Adding new `encoder_no_repeat_ngram_size` to `generate`. (#9984)

Adding new `encoder_no_repeat_ngram_size` to `generate`.

Blenderbot results seemed off compared to original ParlAI script:
`https://parl.ai/projects/recipes/`

. Notably the model seems
to repeat a lot what was said during the conversation.

The actual problem was that `no_repeat_ngram_size` actually applies
to the `encoder_input_ids` but HF's `no_repeat_ngram_size` applies
to the previously generated ids (within the decoder). The history
conversation of blenderbot is within the `encoder` part so that
explains why HF's implementation had the repetitions.

This fix was focused on blenderbot *not* small and added tests
for those because they are quite different in configuration.

This change includes:

- Adding a new EncoderNoRepeatLogitProcessor.
- Adding 1 new arg to `generate` (`encoder_no_repeat_ngram_size`)
- Adding 1 new config parameter `encoder_no_repeat_ngram_size`.
- Adding 2 tests, one for the pipeline (high level, inputs exhibited
repeat behavior, one low level for EncoderNoRepeatLogitProcessor)
- Factored NoRepeatLogitProcessor so that logic could be reused.

Further work:

- Blenderbot conversational pipeline still does not behave correctly
 as they way input is prepared within the pipeline is still incorrect
(follow up PR)
- Blenderbot allows the bot to have personas, which is done by
prepending "your personna: XXXX" to the input, this could be explored
too in a follow up PR.

@patrickvonplaten
@LysandreJik

* Update src/transformers/generation_logits_process.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/generation_utils.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/generation_utils.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/configuration_utils.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Doc quality.

* Fixing test.

* Last fixes.

* Fixing to account for batch_size.

* Update src/transformers/configuration_utils.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/generation_utils.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

src/transformers/configuration_utils.py

@@ -117,6 +117,9 @@ class PretrainedConfig(object):
         - **no_repeat_ngram_size** (:obj:`int`, `optional`, defaults to 0) -- Value that will be used by default in the
           :obj:`generate` method of the model for ``no_repeat_ngram_size``. If set to int > 0, all ngrams of that size
           can only occur once.
+        - **encoder_no_repeat_ngram_size** (:obj:`int`, `optional`, defaults to 0) -- Value that will be used by
+          default in the :obj:`generate` method of the model for ``encoder_no_repeat_ngram_size``. If set to int > 0,
+          all ngrams of that size that occur in the ``encoder_input_ids`` cannot occur in the ``decoder_input_ids``.
         - **bad_words_ids** (:obj:`List[int]`, `optional`) -- List of token ids that are not allowed to be generated
           that will be used by default in the :obj:`generate` method of the model. In order to get the tokens of the
           words that should not appear in the generated text, use :obj:`tokenizer.encode(bad_word,
@@ -205,6 +208,7 @@ class PretrainedConfig(object):
         self.repetition_penalty = kwargs.pop("repetition_penalty", 1.0)
         self.length_penalty = kwargs.pop("length_penalty", 1.0)
         self.no_repeat_ngram_size = kwargs.pop("no_repeat_ngram_size", 0)
+        self.encoder_no_repeat_ngram_size = kwargs.pop("encoder_no_repeat_ngram_size", 0)
         self.bad_words_ids = kwargs.pop("bad_words_ids", None)
         self.num_return_sequences = kwargs.pop("num_return_sequences", 1)
         self.chunk_size_feed_forward = kwargs.pop("chunk_size_feed_forward", 0)

src/transformers/generation_logits_process.py

@@ -235,6 +235,41 @@ class TopKLogitsWarper(LogitsWarper):
         return scores
 
 
+def _get_ngrams(ngram_size: int, prev_input_ids: torch.Tensor, num_hypos: int):
+    generated_ngrams = [{} for _ in range(num_hypos)]
+    for idx in range(num_hypos):
+        gen_tokens = prev_input_ids[idx].tolist()
+        generated_ngram = generated_ngrams[idx]
+        for ngram in zip(*[gen_tokens[i:] for i in range(ngram_size)]):
+            prev_ngram_tuple = tuple(ngram[:-1])
+            generated_ngram[prev_ngram_tuple] = generated_ngram.get(prev_ngram_tuple, []) + [ngram[-1]]
+    return generated_ngrams
+
+
+def _get_generated_ngrams(banned_ngrams, prev_input_ids, ngram_size, cur_len):
+    # Before decoding the next token, prevent decoding of ngrams that have already appeared
+    start_idx = cur_len + 1 - ngram_size
+    ngram_idx = tuple(prev_input_ids[start_idx:cur_len].tolist())
+    return banned_ngrams.get(ngram_idx, [])
+
+
+def _calc_banned_ngram_tokens(
+    ngram_size: int, prev_input_ids: torch.Tensor, num_hypos: int, cur_len: int
+) -> List[Iterable[int]]:
+    """Copied from fairseq for no_repeat_ngram in beam_search"""
+    if cur_len + 1 < ngram_size:
+        # return no banned tokens if we haven't generated no_repeat_ngram_size tokens yet
+        return [[] for _ in range(num_hypos)]
+
+    generated_ngrams = _get_ngrams(ngram_size, prev_input_ids, num_hypos)
+
+    banned_tokens = [
+        _get_generated_ngrams(generated_ngrams[hypo_idx], prev_input_ids[hypo_idx], ngram_size, cur_len)
+        for hypo_idx in range(num_hypos)
+    ]
+    return banned_tokens
+
+
 class NoRepeatNGramLogitsProcessor(LogitsProcessor):
     r"""
     :class:`transformers.LogitsProcessor` that enforces no repetition of n-grams. See `Fairseq
@@ -253,36 +288,53 @@ class NoRepeatNGramLogitsProcessor(LogitsProcessor):
     def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
         num_batch_hypotheses = scores.shape[0]
         cur_len = input_ids.shape[-1]
-        banned_batch_tokens = self._calc_banned_ngram_tokens(input_ids, num_batch_hypotheses, cur_len)
+        banned_batch_tokens = _calc_banned_ngram_tokens(self.ngram_size, input_ids, num_batch_hypotheses, cur_len)
 
         for i, banned_tokens in enumerate(banned_batch_tokens):
             scores[i, banned_tokens] = -float("inf")
 
         return scores
 
-    def _calc_banned_ngram_tokens(
-        self, prev_input_ids: torch.Tensor, num_hypos: int, cur_len: int
-    ) -> List[Iterable[int]]:
-        """Copied from fairseq for no_repeat_ngram in beam_search"""
-        if cur_len + 1 < self.ngram_size:
-            # return no banned tokens if we haven't generated no_repeat_ngram_size tokens yet
-            return [[] for _ in range(num_hypos)]
-        generated_ngrams = [{} for _ in range(num_hypos)]
-        for idx in range(num_hypos):
-            gen_tokens = prev_input_ids[idx].tolist()
-            generated_ngram = generated_ngrams[idx]
-            for ngram in zip(*[gen_tokens[i:] for i in range(self.ngram_size)]):
-                prev_ngram_tuple = tuple(ngram[:-1])
-                generated_ngram[prev_ngram_tuple] = generated_ngram.get(prev_ngram_tuple, []) + [ngram[-1]]
-
-        def _get_generated_ngrams(hypo_idx):
-            # Before decoding the next token, prevent decoding of ngrams that have already appeared
-            start_idx = cur_len + 1 - self.ngram_size
-            ngram_idx = tuple(prev_input_ids[hypo_idx, start_idx:cur_len].tolist())
-            return generated_ngrams[hypo_idx].get(ngram_idx, [])
-
-        banned_tokens = [_get_generated_ngrams(hypo_idx) for hypo_idx in range(num_hypos)]
-        return banned_tokens
+
+class EncoderNoRepeatNGramLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` that enforces no repetition of encoder input ids n-grams for the decoder ids.
+    See `ParlAI <https://github.com/facebookresearch/ParlAI/blob/master/parlai/core/torch_generator_agent.py#L1350>`__.
+
+    Args:
+        encoder_ngram_size (:obj:`int`):
+            All ngrams of size :obj:`ngram_size` can only occur within the encoder input ids.
+        encoder_input_ids (:obj:`int`):
+            The encoder_input_ids that should not be repeated within the decoder ids.
+    """
+
+    def __init__(self, encoder_ngram_size: int, encoder_input_ids: torch.LongTensor):
+        if not isinstance(encoder_ngram_size, int) or encoder_ngram_size <= 0:
+            raise ValueError(
+                f"`encoder_ngram_size` has to be a strictly positive integer, but is {encoder_ngram_size}"
+            )
+        self.ngram_size = encoder_ngram_size
+        if len(encoder_input_ids.shape) == 1:
+            encoder_input_ids = encoder_input_ids.unsqueeze(0)
+        self.batch_size = encoder_input_ids.shape[0]
+        self.generated_ngrams = _get_ngrams(encoder_ngram_size, encoder_input_ids, self.batch_size)
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        # B x num_beams
+        num_hypos = scores.shape[0]
+        num_beams = num_hypos // self.batch_size
+        cur_len = input_ids.shape[-1]
+        banned_batch_tokens = [
+            _get_generated_ngrams(
+                self.generated_ngrams[hypo_idx // num_beams], input_ids[hypo_idx], self.ngram_size, cur_len
+            )
+            for hypo_idx in range(num_hypos)
+        ]
+
+        for i, banned_tokens in enumerate(banned_batch_tokens):
+            scores[i, banned_tokens] = -float("inf")
+
+        return scores
 
 
 class NoBadWordsLogitsProcessor(LogitsProcessor):

src/transformers/generation_utils.py

@@ -23,6 +23,7 @@ from torch.nn import functional as F
 from .file_utils import ModelOutput
 from .generation_beam_search import BeamScorer, BeamSearchScorer
 from .generation_logits_process import (
+    EncoderNoRepeatNGramLogitsProcessor,
     HammingDiversityLogitsProcessor,
     LogitsProcessorList,
     MinLengthLogitsProcessor,
@@ -537,6 +538,8 @@ class GenerationMixin:
         self,
         repetition_penalty: float,
         no_repeat_ngram_size: int,
+        encoder_no_repeat_ngram_size: int,
+        encoder_input_ids: torch.LongTensor,
         bad_words_ids: List[List[int]],
         min_length: int,
         eos_token_id: int,
@@ -555,6 +558,11 @@ class GenerationMixin:
         no_repeat_ngram_size = (
             no_repeat_ngram_size if no_repeat_ngram_size is not None else self.config.no_repeat_ngram_size
         )
+        encoder_no_repeat_ngram_size = (
+            encoder_no_repeat_ngram_size
+            if encoder_no_repeat_ngram_size is not None
+            else self.config.encoder_no_repeat_ngram_size
+        )
         bad_words_ids = bad_words_ids if bad_words_ids is not None else self.config.bad_words_ids
         min_length = min_length if min_length is not None else self.config.min_length
         eos_token_id = eos_token_id if eos_token_id is not None else self.config.eos_token_id
@@ -574,6 +582,13 @@ class GenerationMixin:
             processors.append(RepetitionPenaltyLogitsProcessor(penalty=repetition_penalty))
         if no_repeat_ngram_size is not None and no_repeat_ngram_size > 0:
             processors.append(NoRepeatNGramLogitsProcessor(no_repeat_ngram_size))
+        if encoder_no_repeat_ngram_size is not None and encoder_no_repeat_ngram_size > 0:
+            if self.config.is_encoder_decoder:
+                processors.append(EncoderNoRepeatNGramLogitsProcessor(encoder_no_repeat_ngram_size, encoder_input_ids))
+            else:
+                raise ValueError(
+                    "It's impossible to use `encoder_no_repeat_ngram_size` with decoder-only architecture"
+                )
         if bad_words_ids is not None:
             processors.append(NoBadWordsLogitsProcessor(bad_words_ids, eos_token_id))
         if min_length is not None and eos_token_id is not None and min_length > -1:
@@ -601,6 +616,7 @@ class GenerationMixin:
         eos_token_id: Optional[int] = None,
         length_penalty: Optional[float] = None,
         no_repeat_ngram_size: Optional[int] = None,
+        encoder_no_repeat_ngram_size: Optional[int] = None,
         num_return_sequences: Optional[int] = None,
         decoder_start_token_id: Optional[int] = None,
         use_cache: Optional[bool] = None,
@@ -661,6 +677,9 @@ class GenerationMixin:
                 sequences.
             no_repeat_ngram_size (:obj:`int`, `optional`, defaults to 0):
                 If set to int > 0, all ngrams of that size can only occur once.
+            encoder_no_repeat_ngram_size (:obj:`int`, `optional`, defaults to 0):
+                If set to int > 0, all ngrams of that size that occur in the ``encoder_input_ids`` cannot occur in the
+                ``decoder_input_ids``.
             bad_words_ids(:obj:`List[List[int]]`, `optional`):
                 List of token ids that are not allowed to be generated. In order to get the tokens of the words that
                 should not appear in the generated text, use :obj:`tokenizer(bad_word,
@@ -820,6 +839,9 @@ class GenerationMixin:
             logger.warning(f"Setting `pad_token_id` to `eos_token_id`:{eos_token_id} for open-end generation.")
             pad_token_id = eos_token_id
 
+        # Storing encoder_input_ids for logits_processor that could use them
+        encoder_input_ids = input_ids if self.config.is_encoder_decoder else None
+
         if self.config.is_encoder_decoder:
             # add encoder_outputs to model_kwargs
             model_kwargs = self._prepare_encoder_decoder_kwargs_for_generation(input_ids, model_kwargs)
@@ -862,6 +884,8 @@ class GenerationMixin:
         logits_processor = self._get_logits_processor(
             repetition_penalty=repetition_penalty,
             no_repeat_ngram_size=no_repeat_ngram_size,
+            encoder_no_repeat_ngram_size=encoder_no_repeat_ngram_size,
+            encoder_input_ids=encoder_input_ids,
             bad_words_ids=bad_words_ids,
             min_length=min_length,
             eos_token_id=eos_token_id,
@@ -1638,6 +1662,7 @@ class GenerationMixin:
             beam_idx = beam_outputs["next_beam_indices"]
 
             input_ids = torch.cat([input_ids[beam_idx, :], beam_next_tokens.unsqueeze(-1)], dim=-1)
+
             cur_len = cur_len + 1
 
             model_kwargs = self._update_model_kwargs_for_generation(

src/transformers/models/blenderbot/configuration_blenderbot.py

@@ -128,6 +128,7 @@ class BlenderbotConfig(PretrainedConfig):
         pad_token_id=0,
         bos_token_id=1,
         eos_token_id=2,
+        encoder_no_repeat_ngram_size=3,
         **kwargs
     ):
         super().__init__(
@@ -136,6 +137,7 @@ class BlenderbotConfig(PretrainedConfig):
             eos_token_id=eos_token_id,
             is_encoder_decoder=is_encoder_decoder,
             decoder_start_token_id=decoder_start_token_id,
+            encoder_no_repeat_ngram_size=encoder_no_repeat_ngram_size,
             **kwargs,
         )
 

tests/test_generation_logits_process.py

@@ -27,6 +27,7 @@ if is_torch_available():
     import torch.nn.functional as F
 
     from transformers.generation_logits_process import (
+        EncoderNoRepeatNGramLogitsProcessor,
         HammingDiversityLogitsProcessor,
         LogitsProcessorList,
         MinLengthLogitsProcessor,
@@ -208,6 +209,68 @@ class LogitsProcessorTest(unittest.TestCase):
             torch.isinf(filtered_scores_3_gram).tolist(), [[False, False, False], [True, False, False]]
         )
 
+    def test_encoder_no_repeat_ngram_dist_processor(self):
+        vocab_size = 3
+        num_beams = 2
+        batch_size = 1
+
+        encoder_input_ids = torch.tensor([1, 2, 1, 1], device=torch_device, dtype=torch.long)
+
+        input_ids = torch.tensor([[1, 2, 1], [8, 0, 2]], device=torch_device, dtype=torch.long)
+        scores = self._get_uniform_logits(batch_size * num_beams, vocab_size)
+
+        no_repeat_proc_2_gram = EncoderNoRepeatNGramLogitsProcessor(2, encoder_input_ids=encoder_input_ids)
+        no_repeat_proc_3_gram = EncoderNoRepeatNGramLogitsProcessor(3, encoder_input_ids=encoder_input_ids)
+
+        filtered_scores_2_gram = no_repeat_proc_2_gram(input_ids, scores.clone())
+        filtered_scores_3_gram = no_repeat_proc_3_gram(input_ids, scores.clone())
+
+        # 2-gram would forbid 1st and 2nd token at 1st beam and 1st token (0) at 2nd beam
+        self.assertListEqual(torch.isinf(filtered_scores_2_gram).tolist(), [[False, True, True], [False, True, False]])
+
+        # 3-gram would forbid 1st token at 1st beam and no token at 2nd beam
+        self.assertListEqual(
+            torch.isinf(filtered_scores_3_gram).tolist(), [[False, True, False], [False, False, False]]
+        )
+
+        # Batched input
+        vocab_size = 3
+        num_beams = 2
+        batch_size = 2
+        encoder_input_ids = torch.tensor([[1, 2, 1, 1], [0, 0, 2, 1]], device=torch_device, dtype=torch.long)
+
+        input_ids = torch.tensor([[1, 2, 1], [1, 0, 2], [0, 0, 0], [0, 2, 2]], device=torch_device, dtype=torch.long)
+        scores = self._get_uniform_logits(batch_size * num_beams, vocab_size)
+
+        no_repeat_proc_2_gram = EncoderNoRepeatNGramLogitsProcessor(2, encoder_input_ids=encoder_input_ids)
+        no_repeat_proc_3_gram = EncoderNoRepeatNGramLogitsProcessor(3, encoder_input_ids=encoder_input_ids)
+
+        filtered_scores_2_gram = no_repeat_proc_2_gram(input_ids, scores.clone())
+        filtered_scores_3_gram = no_repeat_proc_3_gram(input_ids, scores.clone())
+
+        # 2gram
+        # Batch 1
+        #   - Beam 1: tokens (1, 2) forbidden
+        #   - Beam 2: tokens (1) forbidden
+        # Batch 2
+        #   - Beam 1: tokens (0, 2) forbidden
+        #   - Beam 2: tokens (1) forbidden
+        self.assertListEqual(
+            torch.isinf(filtered_scores_2_gram).tolist(),
+            [[False, True, True], [False, True, False], [True, False, True], [False, True, False]],
+        )
+
+        # Batch 1
+        #   - Beam 1: tokens (1) forbidden
+        #   - Beam 2: tokens () forbidden
+        # Batch 2
+        #   - Beam 1: tokens (2) forbidden
+        #   - Beam 2: tokens () forbidden
+        self.assertListEqual(
+            torch.isinf(filtered_scores_3_gram).tolist(),
+            [[False, True, False], [False, False, False], [False, False, True], [False, False, False]],
+        )
+
     def test_no_bad_words_dist_processor(self):
         vocab_size = 5
         batch_size = 2

tests/test_pipelines_conversational.py

@@ -276,6 +276,47 @@ class ConversationalPipelineTests(MonoInputPipelineCommonMixin, unittest.TestCas
         self.assertEqual(result.past_user_inputs[1], "Is it an action movie?")
         self.assertEqual(result.generated_responses[1], "It's a comedy.")
 
+    @require_torch
+    @slow
+    def test_integration_torch_conversation_blenderbot_400M(self):
+        tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
+        model = AutoModelForSeq2SeqLM.from_pretrained("facebook/blenderbot-400M-distill")
+        nlp = ConversationalPipeline(model=model, tokenizer=tokenizer)
+
+        conversation_1 = Conversation("hello")
+        result = nlp(
+            conversation_1,
+        )
+        self.assertEqual(
+            result.generated_responses[0],
+            # ParlAI implementation output, we have a different one, but it's our
+            # second best, you can check by using num_return_sequences=10
+            # " Hello! How are you? I'm just getting ready to go to work, how about you?",
+            " Hello! How are you doing today? I just got back from a walk with my dog.",
+        )
+
+        conversation_1 = Conversation(" Lasagne   hello")
+        result = nlp(conversation_1, encoder_no_repeat_ngram_size=3)
+        self.assertEqual(
+            result.generated_responses[0],
+            " Lasagne is my favorite Italian dish. Do you like lasagne?",
+        )
+
+        conversation_1 = Conversation(
+            "Lasagne   hello   Lasagne is my favorite Italian dish. Do you like lasagne?   I like lasagne."
+        )
+        result = nlp(
+            conversation_1,
+            encoder_no_repeat_ngram_size=3,
+        )
+        self.assertEqual(
+            result.generated_responses[0],
+            # ParlAI implementation output, we have a different one, but it's our
+            # second best, you can check by using num_return_sequences=10
+            # " Hello! How are you? I'm just getting ready to go to work, how about you?",
+            " Lasagne is a traditional Italian dish consisting of a yeasted flatbread typically topped with tomato sauce and cheese.",
+        )
+
     @require_torch
     @slow
     def test_integration_torch_conversation_encoder_decoder(self):