Fixes the incorrect token prediction distribution from _all_scores_for_token() in sequence_tagger_model.py

flair/models/sequence_tagger_model.py

@@ -589,7 +589,7 @@ def _all_scores_for_token(self, sentences: List[Sentence], score_tensor: torch.T
         previous = 0
         for length in lengths:
             prob_tags_per_sentence.append(prob_all_tags[previous : previous + length])
-            previous = length
+            previous += length
         return prob_tags_per_sentence
 
     def _get_state_dict(self):

tests/models/test_sequence_tagger.py

@@ -1,4 +1,6 @@
 import pytest
+import torch
+import torch.nn.functional as F
 
 import flair
 from flair.embeddings import FlairEmbeddings, WordEmbeddings
@@ -121,3 +123,36 @@ def test_train_load_use_tagger_disjunct_tags(
         loaded_model.predict([example_sentence, self.empty_sentence])
         loaded_model.predict([self.empty_sentence])
         del loaded_model
+
+    @pytest.mark.integration()
+    def test_all_token_prob_distribution(self, embeddings, corpus):
+        tag_dictionary = corpus.make_label_dictionary("ner", add_unk=False)
+        model = self.build_model(embeddings, tag_dictionary)
+
+        # get features from forward propagation
+        sentences = [corpus.train[i] for i in range(len(corpus.train))]
+
+        # reverse sort all sequences by their length
+        sentences = sorted(sentences, key=len, reverse=True)
+
+        with torch.no_grad():
+            sentence_tensor, lengths = model._prepare_tensors(sentences)
+            features = model.forward(sentence_tensor, lengths)
+
+        # remove previously predicted labels of this type
+        for sentence in sentences:
+            sentence.remove_labels(model.label_type)
+
+        softmax_batch = F.softmax(features, dim=1).cpu()
+        lengths = [len(sentence) for sentence in sentences]
+        all_tokens_prob_distrib = model._all_scores_for_token(sentences, softmax_batch, lengths)
+
+        for i, sen_tokens_prob_distribution in enumerate(all_tokens_prob_distrib):
+            assert len(sen_tokens_prob_distribution) == lengths[i]
+            for token_prob_distrib, token in zip(sen_tokens_prob_distribution, sentences[i]):
+                assert len(token_prob_distrib) == len(model.label_dictionary)
+                score_sum = 0.0
+                for token_label in token_prob_distrib:
+                    assert token_label.data_point == token
+                    score_sum += token_label.score
+                assert abs(score_sum - 1.0) < 1.0e-5