Replace -100s in predictions by the pad token (#22693)

* Replace -100s in predictions by the pad token

* Style

* Try to catch them all

examples/pytorch/question-answering/run_seq2seq_qa.py

@@ -26,6 +26,7 @@ from typing import List, Optional, Tuple
 
 import datasets
 import evaluate
+import numpy as np
 from datasets import load_dataset
 from trainer_seq2seq_qa import QuestionAnsweringSeq2SeqTrainer
 
@@ -614,6 +615,8 @@ def main():
         preds = outputs.predictions
         if isinstance(preds, tuple):
             preds = preds[0]
+        # Replace -100s used for padding as we can't decode them
+        preds = np.where(preds != -100, preds, tokenizer.pad_token_id)
         decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
 
         # Build a map example to its corresponding features.

examples/pytorch/summarization/run_summarization.py

@@ -632,9 +632,9 @@ def main():
         preds, labels = eval_preds
         if isinstance(preds, tuple):
             preds = preds[0]
+        # Replace -100s used for padding as we can't decode them
+        preds = np.where(preds != -100, preds, tokenizer.pad_token_id)
         decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
-        if data_args.ignore_pad_token_for_loss:
-            # Replace -100 in the labels as we can't decode them.
         labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
         decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
 
@@ -714,8 +714,10 @@ def main():
 
         if trainer.is_world_process_zero():
             if training_args.predict_with_generate:
+                predictions = predict_results.predictions
+                predictions = np.where(predictions != -100, predictions, tokenizer.pad_token_id)
                 predictions = tokenizer.batch_decode(
-                    predict_results.predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True
+                    predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True
                 )
                 predictions = [pred.strip() for pred in predictions]
                 output_prediction_file = os.path.join(training_args.output_dir, "generated_predictions.txt")

examples/pytorch/translation/run_translation.py

@@ -543,9 +543,9 @@ def main():
         preds, labels = eval_preds
         if isinstance(preds, tuple):
             preds = preds[0]
+        # Replace -100s used for padding as we can't decode them
+        preds = np.where(preds != -100, preds, tokenizer.pad_token_id)
         decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
-        if data_args.ignore_pad_token_for_loss:
-            # Replace -100 in the labels as we can't decode them.
         labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
         decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
 
@@ -626,8 +626,10 @@ def main():
 
         if trainer.is_world_process_zero():
             if training_args.predict_with_generate:
+                predictions = predict_results.predictions
+                predictions = np.where(predictions != -100, predictions, tokenizer.pad_token_id)
                 predictions = tokenizer.batch_decode(
-                    predict_results.predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True
+                    predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True
                 )
                 predictions = [pred.strip() for pred in predictions]
                 output_prediction_file = os.path.join(training_args.output_dir, "generated_predictions.txt")