boom boom

examples/seq2seq/README.md

@@ -106,7 +106,7 @@ The following command should work on a 16GB GPU:
     --train_batch_size=1 \
     --eval_batch_size=1 \
     --output_dir=xsum_results \
-    --num_train_epochs 1 \
+    --num_train_epochs 6 \
     --model_name_or_path facebook/bart-large
 ```
 
@@ -122,8 +122,8 @@ Best performing command:
 export ENRO_DIR='wmt_en_ro' # Download instructions above
 # export WANDB_PROJECT="MT" # optional
 export MAX_LEN=128
-export BS=4
-./train_mbart_cc25_enro.sh --output_dir enro_finetune_baseline --label_smoothing 0.1 --fp16_opt_level=O1 --logger_name wandb --sortish_sampler
+export BS=8
+./train_mbart_cc25_enro.sh --output_dir enro_finetune_baseline_dropper --label_smoothing 0 --fp16_opt_level=O1 --logger_name wandb --sortish_sampler
 ```
 This should take < 6h/epoch on a 16GB v100 and achieve test BLEU above 26
 To get results in line with fairseq, you need to do some postprocessing. (see `romanian_postprocessing.md`)
@@ -141,7 +141,7 @@ export BS=4
 As you train, `output_dir` will be filled with files, that look kind of like this (comments are mine).
 Some of them are metrics, some of them are checkpoints, some of them are metadata. Here is a quick tour:
 
-```bash
+```
 output_dir
 ├── best_tfmr  # this is a huggingface checkpoint generated by save_pretrained. It is the same model as the PL .ckpt file below
 │   ├── config.json

examples/seq2seq/finetune.py

@@ -14,6 +14,7 @@ from torch.utils.data import DataLoader
 
 from lightning_base import BaseTransformer, add_generic_args, generic_train
 from transformers import MarianTokenizer, MBartTokenizer, T5ForConditionalGeneration
+from transformers.modeling_bart import shift_tokens_right
 
 
 try:
@@ -36,6 +37,7 @@ try:
     )
 
     from .callbacks import Seq2SeqLoggingCallback, get_checkpoint_callback, get_early_stopping_callback
+    from .loss_dropper import LossDropper
 except ImportError:
     from utils import (
         Seq2SeqDataset,
@@ -55,19 +57,21 @@ except ImportError:
         label_smoothed_nll_loss,
     )
     from callbacks import Seq2SeqLoggingCallback, get_checkpoint_callback, get_early_stopping_callback
+    from loss_dropper import LossDropper
 
 logger = logging.getLogger(__name__)
 
 
 class SummarizationModule(BaseTransformer):
     mode = "summarization"
-    loss_names = ["loss"]
+    loss_names = ["loss", "dropper_mask_mean"]
     metric_names = ROUGE_KEYS
     val_metric = "rouge2"
 
     def __init__(self, hparams, **kwargs):
         super().__init__(hparams, num_labels=None, mode=self.mode, **kwargs)
         use_task_specific_params(self.model, "summarization")
+        self.dropper = LossDropper(dropc=.05)
         save_git_info(self.hparams.output_dir)
         self.metrics_save_path = Path(self.output_dir) / "metrics.json"
         self.hparams_save_path = Path(self.output_dir) / "hparams.pkl"
@@ -137,7 +141,10 @@ class SummarizationModule(BaseTransformer):
         pad_token_id = self.tokenizer.pad_token_id
         source_ids, source_mask, target_ids = batch["input_ids"], batch["attention_mask"], batch["decoder_input_ids"]
 
-        if isinstance(self.model, T5ForConditionalGeneration):
+        if "labels" in batch:
+            lm_labels = batch["labels"]
+            decoder_input_ids = shift_tokens_right(lm_labels, pad_token_id)
+        elif isinstance(self.model, T5ForConditionalGeneration):
             decoder_input_ids = self.model._shift_right(target_ids)
             lm_labels = target_ids
         else:
@@ -145,19 +152,32 @@ class SummarizationModule(BaseTransformer):
             lm_labels = target_ids[:, 1:].clone()  # why clone?
 
         outputs = self(source_ids, attention_mask=source_mask, decoder_input_ids=decoder_input_ids, use_cache=False)
+        bs = source_ids.shape[0]
 
         if self.hparams.label_smoothing == 0:
+
             # Same behavior as modeling_bart.py
-            loss_fct = torch.nn.CrossEntropyLoss(ignore_index=pad_token_id)
+            loss_fct = torch.nn.CrossEntropyLoss(reduction='none', ignore_index=pad_token_id)
             lm_logits = outputs[0]
             assert lm_logits.shape[-1] == self.model.config.vocab_size
+
+            #loss_fct = torch.nn.NLLLoss(reduction='none', ignore_index=pad_token_id)
+            #logit_shape =
+            #weights = torch.ones(logit_shape
             loss = loss_fct(lm_logits.view(-1, lm_logits.shape[-1]), lm_labels.view(-1))
+            loss = loss.view(-1, bs)
+            loss = loss.mean(dim=0)
+            mask = self.dropper(loss)
+            loss *= mask
+            loss = loss.mean()
+            return (loss, 1-mask.mean())
+            #loss = loss.view(-1, bs)
         else:
             lprobs = torch.nn.functional.log_softmax(outputs[0], dim=-1)
             loss, nll_loss = label_smoothed_nll_loss(
                 lprobs, lm_labels, self.hparams.label_smoothing, ignore_index=pad_token_id
             )
-        return (loss,)
+        return (loss,torch.tensor(1.))
 
     @property
     def pad(self) -> int:
@@ -302,6 +322,7 @@ class SummarizationModule(BaseTransformer):
             "--task", type=str, default="summarization", required=False, help="# examples. -1 means use all."
         )
         parser.add_argument("--label_smoothing", type=float, default=0.0, required=False)
+        parser.add_argument("--loss_dropper", type=float, default=0.0, required=False)
         parser.add_argument("--src_lang", type=str, default="", required=False)
         parser.add_argument("--tgt_lang", type=str, default="", required=False)
         parser.add_argument(
@@ -316,7 +337,6 @@ class SummarizationModule(BaseTransformer):
 
 class TranslationModule(SummarizationModule):
     mode = "translation"
-    loss_names = ["loss"]
     metric_names = ["bleu"]
     val_metric = "bleu"
 

examples/seq2seq/loss_dropper.py

@@ -0,0 +1,53 @@
+import numpy as np
+import torch.nn as nn
+
+
+class LossDropper(nn.Module):
+    def __init__(
+            self,
+            dropc=0.4,
+            min_count=10000,
+            recompute=10000,
+            verbose=True
+    ):
+        super().__init__()
+        self.keepc = 1. - dropc
+        self.count = 0
+        self.min_count = min_count
+
+        self.recompute = recompute
+        self.last_computed = 0
+        self.percentile_val = 100000000.
+        self.cur_idx = 0
+
+        self.verbose = verbose
+
+        self.vals = np.zeros(self.recompute, dtype=np.float32)
+
+    def forward(self, loss):
+        if loss is None:
+            return loss
+
+        self.last_computed += loss.numel()
+        self.count += loss.numel()
+        if self.count < len(self.vals):
+            self.vals[self.count - loss.numel():self.count] = loss.detach().cpu().numpy().flatten()
+            self.cur_idx += loss.numel()
+            return (loss < np.inf).type(loss.dtype)
+        else:
+            for idx, item in enumerate(loss):
+                self.vals[self.cur_idx] = item
+                self.cur_idx += 1
+                if self.cur_idx >= len(self.vals):
+                    self.cur_idx = 0
+        if self.count < self.min_count:
+            return (loss < np.inf).type(loss.dtype)
+
+        if self.last_computed > self.recompute:
+            self.percentile_val = np.percentile(self.vals, self.keepc * 100)
+            if self.verbose:
+                print('Using cutoff', self.percentile_val)
+            self.last_computed = 0
+
+        mask = (loss < self.percentile_val).type(loss.dtype)
+        return mask

examples/seq2seq/train_distilbart_cnn.sh

@@ -6,7 +6,7 @@ export GAS=1
 
 python finetune.py \
     --learning_rate=3e-5 \
-    --fp16 \
+    --fp16 --fp16_opt_level=O1 \
     --gpus 1 \
     --do_train \
     --do_predict \

examples/seq2seq/train_distilbart_xsum.sh

@@ -6,7 +6,7 @@ python distillation.py \
   --learning_rate=3e-4 \
   --do_train \
   --do_predict \
-  --fp16 \
+  --fp16 --fp16_opt_level=O1 \
   --val_check_interval 0.1 --n_val 1000 \
   --teacher facebook/bart-large-xsum --data_dir $XSUM_DIR \
   --max_target_length=60 --val_max_target_length=60 --test_max_target_length=100 \

examples/seq2seq/train_mbart_cc25_enro.sh

@@ -14,5 +14,4 @@ python finetune.py \
     --task translation \
     --warmup_steps 500 \
     --freeze_embeds \
-    --model_name_or_path=facebook/mbart-large-cc25 \
     "$@"

src/transformers/__init__.py

@@ -299,7 +299,6 @@ if is_torch_available():
         BartModel,
         BartForConditionalGeneration,
         BartForQuestionAnswering,
-        BART_PRETRAINED_MODEL_ARCHIVE_LIST,
     )
     from .modeling_mbart import MBartForConditionalGeneration
     from .modeling_marian import MarianMTModel

src/transformers/modeling_bart.py

@@ -51,15 +51,7 @@ _CONFIG_FOR_DOC = "BartConfig"
 _TOKENIZER_FOR_DOC = "BartTokenizer"
 
 
-BART_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "facebook/bart-base",
-    "facebook/bart-large",
-    "facebook/bart-large-mnli",
-    "facebook/bart-large-cnn",
-    "facebook/bart-large-xsum",
-    "facebook/mbart-large-en-ro",
-    # See all BART models at https://huggingface.co/models?filter=bart
-]
+# See all BART models at https://huggingface.co/models?filter=bart
 
 
 BART_START_DOCSTRING = r"""
@@ -1023,6 +1015,7 @@ class BartForConditionalGeneration(PretrainedBartModel):
 
         if labels is not None:
             use_cache = False
+            decoder_input_ids = shift_tokens_right(labels, self.config.pad_token_id)
 
         outputs = self.model(
             input_ids,

src/transformers/tokenization_bart.py

@@ -33,6 +33,7 @@ _all_bart_models = [
     "facebook/bart-large-cnn",
     "facebook/bart-large-xsum",
     "yjernite/bart_eli5",
+    # This is not exhaustive: see https://huggingface.co/models?filter=bart
 ]
 
 
@@ -133,7 +134,7 @@ class BartTokenizer(RobertaTokenizer):
         # Process tgt_texts
         if max_target_length is None:
             max_target_length = max_length
-        decoder_inputs: BatchEncoding = self(
+        labels = self(
             tgt_texts,
             add_special_tokens=True,
             return_tensors=return_tensors,
@@ -141,10 +142,8 @@ class BartTokenizer(RobertaTokenizer):
             max_length=max_target_length,
             truncation=truncation,
             **kwargs,
-        )
-        for k, v in decoder_inputs.items():
-            model_inputs[f"decoder_{k}"] = v
-
+        )["input_ids"]
+        model_inputs["labels"] = labels
         return model_inputs
 
 
@@ -245,7 +244,7 @@ class BartTokenizerFast(RobertaTokenizerFast):
         # Process tgt_texts
         if max_target_length is None:
             max_target_length = max_length
-        decoder_inputs: BatchEncoding = self(
+        labels = self(
             tgt_texts,
             add_special_tokens=True,
             return_tensors=return_tensors,
@@ -253,8 +252,6 @@ class BartTokenizerFast(RobertaTokenizerFast):
             max_length=max_target_length,
             truncation=truncation,
             **kwargs,
-        )
-        for k, v in decoder_inputs.items():
-            model_inputs[f"decoder_{k}"] = v
-
+        )["input_ids"]
+        model_inputs["labels"] = labels
         return model_inputs

src/transformers/tokenization_mbart.py

@@ -56,6 +56,15 @@ FAIRSEQ_LANGUAGE_CODES = [
 ]
 
 
+def shift_tokens_right(input_ids, pad_token_id):
+    """Shift input ids one token to the right, and wrap the last non pad token (usually <eos>)."""
+    prev_output_tokens = input_ids.clone()
+    index_of_eos = (input_ids.ne(pad_token_id).sum(dim=1) - 1).unsqueeze(-1)
+    prev_output_tokens[:, 0] = input_ids.gather(1, index_of_eos).squeeze()
+    prev_output_tokens[:, 1:] = input_ids[:, :-1]
+    return prev_output_tokens
+
+
 class MBartTokenizer(XLMRobertaTokenizer):
     """
     This inherits from XLMRobertaTokenizer. ``prepare_seq2seq_batch`` should be used to encode inputs.
@@ -98,32 +107,6 @@ class MBartTokenizer(XLMRobertaTokenizer):
         self._additional_special_tokens = list(self.lang_code_to_id.keys())
         self.set_src_lang_special_tokens(kwargs.get("src_lang", "en_XX"))
 
-    def build_inputs_with_special_tokens(
-        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
-    ) -> List[int]:
-        """
-        Build model inputs from a sequence or a pair of sequence for sequence classification tasks
-        by concatenating and adding special tokens. The special tokens depend on calling set_lang.
-        An MBART sequence has the following format, where ``X`` represents the sequence:
-        - ``input_ids`` (for encoder) ``X [eos, src_lang_code]``
-        - ``decoder_input_ids``: (for decoder) ``[tgt_lang_code] X [eos]``
-        BOS is never used.
-        Pairs of sequences are not the expected use case, but they will be handled without a separator.
-
-        Args:
-            token_ids_0 (:obj:`List[int]`):
-                List of IDs to which the special tokens will be added
-            token_ids_1 (:obj:`List[int]`, `optional`, defaults to :obj:`None`):
-                Optional second list of IDs for sequence pairs.
-
-        Returns:
-            :obj:`List[int]`: list of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.
-        """
-        if token_ids_1 is None:
-            return self.prefix_tokens + token_ids_0 + self.suffix_tokens
-        # We don't expect to process pairs, but leave the pair logic for API consistency
-        return self.prefix_tokens + token_ids_0 + token_ids_1 + self.suffix_tokens
-
     def get_special_tokens_mask(
         self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
     ) -> List[int]:
@@ -156,6 +139,32 @@ class MBartTokenizer(XLMRobertaTokenizer):
             return prefix_ones + ([0] * len(token_ids_0)) + suffix_ones
         return prefix_ones + ([0] * len(token_ids_0)) + ([0] * len(token_ids_1)) + suffix_ones
 
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks
+        by concatenating and adding special tokens. The special tokens depend on calling set_lang.
+        An MBART sequence has the following format, where ``X`` represents the sequence:
+        - ``input_ids`` (for encoder) ``X [eos, src_lang_code]``
+        - ``decoder_input_ids``: (for decoder) ``[tgt_lang_code] X [eos]``
+        BOS is never used.
+        Pairs of sequences are not the expected use case, but they will be handled without a separator.
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs to which the special tokens will be added
+            token_ids_1 (:obj:`List[int]`, `optional`, defaults to :obj:`None`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: list of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.
+        """
+        if token_ids_1 is None:
+            return self.prefix_tokens + token_ids_0 + self.suffix_tokens
+        # We don't expect to process pairs, but leave the pair logic for API consistency
+        return self.prefix_tokens + token_ids_0 + token_ids_1 + self.suffix_tokens
+
     @add_start_docstrings_to_callable(PREPARE_SEQ2SEQ_BATCH_DOCSTRING)
     def prepare_seq2seq_batch(
         self,
@@ -251,7 +260,8 @@ class MBartTokenizer(XLMRobertaTokenizer):
         if max_target_length is None:
             max_target_length = max_length
         self.set_tgt_lang_special_tokens(tgt_lang)
-        decoder_inputs: BatchEncoding = self(
+
+        labels = self(
             tgt_texts,
             add_special_tokens=True,
             return_tensors=return_tensors,
@@ -259,10 +269,9 @@ class MBartTokenizer(XLMRobertaTokenizer):
             max_length=max_target_length,
             truncation=True,
             **kwargs,
-        )
-        for k, v in decoder_inputs.items():
-            model_inputs[f"decoder_{k}"] = v
-
+        )["input_ids"]
+        model_inputs["decoder_input_ids"] = shift_tokens_right(labels, self.pad_token_id)
+        model_inputs["labels"] = labels
         self.set_src_lang_special_tokens(src_lang)  # sets to src_lang
         return model_inputs
 
@@ -275,5 +284,5 @@ class MBartTokenizer(XLMRobertaTokenizer):
     def set_tgt_lang_special_tokens(self, lang: str) -> None:
         """Reset the special tokens to the target language setting. Prefix [tgt_lang_code], suffix =[eos]."""
         self.cur_lang_code = self.lang_code_to_id[lang]
-        self.prefix_tokens = [self.cur_lang_code]
-        self.suffix_tokens = [self.eos_token_id]
+        self.prefix_tokens = []
+        self.suffix_tokens = [self.eos_token_id, self.cur_lang_code]

src/transformers/tokenization_pegasus.py

@@ -133,7 +133,8 @@ class PegasusTokenizer(ReformerTokenizer):
             return model_inputs
         if max_target_length is not None:
             tokenizer_kwargs["max_length"] = max_target_length
-        decoder_inputs: BatchEncoding = self(tgt_texts, **tokenizer_kwargs)
-        for k, v in decoder_inputs.items():
-            model_inputs[f"decoder_{k}"] = v
+        labels: BatchEncoding = self(tgt_texts, **tokenizer_kwargs)["input_ids"]
+        model_inputs["labels"] = labels
+        # for k, v in decoder_inputs.items():
+        #    model_inputs[f"decoder_{k}"] = v
         return model_inputs

src/transformers/tokenization_t5.py

@@ -303,7 +303,7 @@ class T5Tokenizer(PreTrainedTokenizer):
         if max_length is None:
             max_length = self.max_len
         self.prefix_tokens = []
-        model_inputs: BatchEncoding = self(
+        model_inputs = self(
             src_texts,
             add_special_tokens=True,
             return_tensors=return_tensors,
@@ -319,7 +319,7 @@ class T5Tokenizer(PreTrainedTokenizer):
             max_target_length = max_length
         # set prefix_tokens for target text
         self.prefix_tokens = [self.pad_token_id]
-        decoder_inputs: BatchEncoding = self(
+        model_inputs["labels"] = self(
             tgt_texts,
             add_special_tokens=True,
             return_tensors=return_tensors,
@@ -327,9 +327,7 @@ class T5Tokenizer(PreTrainedTokenizer):
             max_length=max_target_length,
             truncation=truncation,
             **kwargs,
-        )
-        for k, v in decoder_inputs.items():
-            model_inputs[f"decoder_{k}"] = v
+        )["input_ids"]
 
         self.prefix_tokens = []
         return model_inputs

tests/test_modeling_bart.py

@@ -18,7 +18,7 @@ import unittest
 
 import timeout_decorator  # noqa
 
-from transformers import BatchEncoding, is_torch_available
+from transformers import is_torch_available
 from transformers.file_utils import cached_property
 from transformers.testing_utils import require_torch, slow, torch_device
 
@@ -486,7 +486,7 @@ class BartModelIntegrationTests(unittest.TestCase):
     def test_xsum_summarization_same_as_fairseq(self):
         model = BartForConditionalGeneration.from_pretrained("facebook/bart-large-xsum").to(torch_device)
         self.assertFalse(model.config.is_valid_mbart())
-        tok = BartTokenizer.from_pretrained("facebook/bart-large")
+        tok = self.default_tokenizer
 
         EXPECTED_SUMMARY = "California's largest power company has begun shutting off electricity to thousands of customers in the state."
         dct = tok.batch_encode_plus(
@@ -568,84 +568,6 @@ class BartModelIntegrationTests(unittest.TestCase):
         # TODO(SS): run fairseq again with num_beams=2, min_len=20.
         # TODO(SS): add test case that hits max_length
 
-    def test_prepare_seq2seq_batch(self):
-        tokenizers = [self.default_tokenizer, self.default_tokenizer_fast]
-        src_text = ["A long paragraph for summrization.", "Another paragraph for summrization."]
-        tgt_text = [
-            "Summary of the text.",
-            "Another summary.",
-        ]
-        expected_src_tokens = [0, 250, 251, 17818, 13, 32933, 21645, 1258, 4, 2]
-
-        for tokenizer in tokenizers:
-            batch = tokenizer.prepare_seq2seq_batch(
-                src_text, tgt_texts=tgt_text, max_length=len(expected_src_tokens), return_tensors="pt"
-            )
-            self.assertIsInstance(batch, BatchEncoding)
-
-            self.assertEqual((2, 10), batch.input_ids.shape)
-            self.assertEqual((2, 10), batch.attention_mask.shape)
-            result = batch.input_ids.tolist()[0]
-            self.assertListEqual(expected_src_tokens, result)
-            # Test that special tokens are reset
-
-    def test_empty_target_text(self):
-        tokenizers = [self.default_tokenizer, self.default_tokenizer_fast]
-        src_text = ["A long paragraph for summrization.", "Another paragraph for summrization."]
-        for tokenizer in tokenizers:
-            batch = tokenizer.prepare_seq2seq_batch(src_text, return_tensors="pt")
-            # check if input_ids are returned and no decoder_input_ids
-            self.assertIn("input_ids", batch)
-            self.assertIn("attention_mask", batch)
-            self.assertNotIn("decoder_input_ids", batch)
-            self.assertNotIn("decoder_attention_mask", batch)
-
-    def test_max_target_length(self):
-        tokenizers = [self.default_tokenizer, self.default_tokenizer_fast]
-        src_text = ["A long paragraph for summrization.", "Another paragraph for summrization."]
-        tgt_text = [
-            "Summary of the text.",
-            "Another summary.",
-        ]
-        for tokenizer in tokenizers:
-            batch = tokenizer.prepare_seq2seq_batch(
-                src_text, tgt_texts=tgt_text, max_target_length=32, padding="max_length", return_tensors="pt"
-            )
-            self.assertEqual(32, batch["decoder_input_ids"].shape[1])
-            self.assertEqual(32, batch["decoder_attention_mask"].shape[1])
-
-            # test None max_target_length
-            batch = tokenizer.prepare_seq2seq_batch(
-                src_text, tgt_texts=tgt_text, max_length=32, padding="max_length", return_tensors="pt"
-            )
-            self.assertEqual(32, batch["decoder_input_ids"].shape[1])
-            self.assertEqual(32, batch["decoder_attention_mask"].shape[1])
-
-    def test_outputs_not_longer_than_maxlen(self):
-        tokenizers = [self.default_tokenizer, self.default_tokenizer_fast]
-
-        for tokenizer in tokenizers:
-            batch = tokenizer.prepare_seq2seq_batch(
-                ["I am a small frog" * 1024, "I am a small frog"], return_tensors="pt"
-            )
-            self.assertIsInstance(batch, BatchEncoding)
-            self.assertEqual(batch.input_ids.shape, (2, 1024))
-
-    def test_special_tokens(self):
-        tokenizers = [self.default_tokenizer, self.default_tokenizer_fast]
-        src_text = ["A long paragraph for summrization."]
-        tgt_text = [
-            "Summary of the text.",
-        ]
-        for tokenizer in tokenizers:
-            batch = tokenizer.prepare_seq2seq_batch(src_text, tgt_texts=tgt_text, return_tensors="pt")
-            input_ids = batch["input_ids"]
-            decoder_input_ids = batch["decoder_input_ids"]
-            self.assertTrue((input_ids[:, 0] == tokenizer.bos_token_id).all().item())
-            self.assertTrue((decoder_input_ids[:, 0] == tokenizer.bos_token_id).all().item())
-            self.assertTrue((input_ids[:, -1] == tokenizer.eos_token_id).all().item())
-            self.assertTrue((decoder_input_ids[:, -1] == tokenizer.eos_token_id).all().item())
-
 
 @require_torch
 class TestSinusoidalPositionalEmbeddings(unittest.TestCase):

tests/test_modeling_pegasus.py

@@ -31,9 +31,7 @@ class PegasusXSUMIntegrationTest(AbstractSeq2SeqIntegrationTest):
     @slow
     def test_pegasus_xsum_summary(self):
         assert self.tokenizer.model_max_length == 512
-        inputs = self.tokenizer(self.src_text, return_tensors="pt", truncation=True, max_length=512, padding=True).to(
-            torch_device
-        )
+        inputs = self.tokenizer(self.src_text, return_tensors="pt", truncation=True, padding=True).to(torch_device)
         assert inputs.input_ids.shape == (2, 421)
         translated_tokens = self.model.generate(**inputs)
         decoded = self.tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)

tests/test_tokenization_bart.py

@@ -0,0 +1,90 @@
+import unittest
+
+from transformers import BartTokenizer, BartTokenizerFast, BatchEncoding
+from transformers.file_utils import cached_property
+
+
+class TestTokenizationBart(unittest.TestCase):
+    @cached_property
+    def default_tokenizer(self):
+        return BartTokenizer.from_pretrained("facebook/bart-large")
+
+    @cached_property
+    def default_tokenizer_fast(self):
+        return BartTokenizerFast.from_pretrained("facebook/bart-large")
+
+    def test_prepare_seq2seq_batch(self):
+        tokenizers = [self.default_tokenizer, self.default_tokenizer_fast]
+        src_text = ["A long paragraph for summrization.", "Another paragraph for summrization."]
+        tgt_text = [
+            "Summary of the text.",
+            "Another summary.",
+        ]
+        expected_src_tokens = [0, 250, 251, 17818, 13, 32933, 21645, 1258, 4, 2]
+
+        for tokenizer in tokenizers:
+            batch = tokenizer.prepare_seq2seq_batch(
+                src_text, tgt_texts=tgt_text, max_length=len(expected_src_tokens), return_tensors="pt"
+            )
+            self.assertIsInstance(batch, BatchEncoding)
+
+            self.assertEqual((2, 10), batch.input_ids.shape)
+            self.assertEqual((2, 10), batch.attention_mask.shape)
+            result = batch.input_ids.tolist()[0]
+            self.assertListEqual(expected_src_tokens, result)
+            # Test that special tokens are reset
+
+    def test_empty_target_text(self):
+        tokenizers = [self.default_tokenizer, self.default_tokenizer_fast]
+        src_text = ["A long paragraph for summrization.", "Another paragraph for summrization."]
+        for tokenizer in tokenizers:
+            batch = tokenizer.prepare_seq2seq_batch(src_text, return_tensors="pt")
+            # check if input_ids are returned and no labels
+            self.assertIn("input_ids", batch)
+            self.assertIn("attention_mask", batch)
+            self.assertNotIn("labels", batch)
+            self.assertNotIn("decoder_attention_mask", batch)
+
+    def test_max_target_length(self):
+        tokenizers = [self.default_tokenizer, self.default_tokenizer_fast]
+        src_text = ["A long paragraph for summrization.", "Another paragraph for summrization."]
+        tgt_text = [
+            "Summary of the text.",
+            "Another summary.",
+        ]
+        for tokenizer in tokenizers:
+            batch = tokenizer.prepare_seq2seq_batch(
+                src_text, tgt_texts=tgt_text, max_target_length=32, padding="max_length", return_tensors="pt"
+            )
+            self.assertEqual(32, batch["labels"].shape[1])
+
+            # test None max_target_length
+            batch = tokenizer.prepare_seq2seq_batch(
+                src_text, tgt_texts=tgt_text, max_length=32, padding="max_length", return_tensors="pt"
+            )
+            self.assertEqual(32, batch["labels"].shape[1])
+
+    def test_outputs_not_longer_than_maxlen(self):
+        tokenizers = [self.default_tokenizer, self.default_tokenizer_fast]
+
+        for tokenizer in tokenizers:
+            batch = tokenizer.prepare_seq2seq_batch(
+                ["I am a small frog" * 1024, "I am a small frog"], return_tensors="pt"
+            )
+            self.assertIsInstance(batch, BatchEncoding)
+            self.assertEqual(batch.input_ids.shape, (2, 1024))
+
+    def test_special_tokens(self):
+        tokenizers = [self.default_tokenizer, self.default_tokenizer_fast]
+        src_text = ["A long paragraph for summrization."]
+        tgt_text = [
+            "Summary of the text.",
+        ]
+        for tokenizer in tokenizers:
+            batch = tokenizer.prepare_seq2seq_batch(src_text, tgt_texts=tgt_text, return_tensors="pt")
+            input_ids = batch["input_ids"]
+            labels = batch["labels"]
+            self.assertTrue((input_ids[:, 0] == tokenizer.bos_token_id).all().item())
+            self.assertTrue((labels[:, 0] == tokenizer.bos_token_id).all().item())
+            self.assertTrue((input_ids[:, -1] == tokenizer.eos_token_id).all().item())
+            self.assertTrue((labels[:, -1] == tokenizer.eos_token_id).all().item())

tests/test_tokenization_common.py

@@ -1544,11 +1544,11 @@ class TokenizerTesterMixin:
             src_texts=src_text, tgt_texts=tgt_text, max_length=3, max_target_length=10, return_tensors="pt"
         )
         self.assertEqual(batch.input_ids.shape[1], 3)
-        self.assertEqual(batch.decoder_input_ids.shape[1], 10)
+        self.assertEqual(batch.labels.shape[1], 10)
         # max_target_length will default to max_length if not specified
         batch = tokenizer.prepare_seq2seq_batch(src_text, tgt_texts=tgt_text, max_length=3)
         self.assertEqual(batch.input_ids.shape[1], 3)
-        self.assertEqual(batch.decoder_input_ids.shape[1], 3)
+        self.assertEqual(batch.labels.shape[1], 3)
 
         batch_encoder_only = tokenizer.prepare_seq2seq_batch(
             src_texts=src_text, max_length=3, max_target_length=10, return_tensors="pt"

tests/test_tokenization_mbart.py

@@ -182,3 +182,17 @@ class MBartEnroIntegrationTest(unittest.TestCase):
         self.tokenizer.save_pretrained(tmpdirname)
         new_tok = MBartTokenizer.from_pretrained(tmpdirname)
         self.assertDictEqual(new_tok.fairseq_tokens_to_ids, original_special_tokens)
+
+    def test_batch_fairseq_parity(self):
+        batch: BatchEncoding = self.tokenizer.prepare_seq2seq_batch(
+            self.src_text, tgt_texts=self.tgt_text, return_tensors="pt"
+        )
+        for k in batch:
+            batch[k] = batch[k].tolist()
+        # batch = {k: v.tolist() for k,v in batch.items()}
+        # fairseq batch: https://gist.github.com/sshleifer/cba08bc2109361a74ac3760a7e30e4f4
+        # batch.decoder_inputs_ids[0][0] ==
+        assert batch.input_ids[1][-2:] == [2, EN_CODE]
+        assert batch.decoder_input_ids[1][0] == RO_CODE
+        assert batch.decoder_input_ids[1][-1] == 2
+        assert batch.labels[1][-2:] == [2, RO_CODE]

tests/test_tokenization_pegasus.py

@@ -63,7 +63,6 @@ class PegasusTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
         batch = self.pegasus_large_tokenizer.prepare_seq2seq_batch(src_texts, tgt_texts=tgt_texts, max_target_length=5)
         assert batch.input_ids.shape == (2, 1024)
         assert batch.attention_mask.shape == (2, 1024)
-        assert "decoder_input_ids" in batch  # because tgt_texts was specified
-        assert batch.decoder_input_ids.shape == (2, 5)
-        assert batch.decoder_attention_mask.shape == (2, 5)
-        assert len(batch) == 4  # no extra keys
+        assert "labels" in batch  # because tgt_texts was specified
+        assert batch.labels.shape == (2, 5)
+        assert len(batch) == 3  # input_ids, attention_mask, labels. Other things make by BartModel