MBART: support summarization tasks where max_src_len > max_tgt_len (#6003)

* MBART: support summarization tasks

* fix test

* Style

* add tokenizer test

examples/seq2seq/README.md

@@ -180,6 +180,8 @@ python run_eval.py sshleifer/distilbart-cnn-12-6 $DATA_DIR/val.source dbart_val_
     --task summarization \
     --n_obs 100 \
     --device cuda \
+    --max_source_length 1024 \
+    --max_target_length 56 \
     --fp16 \
     --bs 32
 ```

examples/seq2seq/finetune.py

@@ -105,7 +105,13 @@ class SummarizationModule(BaseTransformer):
         self.hparams.git_sha = get_git_info()["repo_sha"]
         self.num_workers = hparams.num_workers
         self.decoder_start_token_id = None
-        self.dataset_class = Seq2SeqDataset
+        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):
+            self.dataset_class = MBartDataset
+        else:
+            self.dataset_class = Seq2SeqDataset
 
     def freeze_embeds(self):
         """Freeze token embeddings and positional embeddings for bart, just token embeddings for t5."""
@@ -331,11 +337,6 @@ class TranslationModule(SummarizationModule):
         super().__init__(hparams, **kwargs)
         self.dataset_kwargs["src_lang"] = hparams.src_lang
         self.dataset_kwargs["tgt_lang"] = hparams.tgt_lang
-        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):
-            self.dataset_class = MBartDataset
 
     def calc_generative_metrics(self, preds, target) -> dict:
         return calculate_bleu_score(preds, target)

examples/seq2seq/finetune_t5.sh

@@ -8,6 +8,7 @@ python finetune.py \
 --eval_batch_size=$BS \
 --output_dir=$OUTPUT_DIR \
 --max_source_length=512 \
+--max_target_length=56 \
 --val_check_interval=0.1 --n_val=200 \
 --do_train --do_predict \
  $@

examples/seq2seq/test_seq2seq_examples.py

@@ -300,14 +300,17 @@ def test_mbart_dataset_truncation():
     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)
-    trunc = 4
+    max_src_len = 4
+    max_tgt_len = 8
+    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(
         tokenizer,
         data_dir=tmp_dir,
         type_path="train",
-        max_source_length=trunc,
-        max_target_length=1000,  # ignored
+        max_source_length=max_src_len,
+        max_target_length=max_tgt_len,  # ignored
         src_lang=src_lang,
         tgt_lang=tgt_lang,
     )
@@ -316,17 +319,15 @@ def test_mbart_dataset_truncation():
         assert isinstance(batch, dict)
         assert batch["attention_mask"].shape == batch["input_ids"].shape
         # show that articles were trimmed.
-        assert batch["input_ids"].shape[1] == trunc
+        assert batch["input_ids"].shape[1] == max_src_len
         # show that targets are the same len
-        assert batch["decoder_input_ids"].shape[1] == trunc
+        assert batch["decoder_input_ids"].shape[1] == max_tgt_len
         # 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
         assert batch["input_ids"][0, -2].item() == tokenizer.eos_token_id
         assert batch["input_ids"][0, -1].item() == tokenizer.lang_code_to_id[src_lang]
 
-        assert max_len_target > trunc  # Truncated
-        assert max_len_source > trunc
         break  # No need to test every batch
 
 

examples/seq2seq/utils.py

@@ -157,7 +157,8 @@ class MBartDataset(Seq2SeqDataset):
         super().__init__(*args, **kwargs)
         if self.max_source_length != self.max_target_length:
             warnings.warn(
-                f"Mbart will ignore max_target_length = {self.max_target_length} and use {self.max_source_length} for both sides."
+                f"Mbart is using sequence lengths {self.max_source_length}, {self.max_target_length}. "
+                f"Imbalanced sequence lengths may be undesired for translation tasks"
             )
 
     def __getitem__(self, index) -> Dict[str, str]:
@@ -178,6 +179,7 @@ class MBartDataset(Seq2SeqDataset):
             tgt_texts=[x["tgt_texts"] for x in batch],
             tgt_lang=self.tgt_lang,
             max_length=self.max_source_length,
+            max_target_length=self.max_target_length,
         )
         return batch_encoding.data