langfun 0.1.2.dev202411100803__py3-none-any.whl → 0.1.2.dev202411120804__py3-none-any.whl

langfun/core/eval/v2/evaluation_test.py

@@ -0,0 +1,143 @@
+# 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.
+import os
+import tempfile
+import unittest
+
+from langfun.core.eval.v2 import evaluation as evaluation_lib
+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 test_helper
+
+import pyglove as pg
+
+Example = example_lib.Example
+Evaluation = evaluation_lib.Evaluation
+RunId = experiment_lib.RunId
+Run = experiment_lib.Run
+
+
+class EvaluationTest(unittest.TestCase):
+
+  def test_hyper_evaluation(self):
+    exp = test_helper.TestEvaluation(
+        lm=test_helper.TestLLM(offset=pg.oneof(range(3)))
+    )
+    self.assertFalse(exp.is_leaf)
+    self.assertTrue(
+        pg.eq(
+            exp.children,
+            [
+                test_helper.TestEvaluation(lm=test_helper.TestLLM(offset=0)),
+                test_helper.TestEvaluation(lm=test_helper.TestLLM(offset=1)),
+                test_helper.TestEvaluation(lm=test_helper.TestLLM(offset=2)),
+            ]
+        )
+    )
+    self.assertEqual(exp.children[0].num_examples, 10)
+    self.assertEqual(
+        [c.is_leaf for c in exp.children],
+        [True] * len(exp.children)
+    )
+    self.assertEqual(
+        [r.resource_ids() for r in exp.leaf_nodes],
+        [set(['test_llm:0']), set(['test_llm:1']), set(['test_llm:2'])]
+    )
+
+  def test_evaluate(self):
+    exp = test_helper.TestEvaluation()
+    example = exp.evaluate(Example(id=3))
+    self.assertTrue(example.newly_processed)
+    self.assertEqual(example.input, pg.Dict(x=2, y=4, groundtruth=6))
+    self.assertEqual(example.output, 6)
+    self.assertIsNone(example.error)
+    self.assertEqual(example.metadata, {})
+    self.assertEqual(example.metric_metadata, dict(match=True))
+    self.assertIsNotNone(example.usage_summary)
+    self.assertGreater(example.usage_summary.total.total_tokens, 0)
+    self.assertEqual(example.usage_summary.total.num_requests, 1)
+    self.assertIsNotNone(example.execution_status)
+    self.assertIsNotNone(example.start_time)
+    self.assertIsNotNone(example.end_time)
+
+    exp = test_helper.TestEvaluation(lm=test_helper.TestLLM(offset=1))
+    example = exp.evaluate(3)
+    self.assertTrue(example.newly_processed)
+    self.assertEqual(example.input, pg.Dict(x=2, y=4, groundtruth=6))
+    self.assertEqual(example.output, 7)
+    self.assertIsNone(example.error)
+    self.assertEqual(example.metadata, {})
+    self.assertEqual(example.metric_metadata, dict(mismatch=True))
+
+    with self.assertRaisesRegex(ValueError, 'x should not be 5'):
+      _ = exp.evaluate(6, raise_if_has_error=True)
+    example = exp.evaluate(6)
+    self.assertTrue(example.newly_processed)
+    self.assertEqual(example.input, pg.Dict(x=5, y=25, groundtruth=30))
+    self.assertEqual(pg.MISSING_VALUE, example.output)
+    self.assertEqual(example.error.tag, 'ValueError')
+    self.assertEqual(example.metadata, {})
+    self.assertEqual(example.metric_metadata, dict(error='ValueError'))
+
+  def test_evaluate_with_state(self):
+    eval_dir = os.path.join(tempfile.gettempdir(), 'test_eval')
+    pg.io.mkdirs(eval_dir, exist_ok=True)
+    state_file = os.path.join(eval_dir, 'state.jsonl')
+    with pg.io.open_sequence(state_file, 'w') as f:
+      exp = test_helper.TestEvaluation()
+      example = exp.evaluate(3)
+      self.assertTrue(example.newly_processed)
+      self.assertEqual(example.input, pg.Dict(x=2, y=4, groundtruth=6))
+      self.assertEqual(example.output, 6)
+      self.assertEqual(len(exp._state.evaluated_examples), 1)
+      f.add(pg.to_json_str(example))
+
+    exp.reset()
+    self.assertEqual(len(exp._state.evaluated_examples), 0)
+    exp.load_state(state_file)
+    self.assertEqual(len(exp._state.evaluated_examples), 1)
+    example = exp.evaluate(3)
+    self.assertFalse(example.newly_processed)
+    self.assertEqual(example.input, pg.Dict(x=2, y=4, groundtruth=6))
+    self.assertEqual(example.output, 6)
+    self.assertGreater(example.usage_summary.total.total_tokens, 0)
+    self.assertGreater(example.usage_summary.cached.total.total_tokens, 0)
+    self.assertEqual(example.usage_summary.cached.total.num_requests, 1)
+    self.assertEqual(example.usage_summary.uncached.total.total_tokens, 0)
+    self.assertEqual(example.usage_summary.uncached.total.num_requests, 0)
+
+  def test_html_view(self):
+    exp = test_helper.TestEvaluation()
+    self.assertIn(
+        exp.id,
+        exp.to_html(extra_flags=dict(card_view=True, current_run=None)).content
+    )
+    self.assertIn(
+        exp.id,
+        exp.to_html(
+            extra_flags=dict(
+                card_view=False,
+                current_run=Run(
+                    root_dir='/tmp/test_run',
+                    id=RunId.from_id('20241031_1'),
+                    experiment=pg.Ref(exp),
+                )
+            )
+        ).content
+    )
+
+
+if __name__ == '__main__':
+  unittest.main()

langfun/core/eval/v2/example.py

@@ -0,0 +1,286 @@
+# 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 classes for Langfun evaluation."""
+
+import dataclasses
+import inspect
+from typing import Any, Callable
+import langfun.core as lf
+import pyglove as pg
+
+
+@dataclasses.dataclass
+class Example(pg.JSONConvertible, pg.views.HtmlTreeView.Extension):
+  """An item for the evaluation.
+
+  Attributes:
+    id: The 1-based ID of the item in the evaluation set.
+    input: An element returned from the `Evaluable.inputs` functor.
+    output: The output of the `process` method. If `pg.MISSING_VALUE`, it has
+      not been processed yet.
+    metadata: The metadata of the item produced by the `process` method.
+    metric_metadata: The dictionary returned from `Metric.audit`.
+    start_time: The start time of the evaluation item.
+    end_time: The end time of the evaluation item.
+    usage_summary: The summary of LLM usages of the evaluation item.
+    execution_status: The timeit status of the evaluation item.
+  """
+  id: int
+  input: Any = pg.MISSING_VALUE
+  output: Any = pg.MISSING_VALUE
+  error: pg.object_utils.ErrorInfo | None = None
+  metadata: dict[str, Any] = dataclasses.field(default_factory=dict)
+  metric_metadata: dict[str, Any] | None = None
+  # Execution information.
+  newly_processed: bool = True
+  start_time: float | None = None
+  end_time: float | None = None
+  usage_summary: lf.UsageSummary | None = None
+  execution_status: dict[str, pg.object_utils.TimeIt.Status] | None = None
+
+  def __post_init__(self):
+    if self.execution_status is not None:
+      for status in self.execution_status.values():
+        if status.has_error:
+          self.error = status.error
+          break
+
+  @property
+  def is_processed(self) -> bool:
+    """Returns whether the item has been processed."""
+    return pg.MISSING_VALUE != self.output
+
+  @property
+  def has_error(self) -> bool:
+    """Returns whether the item has an error."""
+    return self.error is not None
+
+  @property
+  def elapse(self) -> float | None:
+    """Returns the elapse time of the item."""
+    if self.execution_status is not None:
+      return self.execution_status['evaluate'].elapse
+    return None
+
+  def to_json(self, **kwargs) -> dict[str, Any]:
+    """Returns the JSON representation of the item."""
+    return self.to_json_dict(
+        fields=dict(
+            id=(self.id, None),
+            # NOTE(daiyip): We do not write `input` to JSON as it will be
+            # loaded from the input functor. This allows us to support
+            # non-serializable examples.
+            output=(self.output, pg.MISSING_VALUE),
+            error=(self.error, None),
+            metadata=(self.metadata, {}),
+            metric_metadata=(self.metric_metadata, None),
+            start_time=(self.start_time, None),
+            end_time=(self.end_time, None),
+            usage_summary=(self.usage_summary, None),
+            execution_status=(self.execution_status, None),
+        ),
+        exclude_default=True,
+        **kwargs,
+    )
+
+  @classmethod
+  def from_json(
+      cls,
+      json_value: dict[str, Any],
+      *,
+      example_input_by_id: Callable[[int], Any],
+      **kwargs
+  ) -> 'Example':
+    """Creates an example from the JSON representation."""
+    example_id = json_value.get('id')
+    example_input = example_input_by_id(example_id)
+    json_value['input'] = example_input
+
+    # NOTE(daiyip): We need to load the types of the examples into the
+    # deserialization context, otherwise the deserialization will fail if the
+    # types are not registered.
+    def example_class_defs(example) -> list[type[Any]]:
+      referred_types = set()
+      def _visit(k, v, p):
+        del k, p
+        if inspect.isclass(v):
+          referred_types.add(v)
+        elif isinstance(v, pg.Object):
+          referred_types.add(v.__class__)
+        return pg.TraverseAction.ENTER
+      pg.traverse(example, _visit)
+      return list(referred_types)
+
+    with pg.JSONConvertible.load_types_for_deserialization(
+        *example_class_defs(example_input)
+    ):
+      return cls(
+          **{k: pg.from_json(v, **kwargs) for k, v in json_value.items()}
+      )
+
+  #
+  # HTML rendering.
+  #
+
+  def _html_tree_view_content(
+      self,
+      *,
+      view: pg.views.HtmlTreeView,
+      root_path: pg.KeyPath | None = None,
+      extra_flags: dict[str, Any] | None = None,
+      **kwargs
+  ):
+    root_path = root_path or pg.KeyPath()
+    extra_flags = extra_flags or {}
+    num_examples = extra_flags.get('num_examples', None)
+
+    def _metric_metadata_badge(key, value):
+      if isinstance(value, bool) and bool:
+        text = key
+      else:
+        text = f'{key}:{value}'
+      return pg.views.html.controls.Badge(
+          text,
+          css_classes=[pg.object_utils.camel_to_snake(key, '-')],
+      )
+
+    def _render_header():
+      return pg.Html.element(
+          'div',
+          [
+              pg.Html.element(
+                  'div',
+                  [
+                      # Previous button.
+                      pg.views.html.controls.Label(   # pylint: disable=g-long-ternary
+                          '◀',
+                          link=f'{self.id - 1}.html',
+                          css_classes=['previous'],
+                      ) if self.id > 1 else None,
+                      # Current example ID.
+                      pg.views.html.controls.Label(
+                          f'#{self.id}',
+                          css_classes=['example-id'],
+                      ),
+                      # Next button.
+                      pg.views.html.controls.Label(   # pylint: disable=g-long-ternary
+                          '▶',
+                          link=f'{self.id + 1}.html',
+                          css_classes=['next'],
+                      ) if (num_examples is None
+                            or self.id < num_examples) else None,
+
+                  ]
+              ),
+              pg.Html.element(
+                  'div',
+                  [
+                      # Usage summary.
+                      pg.view(  # pylint: disable=g-long-ternary
+                          self.usage_summary,
+                          extra_flags=dict(as_badge=True)
+                      ) if self.usage_summary is not None else None,
+                      # Metric metadata.
+                      pg.views.html.controls.LabelGroup(
+                          [   # pylint: disable=g-long-ternary
+                              _metric_metadata_badge(k, v)
+                              for k, v in self.metric_metadata.items()
+                          ] if self.metric_metadata else []
+                      ),
+                  ],
+                  css_classes=['example-container'],
+              )
+          ]
+      )
+
+    def _render_content():
+      def _tab(label, key):
+        field = getattr(self, key)
+        if pg.MISSING_VALUE == field or not field:
+          return None
+        return pg.views.html.controls.Tab(
+            label=label,
+            content=view.render(
+                field,
+                root_path=root_path + key,
+                **view.get_passthrough_kwargs(**kwargs),
+            ),
+        )
+      tabs = [
+          _tab('Input', 'input'),
+          _tab('Output', 'output'),
+          _tab('Output Metadata', 'metadata'),
+          _tab('Error', 'error'),
+      ]
+      return pg.views.html.controls.TabControl(
+          [tab for tab in tabs if tab is not None]
+      )
+
+    return pg.Html.element(
+        'div',
+        [
+            _render_header(),
+            _render_content(),
+        ],
+        css_classes=['eval-example']
+    )
+
+  def _html_tree_view_summary(self, *, view, **kwargs):
+    return None
+
+  @classmethod
+  def _html_tree_view_css_styles(cls) -> list[str]:
+    return super()._html_tree_view_css_styles() + [
+        """
+        .example-container {
+          display: block;
+          padding: 10px;
+        }
+        .example-id {
+          font-weight: bold;
+          font-size: 40px;
+          margin: 0 10px;
+          vertical-align: middle;
+        }
+        a.previous, a.next {
+          text-decoration: none;
+          vertical-align: middle;
+          display: inline-block;
+          padding: 8px 8px;
+          color: #DDD;
+        }
+        a.previous:hover, a.next:hover {
+          background-color: #ddd;
+          color: black;
+        }
+        /* Badge styles. */
+        .eval-example .badge.match {
+          color: green;
+          background-color: #dcefbe;
+        }
+        .eval-example .badge.error {
+          color: red;
+          background-color: #fdcccc;
+        }
+        .eval-example .badge.mismatch {
+          color: orange;
+          background-color: #ffefc4;
+        }
+        .eval-example .badge.score {
+          color: blue;
+          background-color: #c4dced;
+        }
+        """
+    ]
+

langfun/core/eval/v2/example_test.py

@@ -0,0 +1,92 @@
+# 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.
+import unittest
+
+from langfun.core.eval.v2 import example as example_lib
+import pyglove as pg
+
+Example = example_lib.Example
+
+
+class ExampleTest(unittest.TestCase):
+
+  def test_basic(self):
+    error = pg.object_utils.ErrorInfo(
+        tag='ValueError',
+        description='Bad input',
+        stacktrace='...',
+    )
+    ex = Example(id=1, execution_status={
+        'evaluate': pg.object_utils.TimeIt.Status(
+            name='evaluation', elapse=1.0, error=error
+        )
+    })
+    self.assertEqual(ex.error, error)
+    self.assertFalse(ex.is_processed)
+    self.assertTrue(ex.has_error)
+    self.assertEqual(ex.elapse, 1.0)
+
+    ex = Example(id=2, output=1)
+    self.assertTrue(ex.is_processed)
+    self.assertFalse(ex.has_error)
+    self.assertIsNone(ex.elapse)
+
+  def test_json_conversion(self):
+    def input_func():
+      class A(pg.Object):
+        x: int
+
+      class B(pg.Object):
+        x: int = 1
+        y: int = 2
+
+      return [
+          pg.Dict(
+              a=A,
+              b=B
+          )
+      ]
+
+    inputs = input_func()
+    ex = Example(
+        id=1,
+        input=inputs[0],
+        output=inputs[0].a(1),
+        metadata=dict(b=inputs[0].b())
+    )
+    json_str = pg.to_json_str(ex)
+    self.assertEqual(
+        pg.from_json_str(
+            json_str,
+            example_input_by_id=lambda i: inputs[i - 1]
+        ),
+        ex
+    )
+
+  def test_html_view(self):
+    ex = Example(
+        id=1,
+        input=pg.Dict(a=1, b=2),
+        output=3,
+        metadata=dict(sum=3),
+        metric_metadata=dict(match=True),
+    )
+    self.assertNotIn(
+        'next',
+        ex.to_html(extra_flags=dict(num_examples=1)).content,
+    )
+
+
+if __name__ == '__main__':
+  unittest.main()