
+      )
+
+    def _parameter_badges():
+      """Renders a tab group for a metric (group)."""
+      return pg.views.html.controls.LabelGroup(
+          [
+              self._parameter_badge(k, v)
+              for k, v in self.non_default_values(flatten=True).items()
+          ],
+          css_classes=['parameter-group'],
+      )
+
+    def _definition_tab() -> pg.views.html.controls.Tab:
+      """Renders a tab for the definition of the evaluation."""
+      return pg.views.html.controls.Tab(
+          label='Definition',
+          content=pg.Html.element(
+              'div',
+              [
+                  pg.views.html.controls.Label(
+                      pg.format(
+                          self,
+                          compact=False,
+                          verbose=False,
+                          use_inferred=True,
+                          hide_frozen=True,
+                          exclude_keys=set(['progress', 'usage_summary'])
+                      ),
+                      css_classes=['eval-definition'],
+                  ),
+              ]
+          )
+      )
+
+    def _metric_tab(metric: metrics_lib.Metric) -> pg.views.html.controls.Tab:
+      """Renders a tab for a metric (group)."""
+      return pg.views.html.controls.Tab(
+          label=f'Metric: {metric.name}',
+          content=pg.Html.element(
+              'div',
+              [
+                  metric.to_html(
+                      extra_flags=dict(
+                          interactive=False,
+                      )
+                  ),
+                  pg.views.html.controls.TabControl(
+                      tabs=[
+                          _metric_value_tab(mv)
+                          for mv in metric.values()
+                      ]
+                  )
+              ]
+          )
+      )
+
+    def _metric_value_tab(
+        metric_value: metric_values_lib.MetricValue
+    ) -> pg.views.html.controls.Tab:
+      """Renders the example links for a metric value."""
+      return pg.views.html.controls.Tab(
+          label=metric_value.sym_path.key,
+          content=pg.Html.element(
+              'div',
+              [
+                  pg.views.html.controls.Label(
+                      str(dp.example_id),
+                      link=self.output_link(run, f'{dp.example_id}.html'),
+                      target='example-view',
+                      css_classes=['example-link'],
+                  )
+                  for dp in metric_value.data_points
+              ]
+          )
+      )
+
+    def _main_tabs() -> pg.Html:
+      return pg.Html.element(
+          'div',
+          [
+              pg.views.html.controls.TabControl(
+                  [
+                      _definition_tab(),
+                  ] + [
+                      _metric_tab(m) for m in self.metrics
+                  ],
+                  selected=1,
+              )
+          ],
+      )
+
+    return pg.Html.element(
+        'div',
+        [
+            _title(),
+            _parameter_badges(),
+            _main_tabs(),
+            pg.Html.element(
+                'iframe', [],
+                name='example-view',
+                src='./1.html',
+                title='Example view.',
+                css_classes=['example-view'],
+            ),
+        ],
+        css_classes=['eval-details'],
+    )
+
+  def _html_tree_view_config(self) -> dict[str, Any]:
+    return dict(
+        css_classes=['eval-card'] if self.is_leaf else None
+    )
+
+  def _html_tree_view_css_styles(self) -> list[str]:
+    return super()._html_tree_view_css_styles() + [
+        """
+        details.eval-card {
+          display: inline-block;
+          border: 0px;
+          box-shadow: rgba(0, 0, 0, 0.16) 0px 1px 4px;
+          margin: 15px;
+        }
+        .eval-card details {
+          border: 0px;
+        }
+        .badge-groups {
+          font-weight: normal;
+          padding: 5px;
+        }
+        .parameter-group {
+          display: inline-grid;
+          grid-template-rows: auto auto;
+          border: 0px;
+          margin-right: 10px;
+        }
+        .parameter.badge {
+          margin: 2px;
+        }
+        .metric-group {
+          display: inline-grid;
+          grid-template-rows: auto auto;
+        }
+        .eval-details .progress-bar > .shade {
+          visibility: hidden;
+          width: 0px;
+          margin: 0px;
+        }
+        .eval-details .progress-label {
+          font-size: 16px;
+          background-color: #eee;
+        }
+        .eval-details .progress-time {
+          font-size: 16px;
+          color: dodgerblue;
+          background-color: #eee;
+          margin-right: 10px;
+        }
+        .eval-details .usage-summary.badge {
+          color: orange;
+          font-size: 16px;
+          background-color: #eee;
+        }
+        .eval-details .experiment-links {
+          display: block;
+          border: 0px;
+          margin: 0px;
+        }
+        .eval-details .tab-button {
+          font-size: large;
+        }
+        .experiment-links .label {
+          color: revert;
+          margin: 0px;
+          padding: 2px;
+        }
+        .eval-details .experiment-id {
+          font-size: 2.0em;
+          font-weight: bold;
+          display: block;
+        }
+        .eval-details .parameter-group {
+          display: inline-block;
+          padding: 5px;
+        }
+        .eval-definition {
+          white-space: pre;
+          background-color: #eee;
+          padding: 15px;
+        }
+        .eval-details .metric-container {
+          display: block;
+          padding: 15px 0px;
+        }
+        .example-link {
+          color: revert;
+        }
+        .example-view {
+          border: 0px;
+          width:100%;
+          height:100%;
+        }
+        """
+    ]
+
+
+class EvaluationState:
+  """Evaluation state."""
+
+  def __init__(self):
+    super().__init__()
+    self._evaluated_examples: dict[int, example_lib.Example] = {}
+
+  def load(
+      self, state_file: str, example_input_by_id: Callable[[int], Any]) -> None:
+    """Loads the state from the example sequence file."""
+    with pg.io.sequence.open_sequence(state_file) as f:
+      for record in f:
+        example = pg.from_json_str(
+            record, example_input_by_id=example_input_by_id
+        )
+        assert isinstance(example, example_lib.Example), example
+        self._evaluated_examples[example.id] = example
+
+  def get(self, example_id: int) -> example_lib.Example | None:
+    """Returns the example with the given ID."""
+    return self._evaluated_examples.get(example_id)
+
+  def update(self, example: example_lib.Example) -> None:
+    """Updates the state with the given example."""
+    self._evaluated_examples[example.id] = example
+
+  @property
+  def evaluated_examples(self) -> dict[int, example_lib.Example]:
+    """Returns the examples in the state."""
+    return self._evaluated_examples
+