Add support for pretraining recurring span selection to Splinter (huggingface#17247)

* Add SplinterForSpanSelection for pre-training recurring span selection.

* Formatting.

* Rename SplinterForSpanSelection to SplinterForPreTraining.

* Ensure repo consistency

* Fixup changes

* Address SplinterForPreTraining PR comments

* Incorporate feedback and derive multiple question tokens per example.

* Update src/transformers/models/splinter/modeling_splinter.py

Co-authored-by: Patrick von Platen <[email protected]>

* Update src/transformers/models/splinter/modeling_splinter.py

Co-authored-by: Patrick von Platen <[email protected]>

Co-authored-by: Jean Vancoppenole <[email protected]>
Co-authored-by: Tobias Günther <[email protected]>
Co-authored-by: Tobias Günther <[email protected]>
Co-authored-by: Patrick von Platen <[email protected]>

Author: 5 people

docs/source/en/model_doc/splinter.mdx

@@ -72,3 +72,8 @@ This model was contributed by [yuvalkirstain](https://huggingface.co/yuvalkirsta
 
 [[autodoc]] SplinterForQuestionAnswering
     - forward
+
+## SplinterForPreTraining
+
+[[autodoc]] SplinterForPreTraining
+    - forward

src/transformers/__init__.py

@@ -1532,6 +1532,7 @@
     _import_structure["models.splinter"].extend(
         [
             "SPLINTER_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "SplinterForPreTraining",
             "SplinterForQuestionAnswering",
             "SplinterLayer",
             "SplinterModel",
@@ -3830,6 +3831,7 @@
         from .models.speech_to_text_2 import Speech2Text2ForCausalLM, Speech2Text2PreTrainedModel
         from .models.splinter import (
             SPLINTER_PRETRAINED_MODEL_ARCHIVE_LIST,
+            SplinterForPreTraining,
             SplinterForQuestionAnswering,
             SplinterLayer,
             SplinterModel,

src/transformers/models/auto/modeling_auto.py

@@ -161,6 +161,7 @@
         ("openai-gpt", "OpenAIGPTLMHeadModel"),
         ("retribert", "RetriBertModel"),
         ("roberta", "RobertaForMaskedLM"),
+        ("splinter", "SplinterForPreTraining"),
         ("squeezebert", "SqueezeBertForMaskedLM"),
         ("t5", "T5ForConditionalGeneration"),
         ("tapas", "TapasForMaskedLM"),

src/transformers/models/splinter/__init__.py

@@ -42,6 +42,7 @@
     _import_structure["modeling_splinter"] = [
         "SPLINTER_PRETRAINED_MODEL_ARCHIVE_LIST",
         "SplinterForQuestionAnswering",
+        "SplinterForPreTraining",
         "SplinterLayer",
         "SplinterModel",
         "SplinterPreTrainedModel",
@@ -68,6 +69,7 @@
     else:
         from .modeling_splinter import (
             SPLINTER_PRETRAINED_MODEL_ARCHIVE_LIST,
+            SplinterForPreTraining,
             SplinterForQuestionAnswering,
             SplinterLayer,
             SplinterModel,

src/transformers/models/splinter/modeling_splinter.py

@@ -16,6 +16,7 @@
 
 
 import math
+from dataclasses import dataclass
 from typing import List, Optional, Tuple, Union
 
 import torch
@@ -24,7 +25,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
-from ...modeling_outputs import BaseModelOutputWithPastAndCrossAttentions, QuestionAnsweringModelOutput
+from ...modeling_outputs import BaseModelOutputWithPastAndCrossAttentions, ModelOutput, QuestionAnsweringModelOutput
 from ...modeling_utils import PreTrainedModel
 from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
 from ...utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
@@ -940,3 +941,171 @@ def forward(
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
         )
+
+
+@dataclass
+class SplinterForPreTrainingOutput(ModelOutput):
+    """
+    Class for outputs of Splinter as a span selection model.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when start and end positions are provided):
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        start_logits (`torch.FloatTensor` of shape `(batch_size, num_questions, sequence_length)`):
+            Span-start scores (before SoftMax).
+        end_logits (`torch.FloatTensor` of shape `(batch_size, num_questions, sequence_length)`):
+            Span-end scores (before SoftMax).
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    start_logits: torch.FloatTensor = None
+    end_logits: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@add_start_docstrings(
+    """
+    Splinter Model for the recurring span selection task as done during the pretraining. The difference to the QA task
+    is that we do not have a question, but multiple question tokens that replace the occurrences of recurring spans
+    instead.
+    """,
+    SPLINTER_START_DOCSTRING,
+)
+class SplinterForPreTraining(SplinterPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.splinter = SplinterModel(config)
+        self.splinter_qass = QuestionAwareSpanSelectionHead(config)
+        self.question_token_id = config.question_token_id
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(
+        SPLINTER_INPUTS_DOCSTRING.format("batch_size, num_questions, sequence_length")
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        question_positions: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, SplinterForPreTrainingOutput]:
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size, num_questions)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`torch.LongTensor` of shape `(batch_size, num_questions)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        question_positions (`torch.LongTensor` of shape `(batch_size, num_questions)`, *optional*):
+            The positions of all question tokens. If given, start_logits and end_logits will be of shape `(batch_size,
+            num_questions, sequence_length)`. If None, the first question token in each sequence in the batch will be
+            the only one for which start_logits and end_logits are calculated and they will be of shape `(batch_size,
+            sequence_length)`.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if question_positions is None and start_positions is not None and end_positions is not None:
+            raise TypeError("question_positions must be specified in order to calculate the loss")
+
+        elif question_positions is None and input_ids is None:
+            raise TypeError("question_positions must be specified when input_embeds is used")
+
+        elif question_positions is None:
+            question_positions = self._prepare_question_positions(input_ids)
+
+        outputs = self.splinter(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+        batch_size, sequence_length, dim = sequence_output.size()
+        # [batch_size, num_questions, sequence_length]
+        start_logits, end_logits = self.splinter_qass(sequence_output, question_positions)
+
+        num_questions = question_positions.size(1)
+        if attention_mask is not None:
+            attention_mask_for_each_question = attention_mask.unsqueeze(1).expand(
+                batch_size, num_questions, sequence_length
+            )
+            start_logits = start_logits + (1 - attention_mask_for_each_question) * -10000.0
+            end_logits = end_logits + (1 - attention_mask_for_each_question) * -10000.0
+
+        total_loss = None
+        # [batch_size, num_questions, sequence_length]
+        if start_positions is not None and end_positions is not None:
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            start_positions.clamp_(0, max(0, sequence_length - 1))
+            end_positions.clamp_(0, max(0, sequence_length - 1))
+
+            # Ignore zero positions in the loss. Splinter never predicts zero
+            # during pretraining and zero is used for padding question
+            # tokens as well as for start and end positions of padded
+            # question tokens.
+            loss_fct = CrossEntropyLoss(ignore_index=self.config.pad_token_id)
+            start_loss = loss_fct(
+                start_logits.view(batch_size * num_questions, sequence_length),
+                start_positions.view(batch_size * num_questions),
+            )
+            end_loss = loss_fct(
+                end_logits.view(batch_size * num_questions, sequence_length),
+                end_positions.view(batch_size * num_questions),
+            )
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[1:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return SplinterForPreTrainingOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def _prepare_question_positions(self, input_ids: torch.Tensor) -> torch.Tensor:
+        rows, flat_positions = torch.where(input_ids == self.config.question_token_id)
+        num_questions = torch.bincount(rows)
+        positions = torch.full(
+            (input_ids.size(0), num_questions.max()),
+            self.config.pad_token_id,
+            dtype=torch.long,
+            device=input_ids.device,
+        )
+        cols = torch.cat([torch.arange(n) for n in num_questions])
+        positions[rows, cols] = flat_positions
+        return positions