[s2s]Use prepare_translation_batch for Marian finetuning (#6293)

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

examples/seq2seq/README.md

@@ -63,7 +63,7 @@ Summarization Tips:
 (It rarely makes sense to start from `bart-large` unless you are a researching finetuning methods).
 
 **Update 2018-07-18**
-Datasets: Seq2SeqDataset will be used for all models besides MBart, for which MBartDataset will be used.**
+Datasets: `Seq2SeqDataset` should be used for all tokenizers without a `prepare_translation_batch` method. For those who do (like Marian, MBart), `TranslationDataset` should be used.**
 A new dataset is needed to support multilingual tasks.
 
 

examples/seq2seq/finetune.py

@@ -14,7 +14,7 @@ import torch
 from torch.utils.data import DataLoader
 
 from lightning_base import BaseTransformer, add_generic_args, generic_train
-from transformers import MBartTokenizer, T5ForConditionalGeneration, get_linear_schedule_with_warmup
+from transformers import MarianTokenizer, MBartTokenizer, T5ForConditionalGeneration, get_linear_schedule_with_warmup
 
 
 try:
@@ -32,7 +32,7 @@ try:
         ROUGE_KEYS,
         calculate_bleu_score,
         Seq2SeqDataset,
-        MBartDataset,
+        TranslationDataset,
         label_smoothed_nll_loss,
     )
 
@@ -40,7 +40,7 @@ try:
 except ImportError:
     from utils import (
         Seq2SeqDataset,
-        MBartDataset,
+        TranslationDataset,
         assert_all_frozen,
         use_task_specific_params,
         lmap,
@@ -108,8 +108,8 @@ class SummarizationModule(BaseTransformer):
         if self.model.config.decoder_start_token_id is None and isinstance(self.tokenizer, MBartTokenizer):
             self.decoder_start_token_id = self.tokenizer.lang_code_to_id[hparams.tgt_lang]
             self.model.config.decoder_start_token_id = self.decoder_start_token_id
-        if isinstance(self.tokenizer, MBartTokenizer):
+        if isinstance(self.tokenizer, MBartTokenizer) or isinstance(self.tokenizer, MarianTokenizer):
-            self.dataset_class = MBartDataset
+            self.dataset_class = TranslationDataset
         else:
             self.dataset_class = Seq2SeqDataset
 

examples/seq2seq/test_seq2seq_examples.py

@@ -14,14 +14,14 @@ from pytest import param
 from torch.utils.data import DataLoader
 
 import lightning_base
-from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, MBartTokenizer
+from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
 from transformers.testing_utils import require_multigpu
 
 from .distillation import distill_main, evaluate_checkpoint
 from .finetune import SummarizationModule, main
 from .pack_dataset import pack_data_dir
 from .run_eval import generate_summaries_or_translations, run_generate
-from .utils import MBartDataset, Seq2SeqDataset, label_smoothed_nll_loss, lmap, load_json
+from .utils import Seq2SeqDataset, TranslationDataset, label_smoothed_nll_loss, lmap, load_json
 
 
 logging.basicConfig(level=logging.DEBUG)
@@ -406,8 +406,9 @@ def test_pack_dataset():
     assert orig_paths == new_paths
 
 
-def test_mbart_dataset_truncation():
+@pytest.mark.parametrize(["tok_name"], [pytest.param(MBART_TINY), pytest.param(MARIAN_TINY)])
-    tokenizer = MBartTokenizer.from_pretrained(MBART_TINY)
+def test_mbart_dataset_truncation(tok_name):
+    tokenizer = AutoTokenizer.from_pretrained(tok_name)
     tmp_dir = make_test_data_dir()
     max_len_source = max(len(tokenizer.encode(a)) for a in ARTICLES)
     max_len_target = max(len(tokenizer.encode(a)) for a in SUMMARIES)
@@ -416,7 +417,7 @@ def test_mbart_dataset_truncation():
     assert max_len_target > max_src_len  # Truncated
     assert max_len_source > max_src_len
     src_lang, tgt_lang = "ro_RO", "de_DE"  # NOT WHAT IT WAS TRAINED ON
-    train_dataset = MBartDataset(
+    train_dataset = TranslationDataset(
         tokenizer,
         data_dir=tmp_dir,
         type_path="train",
@@ -433,6 +434,8 @@ def test_mbart_dataset_truncation():
         assert batch["input_ids"].shape[1] == max_src_len
         # show that targets are the same len
         assert batch["decoder_input_ids"].shape[1] == max_tgt_len
+        if tok_name == MARIAN_TINY:
+            continue
         # check language codes in correct place
         assert batch["decoder_input_ids"][0, 0].item() == tokenizer.lang_code_to_id[tgt_lang]
         assert batch["decoder_input_ids"][0, -1].item() == tokenizer.eos_token_id

examples/seq2seq/utils.py

@@ -146,7 +146,9 @@ class Seq2SeqDataset(Dataset):
         return SortishSampler(self.src_lens, batch_size)
 
 
-class MBartDataset(Seq2SeqDataset):
+class TranslationDataset(Seq2SeqDataset):
+    """A dataset that calls prepare_translation_batch."""
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         if self.max_source_length != self.max_target_length:

src/transformers/tokenization_marian.py

@@ -127,10 +127,12 @@ class MarianTokenizer(PreTrainedTokenizer):
         src_texts: List[str],
         tgt_texts: Optional[List[str]] = None,
         max_length: Optional[int] = None,
+        max_target_length: Optional[int] = None,
         pad_to_max_length: bool = True,
         return_tensors: str = "pt",
         truncation_strategy="only_first",
         padding="longest",
+        **unused,
     ) -> BatchEncoding:
         """Prepare model inputs for translation. For best performance, translate one sentence at a time.
         Arguments:
@@ -162,6 +164,9 @@ class MarianTokenizer(PreTrainedTokenizer):
         if tgt_texts is None:
             return model_inputs
 
+        if max_target_length is not None:
+            tokenizer_kwargs["max_length"] = max_target_length
+
         self.current_spm = self.spm_target
         decoder_inputs: BatchEncoding = self(tgt_texts, **tokenizer_kwargs)
         for k, v in decoder_inputs.items():