Update README.md

README.md

@@ -73,40 +73,56 @@ reuters = Dataset.from_hub("AutoIntent/reuters")
 This dataset is taken from `ucirvine/reuters21578` and formatted with our [AutoIntent Library](https://deeppavlov.github.io/AutoIntent/index.html):
 
 ```python
-from collections import defaultdict
 from autointent import Dataset
 import datasets
 
-# load original data
-reuters = datasets.load_dataset("ucirvine/reuters21578", "ModHayes", trust_remote_code=True)
-
-# remove low-resource classes
-counter = defaultdict(int)
-for batch in reuters["train"].iter(batch_size=16):
-    for labels in batch["topics"]:
-        for lab in labels:
-            counter[lab] += 1
-names_to_remove = [name for name, cnt in counter.items() if cnt < 10]
-
-intent_names = sorted(set(name for intents in reuters["train"]["topics"] for name in intents))
-for n in names_to_remove:
-    intent_names.remove(n)
-name_to_id = {name: i for i, name in enumerate(intent_names)}
-
-# extract only texts and labels
-def transform(ds: datasets.Dataset) -> list[dict]:
-    def _transform(example: dict):
-        return {
-            "utterance": example["text"],
-            "label": [int(name in example["topics"]) for name in intent_names if name not in names_to_remove]
-        }
-    ds = ds.map(_transform, remove_columns=ds.features.keys())
+
+
+def get_intents_info(ds: datasets.DatasetDict) -> list[str]:
+    return sorted(set(name for intents in ds["train"]["topics"] for name in intents))
+
+def parse(ds: datasets.Dataset, intent_names: list[str]) -> list[dict]:
+    return [{
+        "utterance": example["text"],
+        "label": [int(name in example["topics"]) for name in intent_names]
+    } for example in ds]
+
+
+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: datasets.Dataset, 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
+
+def remove_oos(ds: list[dict]):
     return [sample for sample in ds if sum(sample["label"]) != 0]
 
-train = transform(reuters["train"])
-test = transform(reuters["test"])
 
-# format
-intents = [{"id": i, "name": name} for i, name in enumerate(intent_names)]
-reuters_converted = Dataset.from_dict({"intents": intents, "train": train, "test": test})
+if __name__ == "__main__":
+    reuters = datasets.load_dataset("ucirvine/reuters21578", "ModHayes", trust_remote_code=True)
+    intent_names = get_intents_info(reuters)
+    train_parsed = parse(reuters["train"], intent_names)
+    test_parsed = parse(reuters["test"], intent_names)
+    mask = get_low_resource_classes_mask(train_parsed, intent_names)
+    intent_names = [name for i, name in enumerate(intent_names) if not mask[i]]
+    train_filtered = remove_oos(remove_low_resource_classes(train_parsed, mask))
+    test_filtered = remove_oos(remove_low_resource_classes(test_parsed, mask))
+
+    intents = [{"id": i, "name": name} for i, name in enumerate(intent_names)]
+    reuters_converted = Dataset.from_dict({"intents": intents, "train": train_filtered, "test": test_filtered})
 ```