Update README.md

README.md

@@ -74,21 +74,12 @@ from datasets import Dataset as HFDataset
 from datasets import load_dataset
 
 from autointent import Dataset
-from autointent.schemas import Intent, Sample
+from autointent.schemas import Intent
 
-# these classes contain too few sampls
-names_to_remove = [
-    "partnerships & alliances",
-    "patent publication",
-    "subsidiary establishment",
-    "department establishment",
-]
 
 def extract_intents_data(events_dataset: HFDataset) -> list[Intent]:
     """Extract intent names and assign ids to them."""
     intent_names = sorted({name for intents in events_dataset["train"]["all_labels"] for name in intents})
-    for n in names_to_remove:
-        intent_names.remove(n)
     return [Intent(id=i,name=name) for i, name in enumerate(intent_names)]
 
 
@@ -105,30 +96,42 @@ def converting_mapping(example: dict, intents_data: list[Intent]) -> dict[str, s
     return res
 
 
-def convert_events(events_split: HFDataset, intents_data: dict[str, int]) -> list[Sample]:
+def convert_events(events_split: HFDataset, intents_data: dict[str, int]) -> list[dict]:
     """Convert one split into desired format."""
     events_split = events_split.map(
         converting_mapping, remove_columns=events_split.features.keys(),
         fn_kwargs={"intents_data": intents_data}
     )
 
-    samples = []
-    for sample in events_split.to_list():
-        if sample["utterance"] is None:
-            continue
-        samples.append(sample)
+    return [sample for sample in events_split if sample["utterance"] is not None]
 
-    mask = [sample["label"] is None for sample in samples]
-    n_oos_samples = sum(mask)
-    n_in_domain_samples = len(samples) - n_oos_samples
-    
-    print(f"{n_oos_samples=}")
-    print(f"{n_in_domain_samples=}\n")
 
-    # actually there are too few oos samples to include them, so filter out
-    samples = list(filter(lambda sample: sample["label"] is not None, samples))
+def get_low_resource_classes_mask(ds: list[dict], intent_names: list[str], fraction_thresh: float = 0.01) -> list[bool]:
+    res = [0] * len(intent_names)
+    for sample in ds:
+        for i, indicator in enumerate(sample["label"]):
+            res[i] += indicator
+    for i in range(len(intent_names)):
+        res[i] /= len(ds)
+    return [(frac < fraction_thresh) for frac in res]
+
+
+def remove_low_resource_classes(ds: list[dict], mask: list[bool]) -> list[dict]:
+    res = []
+    for sample in ds:
+        if sum(sample["label"]) == 1 and mask[sample["label"].index(1)]:
+            continue
+        sample["label"] = [
+            indicator for indicator, low_resource in
+            zip(sample["label"], mask, strict=True) if not low_resource
+        ]
+        res.append(sample)
+    return res
+
 
-    return [Sample(**sample) for sample in samples]
+def remove_oos(ds: list[dict]):
+    return [sample for sample in ds if sum(sample["label"]) != 0]
+    
 
 if __name__ == "__main__":
     # `load_dataset` might not work
@@ -140,6 +143,11 @@ if __name__ == "__main__":
     train_samples = convert_events(events_dataset["train"], intents_data)
     test_samples = convert_events(events_dataset["test"], intents_data)
 
+    intents_names = [intent.name for intent in intents_data]
+    mask = get_low_resource_classes_mask(train_samples, intents_names)
+    train_samples = remove_oos(remove_low_resource_classes(train_samples, mask))
+    test_samples = remove_oos(remove_low_resource_classes(test_samples, mask))
+
     events_converted = Dataset.from_dict(
         {"train": train_samples, "test": test_samples, "intents": intents_data}
     )