Add test for tags (#67)

Author: Samoed

autointent/modules/prediction/_adaptive.py

@@ -146,7 +146,11 @@ def dump(self, path: str) -> None:
         """
         dump_dir = Path(path)
 
-        metadata = AdaptivePredictorDumpMetadata(r=self._r, tags=self.tags, n_classes=self.n_classes)
+        metadata = AdaptivePredictorDumpMetadata(
+            r=self._r,
+            tags=[t.model_dump() for t in self.tags] if self.tags else None,  # type: ignore[misc]
+            n_classes=self.n_classes,
+        )
 
         with (dump_dir / self.metadata_dict_name).open("w") as file:
             json.dump(metadata, file, indent=4)

autointent/modules/prediction/_threshold.py

@@ -164,9 +164,11 @@ def dump(self, path: str) -> None:
         )
 
         dump_dir = Path(path)
+        metadata_json = self.metadata
+        metadata_json["tags"] = [tag.model_dump() for tag in metadata_json["tags"]] if metadata_json["tags"] else None  # type: ignore[misc]
 
         with (dump_dir / self.metadata_dict_name).open("w") as file:
-            json.dump(self.metadata, file, indent=4)
+            json.dump(metadata_json, file, indent=4)
 
     def load(self, path: str) -> None:
         """

autointent/modules/prediction/_tunable.py

@@ -170,9 +170,11 @@ def dump(self, path: str) -> None:
         )
 
         dump_dir = Path(path)
+        metadata_json = self.metadata
+        metadata_json["tags"] = [tag.model_dump() for tag in metadata_json["tags"]] if metadata_json["tags"] else None  # type: ignore[misc]
 
         with (dump_dir / self.metadata_dict_name).open("w") as file:
-            json.dump(self.metadata, file, indent=4)
+            json.dump(metadata_json, file, indent=4)
 
     def load(self, path: str) -> None:
         """
@@ -183,9 +185,11 @@ def load(self, path: str) -> None:
         dump_dir = Path(path)
 
         with (dump_dir / self.metadata_dict_name).open() as file:
-            metadata: TunablePredictorDumpMetadata = json.load(file)
+            metadata = json.load(file)
 
-        self.metadata = metadata
+        metadata["tags"] = [Tag(**tag) for tag in metadata["tags"]] if metadata["tags"] else None
+
+        self.metadata: TunablePredictorDumpMetadata = metadata
         self.thresh = np.array(metadata["thresh"])
         self.multilabel = metadata["multilabel"]
         self.tags = metadata["tags"]

autointent/modules/prediction/_utils.py

@@ -21,23 +21,31 @@ def apply_tags(labels: npt.NDArray[Any], scores: npt.NDArray[Any], tags: list[Ta
     :param tags: List of `Tag` objects, where each tag specifies mutually exclusive intent IDs.
     :return: Adjusted array of shape (n_samples, n_classes) with binary labels.
     """
-    n_samples, _ = labels.shape
-    res = np.copy(labels)
-
-    for i in range(n_samples):
-        sample_labels = labels[i].astype(bool)
-        sample_scores = scores[i]
-
-        for tag in tags:
-            if any(sample_labels[idx] for idx in tag.intent_ids):
-                # Find the index of the class with the highest score among the tagged indices
-                max_score_index = max(tag.intent_ids, key=lambda idx: sample_scores[idx])
-                # Set all other tagged indices to 0 in the result
-                for idx in tag.intent_ids:
-                    if idx != max_score_index:
-                        res[i, idx] = 0
-
-    return res
+    labels = labels.copy()
+
+    for tag in tags:
+        intent_ids = tag.intent_ids
+
+        labels_sub = labels[:, intent_ids]
+        scores_sub = scores[:, intent_ids]
+
+        assigned = labels_sub == 1
+        num_assigned = assigned.sum(axis=1)
+
+        assigned_scores = np.where(assigned, scores_sub, -np.inf)
+
+        samples_to_adjust = np.where(num_assigned > 1)[0]
+
+        if samples_to_adjust.size > 0:
+            assigned_scores_adjust = assigned_scores[samples_to_adjust, :]
+            idx_max_adjust = assigned_scores_adjust.argmax(axis=1)
+
+            labels_sub[samples_to_adjust, :] = 0
+            labels_sub[samples_to_adjust, idx_max_adjust] = 1
+
+        labels[:, intent_ids] = labels_sub
+
+    return labels
 
 
 class WrongClassificationError(Exception):

tests/assets/configs/multiclass.yaml

@@ -29,7 +29,7 @@
   metric: prediction_accuracy
   search_space:
     - module_type: threshold
-      thresh: [0.5, [0.5, 0.5, 0.5]]
+      thresh: [0.5, [0.5, 0.5, 0.5, 0.5]]
     - module_type: tunable
     - module_type: argmax
     - module_type: jinoos

tests/assets/configs/multilabel.yaml

@@ -25,6 +25,6 @@
   metric: prediction_accuracy
   search_space:
     - module_type: threshold
-      thresh: [0.5, [0.5, 0.5, 0.5]]
+      thresh: [0.5, [0.5, 0.5, 0.5, 0.5]]
     - module_type: tunable
     - module_type: adaptive

tests/assets/data/clinc_subset.json

@@ -14,6 +14,14 @@
             "id": 2,
             "name": "alarm",
             "description": "User wants to set or manage an alarm."
+        },
+        {
+            "id": 3,
+            "name": "alarm reservation",
+            "tags": ["alarm", "reservation"],
+            "regexp_full_match": [],
+            "regexp_partial_match": [],
+            "description": "User wants to set or manage an alarm second time."
         }
     ],
     "train": [
@@ -138,20 +146,125 @@
             "label": 2
         },
         {
-            "utterance": "how much is an overdraft fee for bank"
+            "utterance": "how much is an overdraft fee for bank",
+            "label": 3
+        },
+        {
+            "utterance": "where is the dipstick",
+            "label": 3
+        },
+        {
+            "utterance": "where is the dipstick",
+            "label": 3
+        },
+        {
+            "utterance": "where is the dipstick",
+            "label": 3
+        },
+        {
+            "utterance": "where is the dipstick",
+            "label": 3
+        },
+        {
+            "utterance": "where is the dipstick",
+            "label": 3
+        },
+        {
+            "utterance": "where is the dipstick",
+            "label": 3
+        },
+        {
+            "utterance": "how much is 1 share of aapl"
+        },
+        {
+            "utterance": "how is glue made"
+        },
+        {
+            "utterance": "how much is 1 share of aapl"
+        },
+        {
+            "utterance": "how is glue made"
+        },
+        {
+            "utterance": "how much is 1 share of aapl"
+        },
+        {
+            "utterance": "how is glue made"
+        },
+        {
+            "utterance": "how much is 1 share of aapl"
+        },
+        {
+            "utterance": "how is glue made"
+        }
+    ],
+    "test": [
+        {
+            "utterance": "can i make a reservation for redrobin",
+            "label": 0
+        },
+        {
+            "utterance": "does redrobin do reservations",
+            "label": 0
+        },
+        {
+            "utterance": "does acero in maplewood allow reservations",
+            "label": 0
+        },
+        {
+            "utterance": "i think my account is blocked",
+            "label": 1
+        },
+        {
+            "utterance": "why is my bank account stopping all transactions from going through",
+            "label": 1
+        },
+        {
+            "utterance": "what would cause me to be locked out of my bank account",
+            "label": 1
+        },
+        {
+            "utterance": "find out the reason why am i locked out of my bank account",
+            "label": 1
+        },
+        {
+            "utterance": "make sure my alarm is set for three thirty in the morning",
+            "label": 2
+        },
+        {
+            "utterance": "please set an alarm for mid day",
+            "label": 2
+        },
+        {
+            "utterance": "have an alarm set for three in the morning",
+            "label": 2
+        },
+        {
+            "utterance": "set an alarm for me for 10:00 and another one set for 4:00",
+            "label": 2
         },
         {
-            "utterance": "why are exponents preformed before multiplication in the order of operations"
+            "utterance": "set an alarm to go to sleep and another to wake up",
+            "label": 2
         },
         {
-            "utterance": "what size wipers does this car take"
+            "utterance": "how much is an overdraft fee for bank",
+            "label": 3
         },
         {
-            "utterance": "where is the dipstick"
+            "utterance": "why are exponents preformed before multiplication in the order of operations",
+            "label": 3
+        },
+        {
+            "utterance": "what size wipers does this car take",
+            "label": 3
         },
         {
             "utterance": "how much is 1 share of aapl"
         },
+        {
+            "utterance": "how is glue made"
+        },
         {
             "utterance": "how is glue made"
         }

tests/context/datahandler/test_stratificaiton.py

@@ -12,8 +12,8 @@ def test_train_test_split(dataset):
 
     assert Split.TRAIN in dataset
     assert Split.TEST in dataset
-    assert dataset[Split.TRAIN].num_rows == 24
-    assert dataset[Split.TEST].num_rows == 6
+    assert dataset[Split.TRAIN].num_rows == 29
+    assert dataset[Split.TEST].num_rows == 8
     assert dataset.get_n_classes(Split.TRAIN) == dataset.get_n_classes(Split.TEST)
 
 
@@ -28,6 +28,6 @@ def test_multilabel_train_test_split(dataset):
 
     assert Split.TRAIN in dataset
     assert Split.TEST in dataset
-    assert dataset[Split.TRAIN].num_rows == 24
-    assert dataset[Split.TEST].num_rows == 6
+    assert dataset[Split.TRAIN].num_rows == 30
+    assert dataset[Split.TEST].num_rows == 7
     assert dataset.get_n_classes(Split.TRAIN) == dataset.get_n_classes(Split.TEST)

tests/modules/prediction/conftest.py

@@ -1,3 +1,4 @@
+import numpy as np
 import pytest
 
 from autointent.context.data_handler import DataHandler
@@ -43,3 +44,8 @@ def multilabel_fit_data(dataset):
     scores = scorer.predict(data_handler.validation_utterances(1) + data_handler.oos_utterances(1))
     labels = data_handler.validation_labels(1) + [[0] * data_handler.n_classes] * len(data_handler.oos_utterances(1))
     return scores, labels
+
+
+@pytest.fixture
+def scores():
+    return np.array([[0.05, 0.9, 0, 0.05], [0.8, 0, 0.1, 0.1], [0, 0.2, 0.7, 0.1]])

tests/modules/prediction/test_adaptive.py

@@ -8,9 +8,9 @@
 def test_multilabel(multilabel_fit_data):
     predictor = AdaptivePredictor()
     predictor.fit(*multilabel_fit_data)
-    scores = np.array([[0.2, 0.9, 0], [0.8, 0, 0.6], [0, 0.4, 0.7]])
+    scores = np.array([[0.2, 0.9, 0, 0], [0.8, 0, 0.6, 0], [0, 0.4, 0.7, 0]])
     predictions = predictor.predict(scores)
-    desired = np.array([[0, 1, 0], [1, 0, 1], [0, 1, 1]])
+    desired = np.array([[0, 1, 0, 0], [1, 0, 1, 0], [0, 1, 1, 0]])
 
     np.testing.assert_array_equal(predictions, desired)