Merge pull request #8 from sign-language-processing/copilot/ensure-float-return-values

Author: AmitMY

signwriting_evaluation/metrics/similarity.py

@@ -129,7 +129,7 @@ def length_acc(self, hyp: Sign, ref: Sign) -> float:
     def error_rate(self, hyp: Sign, ref: Sign) -> float:
         # Calculate the evaluate score for a given hypothesis and ref.
         if not hyp["symbols"] or not ref["symbols"]:
-            return 1
+            return 1.0
 
         cost_matrix = np.array(
             [self.symbols_score(first, second) for first in hyp["symbols"] for second in ref["symbols"]])
@@ -150,7 +150,7 @@ def score_single_sign(self, hypothesis: str, reference: str) -> float:
 
     def score(self, hypothesis: Optional[str], reference: Optional[str]) -> float:
         if hypothesis is None or reference is None:
-            return 0
+            return 0.0
 
         # Here, hypothesis and reference are both FSW strings of potentially different number of signs
         hypothesis_signs = text_to_signs(hypothesis)

signwriting_evaluation/metrics/test_float_returns.py

@@ -0,0 +1,68 @@
+import unittest
+
+from signwriting_evaluation.metrics.similarity import SignWritingSimilarityMetric
+from signwriting_evaluation.metrics.bleu import SignWritingBLEU
+from signwriting_evaluation.metrics.chrf import SignWritingCHRF
+
+
+class TestMetricsReturnFloat(unittest.TestCase):
+    """Test that all metrics return float values, never integers."""
+
+    def setUp(self):
+        self.similarity = SignWritingSimilarityMetric()
+        self.bleu = SignWritingBLEU()
+        self.chrf = SignWritingCHRF()
+
+    def test_similarity_returns_float(self):
+        """Test that similarity metric always returns float."""
+        # Test with None values (edge case that was returning int)
+        score = self.similarity.score(None, None)
+        self.assertIsInstance(score, float, "Similarity score with None inputs should be float")
+
+        # Test with valid FSW strings
+        hypothesis = "M530x538S37602508x462S15a11493x494S20e00488x510S22f03469x517"
+        reference = "M519x534S37900497x466S3770b497x485S15a51491x501S22f03481x513"
+        score = self.similarity.score(hypothesis, reference)
+        self.assertIsInstance(score, float, "Similarity score should be float")
+
+        # Test with None hypothesis
+        score = self.similarity.score(None, reference)
+        self.assertIsInstance(score, float, "Similarity score with None hypothesis should be float")
+
+        # Test with None reference
+        score = self.similarity.score(hypothesis, None)
+        self.assertIsInstance(score, float, "Similarity score with None reference should be float")
+
+    def test_bleu_returns_float(self):
+        """Test that BLEU metric always returns float."""
+        hypothesis = "M530x538S37602508x462S15a11493x494S20e00488x510S22f03469x517"
+        reference = "M519x534S37900497x466S3770b497x485S15a51491x501S22f03481x513"
+        score = self.bleu.score(hypothesis, reference)
+        self.assertIsInstance(score, float, "BLEU score should be float")
+
+    def test_chrf_returns_float(self):
+        """Test that CHRF metric always returns float."""
+        hypothesis = "M530x538S37602508x462S15a11493x494S20e00488x510S22f03469x517"
+        reference = "M519x534S37900497x466S3770b497x485S15a51491x501S22f03481x513"
+        score = self.chrf.score(hypothesis, reference)
+        self.assertIsInstance(score, float, "CHRF score should be float")
+
+    def test_similarity_score_all_returns_float(self):
+        """Test that score_all returns list of lists of floats."""
+        hypotheses = ["M530x538S37602508x462S15a11493x494S20e00488x510S22f03469x517"]
+        references = ["M519x534S37900497x466S3770b497x485S15a51491x501S22f03481x513"]
+        scores = self.similarity.score_all(hypotheses, references, progress_bar=False)
+        for row in scores:
+            for score in row:
+                self.assertIsInstance(score, float, "Each score in score_all should be float")
+
+    def test_similarity_corpus_score_returns_float(self):
+        """Test that corpus_score returns float."""
+        hypotheses = ["M530x538S37602508x462S15a11493x494S20e00488x510S22f03469x517"]
+        references = [["M519x534S37900497x466S3770b497x485S15a51491x501S22f03481x513"]]
+        score = self.similarity.corpus_score(hypotheses, references)
+        self.assertIsInstance(score, float, "Corpus score should be float")
+
+
+if __name__ == '__main__':
+    unittest.main()