Only resize embeddings when necessary (#20043)

* Only resize embeddings when necessary

* Add comment

examples/pytorch/language-modeling/run_clm.py

@@ -387,7 +387,11 @@ def main():
         n_params = sum(dict((p.data_ptr(), p.numel()) for p in model.parameters()).values())
         logger.info(f"Training new model from scratch - Total size={n_params/2**20:.2f}M params")
 
-    model.resize_token_embeddings(len(tokenizer))
+    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+    # on a small vocab and want a smaller embedding size, remove this test.
+    embedding_size = model.get_input_embeddings().weight.shape[0]
+    if len(tokenizer) > embedding_size:
+        model.resize_token_embeddings(len(tokenizer))
 
     # Preprocessing the datasets.
     # First we tokenize all the texts.

examples/pytorch/language-modeling/run_clm_no_trainer.py

@@ -378,7 +378,11 @@ def main():
         logger.info("Training new model from scratch")
         model = AutoModelForCausalLM.from_config(config)
 
-    model.resize_token_embeddings(len(tokenizer))
+    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+    # on a small vocab and want a smaller embedding size, remove this test.
+    embedding_size = model.get_input_embeddings().weight.shape[0]
+    if len(tokenizer) > embedding_size:
+        model.resize_token_embeddings(len(tokenizer))
 
     # Preprocessing the datasets.
     # First we tokenize all the texts.

examples/pytorch/language-modeling/run_mlm.py

@@ -389,7 +389,11 @@ def main():
         logger.info("Training new model from scratch")
         model = AutoModelForMaskedLM.from_config(config)
 
-    model.resize_token_embeddings(len(tokenizer))
+    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+    # on a small vocab and want a smaller embedding size, remove this test.
+    embedding_size = model.get_input_embeddings().weight.shape[0]
+    if len(tokenizer) > embedding_size:
+        model.resize_token_embeddings(len(tokenizer))
 
     # Preprocessing the datasets.
     # First we tokenize all the texts.

examples/pytorch/language-modeling/run_mlm_no_trainer.py

@@ -383,7 +383,11 @@ def main():
         logger.info("Training new model from scratch")
         model = AutoModelForMaskedLM.from_config(config)
 
-    model.resize_token_embeddings(len(tokenizer))
+    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+    # on a small vocab and want a smaller embedding size, remove this test.
+    embedding_size = model.get_input_embeddings().weight.shape[0]
+    if len(tokenizer) > embedding_size:
+        model.resize_token_embeddings(len(tokenizer))
 
     # Preprocessing the datasets.
     # First we tokenize all the texts.

examples/pytorch/language-modeling/run_plm.py

@@ -376,7 +376,11 @@ def main():
         logger.info("Training new model from scratch")
         model = XLNetLMHeadModel(config)
 
-    model.resize_token_embeddings(len(tokenizer))
+    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+    # on a small vocab and want a smaller embedding size, remove this test.
+    embedding_size = model.get_input_embeddings().weight.shape[0]
+    if len(tokenizer) > embedding_size:
+        model.resize_token_embeddings(len(tokenizer))
 
     # Preprocessing the datasets.
     # First we tokenize all the texts.

examples/pytorch/multiple-choice/run_swag_no_trainer.py

@@ -398,7 +398,11 @@ def main():
         logger.info("Training new model from scratch")
         model = AutoModelForMultipleChoice.from_config(config)
 
-    model.resize_token_embeddings(len(tokenizer))
+    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+    # on a small vocab and want a smaller embedding size, remove this test.
+    embedding_size = model.get_input_embeddings().weight.shape[0]
+    if len(tokenizer) > embedding_size:
+        model.resize_token_embeddings(len(tokenizer))
 
     # Preprocessing the datasets.
     # First we tokenize all the texts.

examples/pytorch/question-answering/run_seq2seq_qa.py

@@ -380,7 +380,11 @@ def main():
         use_auth_token=True if model_args.use_auth_token else None,
     )
 
-    model.resize_token_embeddings(len(tokenizer))
+    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+    # on a small vocab and want a smaller embedding size, remove this test.
+    embedding_size = model.get_input_embeddings().weight.shape[0]
+    if len(tokenizer) > embedding_size:
+        model.resize_token_embeddings(len(tokenizer))
 
     if model.config.decoder_start_token_id is None:
         raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")

examples/pytorch/summarization/run_summarization.py

@@ -422,7 +422,11 @@ def main():
         use_auth_token=True if model_args.use_auth_token else None,
     )
 
-    model.resize_token_embeddings(len(tokenizer))
+    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+    # on a small vocab and want a smaller embedding size, remove this test.
+    embedding_size = model.get_input_embeddings().weight.shape[0]
+    if len(tokenizer) > embedding_size:
+        model.resize_token_embeddings(len(tokenizer))
 
     if model.config.decoder_start_token_id is None and isinstance(tokenizer, (MBartTokenizer, MBartTokenizerFast)):
         if isinstance(tokenizer, MBartTokenizer):

examples/pytorch/summarization/run_summarization_no_trainer.py

@@ -439,7 +439,11 @@ def main():
         logger.info("Training new model from scratch")
         model = AutoModelForSeq2SeqLM.from_config(config)
 
-    model.resize_token_embeddings(len(tokenizer))
+    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+    # on a small vocab and want a smaller embedding size, remove this test.
+    embedding_size = model.get_input_embeddings().weight.shape[0]
+    if len(tokenizer) > embedding_size:
+        model.resize_token_embeddings(len(tokenizer))
     if model.config.decoder_start_token_id is None:
         raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
 

examples/pytorch/token-classification/run_ner_no_trainer.py

@@ -414,7 +414,13 @@ def main():
         logger.info("Training new model from scratch")
         model = AutoModelForTokenClassification.from_config(config)
 
-    model.resize_token_embeddings(len(tokenizer))
+    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+    # on a small vocab and want a smaller embedding size, remove this test.
+    embedding_size = model.get_input_embeddings().weight.shape[0]
+    if len(tokenizer) > embedding_size:
+        embedding_size = model.get_input_embeddings().weight.shape[0]
+    if len(tokenizer) > embedding_size:
+        model.resize_token_embeddings(len(tokenizer))
 
     # Model has labels -> use them.
     if model.config.label2id != PretrainedConfig(num_labels=num_labels).label2id:

examples/pytorch/translation/run_translation.py

@@ -380,7 +380,11 @@ def main():
         use_auth_token=True if model_args.use_auth_token else None,
     )
 
-    model.resize_token_embeddings(len(tokenizer))
+    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+    # on a small vocab and want a smaller embedding size, remove this test.
+    embedding_size = model.get_input_embeddings().weight.shape[0]
+    if len(tokenizer) > embedding_size:
+        model.resize_token_embeddings(len(tokenizer))
 
     # Set decoder_start_token_id
     if model.config.decoder_start_token_id is None and isinstance(tokenizer, (MBartTokenizer, MBartTokenizerFast)):

examples/pytorch/translation/run_translation_no_trainer.py

@@ -411,7 +411,11 @@ def main():
         logger.info("Training new model from scratch")
         model = AutoModelForSeq2SeqLM.from_config(config)
 
-    model.resize_token_embeddings(len(tokenizer))
+    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+    # on a small vocab and want a smaller embedding size, remove this test.
+    embedding_size = model.get_input_embeddings().weight.shape[0]
+    if len(tokenizer) > embedding_size:
+        model.resize_token_embeddings(len(tokenizer))
 
     # Set decoder_start_token_id
     if model.config.decoder_start_token_id is None and isinstance(tokenizer, (MBartTokenizer, MBartTokenizerFast)):

examples/tensorflow/language-modeling/run_clm.py

@@ -473,7 +473,11 @@ def main():
             logger.info("Training new model from scratch")
             model = TFAutoModelForCausalLM.from_config(config)
 
-        model.resize_token_embeddings(len(tokenizer))
+        # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+        # on a small vocab and want a smaller embedding size, remove this test.
+        embedding_size = model.get_input_embeddings().weight.shape[0]
+        if len(tokenizer) > embedding_size:
+            model.resize_token_embeddings(len(tokenizer))
         # endregion
 
         # region TF Dataset preparation

examples/tensorflow/language-modeling/run_mlm.py

@@ -489,7 +489,11 @@ def main():
             logger.info("Training new model from scratch")
             model = TFAutoModelForMaskedLM.from_config(config)
 
-        model.resize_token_embeddings(len(tokenizer))
+        # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+        # on a small vocab and want a smaller embedding size, remove this test.
+        embedding_size = model.get_input_embeddings().weight.shape[0]
+        if len(tokenizer) > embedding_size:
+            model.resize_token_embeddings(len(tokenizer))
         # endregion
 
         # region TF Dataset preparation

examples/tensorflow/summarization/run_summarization.py

@@ -516,7 +516,11 @@ def main():
             use_auth_token=True if model_args.use_auth_token else None,
         )
 
-        model.resize_token_embeddings(len(tokenizer))
+        # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+        # on a small vocab and want a smaller embedding size, remove this test.
+        embedding_size = model.get_input_embeddings().weight.shape[0]
+        if len(tokenizer) > embedding_size:
+            model.resize_token_embeddings(len(tokenizer))
         # endregion
 
         # region Prepare TF Dataset objects

examples/tensorflow/token-classification/run_ner.py

@@ -385,7 +385,11 @@ def main():
             logger.info("Training new model from scratch")
             model = TFAutoModelForTokenClassification.from_config(config)
 
-        model.resize_token_embeddings(len(tokenizer))
+        # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+        # on a small vocab and want a smaller embedding size, remove this test.
+        embedding_size = model.get_input_embeddings().weight.shape[0]
+        if len(tokenizer) > embedding_size:
+            model.resize_token_embeddings(len(tokenizer))
         # endregion
 
         # region Create TF datasets

examples/tensorflow/translation/run_translation.py

@@ -469,7 +469,11 @@ def main():
             use_auth_token=True if model_args.use_auth_token else None,
         )
 
-        model.resize_token_embeddings(len(tokenizer))
+        # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
+        # on a small vocab and want a smaller embedding size, remove this test.
+        embedding_size = model.get_input_embeddings().weight.shape[0]
+        if len(tokenizer) > embedding_size:
+            model.resize_token_embeddings(len(tokenizer))
         if isinstance(tokenizer, tuple(MULTILINGUAL_TOKENIZERS)):
             model.config.forced_bos_token_id = forced_bos_token_id
         # endregion