Add Epsilon- and Eta-Sampling (#21121)

* Add epsilon- and eta-sampling.

Add epsilon- and eta-sampling, following the official code from https://github.com/john-hewitt/truncation-sampling and adapting to be more configurable, as required by Huggingface transformers.

* Add unit tests for epsilon- and eta-sampling.

* Black: fix code formatting.

* Fix docstring spacing.

* Clean up newlines.

* Fix implementation bugs and their associated tests.

* Remove epsilon- and eta-sampling parameters from PretrainedConfig.

* Clarify and clean up the documentation.

* Remove parameters for PretrainedConfig test.

src/transformers/generation/__init__.py

@@ -43,6 +43,8 @@ else:
         "ConstrainedBeamSearchScorer",
     ]
     _import_structure["logits_process"] = [
+        "EpsilonLogitsWarper",
+        "EtaLogitsWarper",
         "ForcedBOSTokenLogitsProcessor",
         "ForcedEOSTokenLogitsProcessor",
         "HammingDiversityLogitsProcessor",
@@ -162,6 +164,8 @@ if TYPE_CHECKING:
         from .beam_search import BeamHypotheses, BeamScorer, BeamSearchScorer, ConstrainedBeamSearchScorer
         from .logits_process import (
             EncoderNoRepeatNGramLogitsProcessor,
+            EpsilonLogitsWarper,
+            EtaLogitsWarper,
             ExponentialDecayLengthPenalty,
             ForcedBOSTokenLogitsProcessor,
             ForcedEOSTokenLogitsProcessor,

src/transformers/generation/configuration_utils.py

@@ -109,6 +109,18 @@ class GenerationConfig(PushToHubMixin):
             generated. If set to float < 1, the smallest set of the most locally typical tokens with probabilities that
             add up to `typical_p` or higher are kept for generation. See [this
             paper](https://arxiv.org/pdf/2202.00666.pdf) for more details.
+        epsilon_cutoff (`float`, *optional*, defaults to 0.0):
+            If set to float strictly between 0 and 1, only tokens with a conditional probability greater than
+            `epsilon_cutoff` will be sampled. In the paper, suggested values range from 3e-4 to 9e-4, depending on the
+            size of the model. See [Truncation Sampling as Language Model
+            Desmoothing](https://arxiv.org/abs/2210.15191) for more details.
+        eta_cutoff (`float`, *optional*, defaults to 0.0):
+            Eta sampling is a hybrid of locally typical sampling and epsilon sampling. If set to float strictly between
+            0 and 1, a token is only considered if it is greater than either `eta_cutoff` or `sqrt(eta_cutoff) *
+            exp(-entropy(softmax(next_token_logits)))`. The latter term is intuitively the expected next token
+            probability, scaled by `sqrt(eta_cutoff)`. In the paper, suggested values range from 3e-4 to 2e-3,
+            depending on the size of the model. See [Truncation Sampling as Language Model
+            Desmoothing](https://arxiv.org/abs/2210.15191) for more details.
         diversity_penalty (`float`, *optional*, defaults to 0.0):
             This value is subtracted from a beam's score if it generates a token same as any beam from other group at a
             particular time. Note that `diversity_penalty` is only effective if `group beam search` is enabled.
@@ -223,6 +235,8 @@ class GenerationConfig(PushToHubMixin):
         self.top_k = kwargs.pop("top_k", 50)
         self.top_p = kwargs.pop("top_p", 1.0)
         self.typical_p = kwargs.pop("typical_p", 1.0)
+        self.epsilon_cutoff = kwargs.pop("epsilon_cutoff", 0.0)
+        self.eta_cutoff = kwargs.pop("eta_cutoff", 0.0)
         self.diversity_penalty = kwargs.pop("diversity_penalty", 0.0)
         self.repetition_penalty = kwargs.pop("repetition_penalty", 1.0)
         self.length_penalty = kwargs.pop("length_penalty", 1.0)

src/transformers/generation/logits_process.py

@@ -138,7 +138,6 @@ class MinNewTokensLengthLogitsProcessor(LogitsProcessor):
     """
 
     def __init__(self, prompt_length_to_skip: int, min_new_tokens: int, eos_token_id: int):
-
         for arg_name, arg_value in [
             ("prompt_length_to_skip", prompt_length_to_skip),
             ("min_new_tokens", min_new_tokens),
@@ -152,7 +151,6 @@ class MinNewTokensLengthLogitsProcessor(LogitsProcessor):
         self.eos_token_id = eos_token_id
 
     def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
-
         new_tokens_length = input_ids.shape[-1] - self.prompt_length_to_skip
         if new_tokens_length < self.min_new_tokens:
             scores[:, self.eos_token_id] = -float("inf")
@@ -297,7 +295,6 @@ class TypicalLogitsWarper(LogitsWarper):
         self.min_tokens_to_keep = min_tokens_to_keep
 
     def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
-
         # calculate entropy
         normalized = torch.nn.functional.log_softmax(scores, dim=-1)
         p = torch.exp(normalized)
@@ -322,6 +319,90 @@ class TypicalLogitsWarper(LogitsWarper):
         return scores
 
 
+class EpsilonLogitsWarper(LogitsWarper):
+    r"""
+    [`LogitsWarper`] that performs epsilon-sampling, i.e. restricting to tokens with `prob >= epsilon`. Takes the
+    largest min_tokens_to_keep tokens if no tokens satisfy this constraint. See [Truncation Sampling as Language Model
+    Desmoothing](https://arxiv.org/abs/2210.15191) for more information.
+
+    Args:
+        epsilon (`float`):
+            If set to > 0, only the most tokens with probabilities `epsilon` or higher are kept for generation.
+        filter_value (`float`, *optional*, defaults to `-float("Inf")`):
+            All filtered values will be set to this float value.
+        min_tokens_to_keep (`int`, *optional*, defaults to 1):
+            Minimum number of tokens that cannot be filtered.
+    """
+
+    def __init__(self, epsilon: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
+        epsilon = float(epsilon)
+        if epsilon <= 0 or epsilon >= 1:
+            raise ValueError(f"`epsilon_cutoff` has to be a float > 0 and < 1, but is {epsilon}")
+
+        min_tokens_to_keep = int(min_tokens_to_keep)
+        if min_tokens_to_keep < 1:
+            raise ValueError(
+                f"`min_tokens_to_keep` has to be a strictly positive integer, but is {min_tokens_to_keep}"
+            )
+
+        self.epsilon = epsilon
+        self.filter_value = filter_value
+        self.min_tokens_to_keep = min_tokens_to_keep
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        # Determine which indices to remove
+        probabilities = scores.softmax(dim=-1)
+        indices_to_remove = probabilities < self.epsilon
+
+        # Keep the words with the 'min_tokens_to_keep'-highest probabilities
+        top_k = min(self.min_tokens_to_keep, scores.size(-1))  # Safety check
+        indices_to_remove = indices_to_remove & (scores < torch.topk(scores, top_k)[0][..., -1, None])
+
+        scores = scores.masked_fill(indices_to_remove, self.filter_value)
+        return scores
+
+
+class EtaLogitsWarper(LogitsWarper):
+    r"""
+    [`LogitsWarper`] that performs eta-sampling, i.e. calculates a dynamic cutoff `eta := min(epsilon, sqrt(epsilon,
+    e^-entropy(probabilities)))` and restricts to tokens with `prob >= eta`. Takes the largest min_tokens_to_keep
+    tokens if no tokens satisfy this constraint. See [Truncation Sampling as Language Model
+    Desmoothing](https://arxiv.org/abs/2210.15191) for more information.
+
+    Args:
+        min_tokens_to_keep (`int`, *optional*, defaults to 1):
+            Minimum number of tokens that cannot be filtered."""
+
+    def __init__(self, epsilon: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
+        epsilon = float(epsilon)
+        if epsilon <= 0 or epsilon >= 1:
+            raise ValueError(f"`eta_cutoff` has to be a float > 0 and < 1, but is {epsilon}")
+
+        min_tokens_to_keep = int(min_tokens_to_keep)
+        if min_tokens_to_keep < 1:
+            raise ValueError(
+                f"`min_tokens_to_keep` has to be a strictly positive integer, but is {min_tokens_to_keep}"
+            )
+
+        self.epsilon = torch.tensor(epsilon)
+        self.filter_value = filter_value
+        self.min_tokens_to_keep = min_tokens_to_keep
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        # Calculate the adaptive cutoff
+        probabilities = scores.softmax(dim=-1)
+        entropy = torch.distributions.Categorical(probs=probabilities).entropy()
+        eta = torch.min(self.epsilon, torch.sqrt(self.epsilon) * torch.exp(-entropy))[..., None]
+        indices_to_remove = probabilities < eta
+
+        # Keep the words with the 'min_tokens_to_keep'-highest probabilities
+        top_k = min(self.min_tokens_to_keep, scores.size(-1))  # Safety check
+        indices_to_remove = indices_to_remove & (scores < torch.topk(scores, top_k)[0][..., -1, None])
+
+        scores = scores.masked_fill(indices_to_remove, self.filter_value)
+        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):
@@ -438,7 +519,6 @@ class NoBadWordsLogitsProcessor(LogitsProcessor):
     """
 
     def __init__(self, bad_words_ids: List[List[int]], eos_token_id: Union[int, List[int]]):
-
         if not isinstance(bad_words_ids, List) or len(bad_words_ids) == 0:
             raise ValueError(f"`bad_words_ids` has to be a non-empty list, but is {bad_words_ids}.")
         if any(not isinstance(bad_word_ids, list) for bad_word_ids in bad_words_ids):

src/transformers/generation/utils.py

@@ -39,6 +39,8 @@ from .beam_search import BeamScorer, BeamSearchScorer, ConstrainedBeamSearchScor
 from .configuration_utils import GenerationConfig
 from .logits_process import (
     EncoderNoRepeatNGramLogitsProcessor,
+    EpsilonLogitsWarper,
+    EtaLogitsWarper,
     ExponentialDecayLengthPenalty,
     ForcedBOSTokenLogitsProcessor,
     ForcedEOSTokenLogitsProcessor,
@@ -750,23 +752,22 @@ class GenerationMixin:
         # all samplers can be found in `generation_utils_samplers.py`
         if generation_config.temperature is not None and generation_config.temperature != 1.0:
             warpers.append(TemperatureLogitsWarper(generation_config.temperature))
+        min_tokens_to_keep = 2 if generation_config.num_beams > 1 else 1
         if generation_config.top_k is not None and generation_config.top_k != 0:
-            warpers.append(
+            warpers.append(TopKLogitsWarper(top_k=generation_config.top_k, min_tokens_to_keep=min_tokens_to_keep))
-                TopKLogitsWarper(
-                    top_k=generation_config.top_k, min_tokens_to_keep=(2 if generation_config.num_beams > 1 else 1)
-                )
-            )
         if generation_config.top_p is not None and generation_config.top_p < 1.0:
-            warpers.append(
+            warpers.append(TopPLogitsWarper(top_p=generation_config.top_p, min_tokens_to_keep=min_tokens_to_keep))
-                TopPLogitsWarper(
-                    top_p=generation_config.top_p, min_tokens_to_keep=(2 if generation_config.num_beams > 1 else 1)
-                )
-            )
         if generation_config.typical_p is not None and generation_config.typical_p < 1.0:
             warpers.append(
-                TypicalLogitsWarper(
+                TypicalLogitsWarper(mass=generation_config.typical_p, min_tokens_to_keep=min_tokens_to_keep)
-                    mass=generation_config.typical_p, min_tokens_to_keep=(2 if generation_config.num_beams > 1 else 1)
+            )
-                )
+        if generation_config.epsilon_cutoff is not None and 0.0 < generation_config.epsilon_cutoff < 1.0:
+            warpers.append(
+                EpsilonLogitsWarper(epsilon=generation_config.epsilon_cutoff, min_tokens_to_keep=min_tokens_to_keep)
+            )
+        if generation_config.eta_cutoff is not None and 0.0 < generation_config.eta_cutoff < 1.0:
+            warpers.append(
+                EtaLogitsWarper(epsilon=generation_config.eta_cutoff, min_tokens_to_keep=min_tokens_to_keep)
             )
         # `LogitNormalization` should always be the last logit processor, when present
         if generation_config.renormalize_logits is True:
@@ -1311,7 +1312,6 @@ class GenerationMixin:
             )
 
         elif is_contrastive_search_gen_mode:
-
             if generation_config.num_return_sequences > 1:
                 raise ValueError(
                     f"num_return_sequences has to be 1, but is {generation_config.num_return_sequences} when doing"
@@ -1716,7 +1716,6 @@ class GenerationMixin:
             # if the first step in the loop, encode all the prefix and obtain: (1) past_key_values;
             # (2) last_hidden_states; (3) logit_for_next_step; (4) update model kwargs for the next step
             if model_kwargs.get("past_key_values") is None:
-
                 # prepare inputs
                 model_kwargs["use_cache"] = True
                 model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)

tests/generation/test_logits_process.py

@@ -28,6 +28,8 @@ if is_torch_available():
 
     from transformers.generation import (
         EncoderNoRepeatNGramLogitsProcessor,
+        EpsilonLogitsWarper,
+        EtaLogitsWarper,
         ExponentialDecayLengthPenalty,
         ForcedBOSTokenLogitsProcessor,
         ForcedEOSTokenLogitsProcessor,
@@ -288,6 +290,80 @@ class LogitsProcessorTest(unittest.TestCase):
         # first batch should keep two tokens, second batch would keep only 1, but due to `min_tokens_to_keep=2` keeps 2.
         self.assertListEqual((filtered_dist != 0.0).to(torch.long).sum(dim=-1).tolist(), [2, 2])
 
+    def test_epsilon_dist_warper(self):
+        input_ids = None
+        vocab_size = 10
+        batch_size = 2
+
+        # create distribution and take log (inverse to Softmax as taken in TopPLogitsWarper)
+        dist = torch.log(
+            torch.tensor(
+                [[0.87, 0.099, 0.001, 0.03], [0.4, 0.299, 0.101, 0.2]], device=torch_device, dtype=torch.float
+            )
+        )
+
+        epsilon_warp = EpsilonLogitsWarper(0.1)
+        filtered_dist = torch.exp(epsilon_warp(input_ids, dist))
+
+        # dist should be filtered to only keep values with proba >= 0.1
+        # exp (-inf) => 0
+        EXPECTED_FILTERED_DIST = torch.tensor(
+            [[0.87, 0, 0, 0], [0.4, 0.299, 0.101, 0.2]], device=torch_device, dtype=torch.float
+        )
+        self.assertTrue(torch.allclose(filtered_dist, EXPECTED_FILTERED_DIST, atol=1e-3))
+
+        # check edge cases with negative and extreme logits
+        ramp_logits = torch.arange(vocab_size, device=torch_device, dtype=torch.float).unsqueeze(0).repeat(
+            batch_size, 1
+        ) - (vocab_size // 2)
+
+        # make ramp_logits more extreme
+        ramp_logits[1] = ramp_logits[1] * 100.0
+
+        # make sure at least 2 tokens are kept
+        epsilon_warp = EpsilonLogitsWarper(5e-2, min_tokens_to_keep=2, filter_value=0.0)
+        filtered_dist = epsilon_warp(input_ids, ramp_logits)
+
+        # first batch should keep 3 tokens, second batch would keep only 1, but due to `min_tokens_to_keep=2` keeps 2.
+        self.assertListEqual((filtered_dist != 0.0).to(torch.long).sum(dim=-1).tolist(), [3, 2])
+
+    def test_eta_dist_warper(self):
+        input_ids = None
+        vocab_size = 10
+        batch_size = 2
+
+        # create distribution and take log (inverse to Softmax as taken in TopPLogitsWarper)
+        dist = torch.log(
+            torch.tensor([[0.0, 0.1, 0.8, 0.1], [0.01, 0.04, 0.9, 0.05]], device=torch_device, dtype=torch.float)
+        )
+
+        eta_warp = EtaLogitsWarper(0.0625)
+        filtered_dist = torch.exp(eta_warp(input_ids, dist))
+
+        # dist should be filtered to only keep values with proba >= min(0.0625, sqrt(0.0625) * e^-H(p))
+        # min(0.0625, 0.1320) is the cutoff for the first row and min(0.0625, 0.1644) is for the second
+        # where H is the entropy function and p is the probability vector.
+        # exp (-inf) => 0
+        EXPECTED_FILTERED_DIST = torch.tensor(
+            [[0.0, 0.1, 0.8, 0.1], [0.0, 0.0, 0.9, 0.0]], device=torch_device, dtype=torch.float
+        )
+        self.assertTrue(torch.allclose(filtered_dist, EXPECTED_FILTERED_DIST, atol=1e-3))
+
+        # check edge cases with negative and extreme logits
+        ramp_logits = torch.arange(vocab_size, device=torch_device, dtype=torch.float).unsqueeze(0).repeat(
+            batch_size, 1
+        ) - (vocab_size // 2)
+
+        # make ramp_logits more extreme
+        ramp_logits[1] = ramp_logits[1] * 100.0
+
+        # make sure at least 2 tokens are kept
+        eta_warp = EtaLogitsWarper(0.1, min_tokens_to_keep=2, filter_value=0.0)
+        filtered_dist = eta_warp(input_ids, ramp_logits)
+
+        # first batch should keep 2 tokens, second batch would keep only 1, but due to `min_tokens_to_keep=2` keeps 2.
+        self.assertListEqual((filtered_dist != 0.0).to(torch.long).sum(dim=-1).tolist(), [2, 2])
+
     def test_no_repeat_ngram_dist_processor(self):
         vocab_size = 3
         batch_size = 2