💡 All-in-one AI framework for semantic search, LLM orchestration and language model workflows
Python agentsaiai-agentsembeddingsinformation-retrievallanguage-modellarge-language-modelsllmnlppythonragretrieval-augmented-generationsearchsearch-enginesemantic-searchsentence-embeddingstransformerstxtaivector-databasevector-search
Update labels pipeline to filter text classification output, closes #116
davidmezzetti5b86e422e05c
diff --git a/src/python/txtai/pipeline/labels.py b/src/python/txtai/pipeline/labels.pyindex bebc4cf0..71214059 100644--- a/src/python/txtai/pipeline/labels.py+++ b/src/python/txtai/pipeline/labels.py@@ -21,8 +21,8 @@ class Labels(HFPipeline):def __call__(self, text, labels=None, multilabel=False):"""Applies a text classifier to text. Returns a list of (id, score) sorted by highest score,- where id is the index in labels. For zero shot classification, a list of labels is required,- otherwise the labels indices are taken from the trained text classification model.+ where id is the index in labels. For zero shot classification, a list of labels is required.+ For text classification models, a list of labels is optional, otherwise all trained labels are returned.This method supports text as a string or a list. If the input is a string, the returntype is a 1D list of (id, score). If text is a list, a 2D list of (id, score) is@@ -30,7 +30,7 @@ class Labels(HFPipeline):Args:text: text|list- labels: list of labels (zero shot classification only)+ labels: list of labelsmultilabel: labels are independent if True, otherwise scores are normalized to sum to 1 per text itemReturns:@@ -54,6 +54,9 @@ class Labels(HFPipeline):if self.dynamic:scores.append([(labels.index(label), result["scores"][x]) for x, label in enumerate(result["labels"])])else:+ # Filter results using labels, if provided+ result = self.limit(result, labels)+scores.append(sorted(enumerate(result), key=lambda x: x[1], reverse=True))return scores[0] if isinstance(text, str) else scores@@ -94,3 +97,34 @@ class Labels(HFPipeline):"""return list(self.pipeline.model.config.id2label.values())++ def limit(self, result, labels):+ """+ Filter result using labels. If labels is None, original result is returned.++ Args:+ result: results array+ labels: list of labels or None++ Returns:+ filtered results+ """++ if labels:+ config = self.pipeline.model.config+ indices = []+ for label in labels:+ # Lookup label keys from model config+ if label.isdigit():+ label = int(label)+ keys = list(config.id2label.keys())+ else:+ label = label.lower()+ keys = [x.lower() for x in config.label2id.keys()]++ # Get index, default to 0 if not found+ indices.append(keys.index(label) if label in keys else 0)++ return [result[i] for i in indices]++ return resultdiff --git a/test/python/testlabels.py b/test/python/testlabels.pyindex 95de9936..8b168002 100644--- a/test/python/testlabels.py+++ b/test/python/testlabels.py@@ -75,6 +75,17 @@ class TestLabels(unittest.TestCase):results = [r[0][0] for r in self.similarity(["feel good story", "climate change"], self.data)]self.assertEqual(results, [4, 1])+ def testSimilarityFixed(self):+ """+ Test similarity with a fixed label text classification model+ """++ similarity = Similarity(dynamic=False)++ # Test with query as label text and label id+ self.assertLessEqual(similarity("negative", ["This is the best sentence ever"])[0][1], 0.1)+ self.assertLessEqual(similarity("0", ["This is the best sentence ever"])[0][1], 0.1)+def testSimilarityLong(self):"""Test similarity with long text