Merge pull request #1238 from zalandoresearch/GH-1115-warning-validat…

…ion-data

GH-1115: Add warning if validation data too small

flair/embeddings.py

@@ -581,7 +581,6 @@ def _add_embeddings_internal(self, sentences: List[Sentence]) -> List[Sentence]:
         for i, sentence in enumerate(sentences):
 
             language_code = sentence.get_language_code()
-            print(language_code)
             supported = [
                 "en",
                 "de",

flair/models/language_model.py

@@ -354,8 +354,6 @@ def generate_text(
             if not self.is_forward_lm:
                 text = text[::-1]
 
-            text = text.encode("utf-8")
-
             return text, log_prob
 
     def calculate_perplexity(self, text: str) -> float:

flair/trainers/language_model_trainer.py

@@ -278,6 +278,13 @@ def train(
 
         val_data = self._batchify(self.corpus.valid, mini_batch_size)
 
+        # error message if the validation dataset is too small
+        if val_data.size(0) == 1:
+            raise RuntimeError(
+                f"ERROR: Your validation dataset is too small. For your mini_batch_size, the data needs to "
+                f"consist of at least {mini_batch_size * 2} characters!"
+            )
+
         base_path.mkdir(parents=True, exist_ok=True)
         loss_txt = base_path / "loss.txt"
         savefile = base_path / "best-lm.pt"