image-match/python/py/Lib/site-packages/transformers/models/layoutlm/modeling_layoutlm.py

# Copyright 2018 The Microsoft Research Asia LayoutLM Team Authors and the HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch LayoutLM model."""

from collections.abc import Callable

import torch
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss

from ... import initialization as init
from ...activations import ACT2FN
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPooling,
    MaskedLMOutput,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutput,
    TokenClassifierOutput,
)
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...pytorch_utils import apply_chunking_to_forward
from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
from ...utils.generic import merge_with_config_defaults
from ...utils.output_capturing import capture_outputs
from .configuration_layoutlm import LayoutLMConfig


logger = logging.get_logger(__name__)


LayoutLMLayerNorm = nn.LayerNorm


class LayoutLMEmbeddings(nn.Module):
    """Construct the embeddings from word, position and token_type embeddings."""

    def __init__(self, config):
        super().__init__()
        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
        self.x_position_embeddings = nn.Embedding(config.max_2d_position_embeddings, config.hidden_size)
        self.y_position_embeddings = nn.Embedding(config.max_2d_position_embeddings, config.hidden_size)
        self.h_position_embeddings = nn.Embedding(config.max_2d_position_embeddings, config.hidden_size)
        self.w_position_embeddings = nn.Embedding(config.max_2d_position_embeddings, config.hidden_size)
        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)

        self.LayerNorm = LayoutLMLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

        self.register_buffer(
            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
        )

    def forward(
        self,
        input_ids=None,
        bbox=None,
        token_type_ids=None,
        position_ids=None,
        inputs_embeds=None,
    ):
        if input_ids is not None:
            input_shape = input_ids.size()
        else:
            input_shape = inputs_embeds.size()[:-1]

        seq_length = input_shape[1]

        device = input_ids.device if input_ids is not None else inputs_embeds.device

        if position_ids is None:
            position_ids = self.position_ids[:, :seq_length]

        if token_type_ids is None:
            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)

        if inputs_embeds is None:
            inputs_embeds = self.word_embeddings(input_ids)

        words_embeddings = inputs_embeds
        position_embeddings = self.position_embeddings(position_ids)
        try:
            left_position_embeddings = self.x_position_embeddings(bbox[:, :, 0])
            upper_position_embeddings = self.y_position_embeddings(bbox[:, :, 1])
            right_position_embeddings = self.x_position_embeddings(bbox[:, :, 2])
            lower_position_embeddings = self.y_position_embeddings(bbox[:, :, 3])
        except IndexError as e:
            raise IndexError("The `bbox`coordinate values should be within 0-1000 range.") from e

        h_position_embeddings = self.h_position_embeddings(bbox[:, :, 3] - bbox[:, :, 1])
        w_position_embeddings = self.w_position_embeddings(bbox[:, :, 2] - bbox[:, :, 0])
        token_type_embeddings = self.token_type_embeddings(token_type_ids)

        embeddings = (
            words_embeddings
            + position_embeddings
            + left_position_embeddings
            + upper_position_embeddings
            + right_position_embeddings
            + lower_position_embeddings
            + h_position_embeddings
            + w_position_embeddings
            + token_type_embeddings
        )
        embeddings = self.LayerNorm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings


# Copied from transformers.models.align.modeling_align.eager_attention_forward
def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor | None,
    scaling: float,
    dropout: float = 0.0,
    **kwargs,
):
    attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling
    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)

    attn_output = torch.matmul(attn_weights, value)
    attn_output = attn_output.transpose(1, 2).contiguous()
    return attn_output, attn_weights


# Copied from transformers.models.align.modeling_align.AlignTextSelfAttention with AlignText->LayoutLM
class LayoutLMSelfAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
            raise ValueError(
                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
                f"heads ({config.num_attention_heads})"
            )

        self.config = config
        self.num_attention_heads = config.num_attention_heads
        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.query = nn.Linear(config.hidden_size, self.all_head_size)
        self.key = nn.Linear(config.hidden_size, self.all_head_size)
        self.value = nn.Linear(config.hidden_size, self.all_head_size)

        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
        self.attention_dropout = config.attention_probs_dropout_prob
        self.scaling = self.attention_head_size**-0.5

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.FloatTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.attention_head_size)

        query_states = self.query(hidden_states).view(hidden_shape).transpose(1, 2)
        key_states = self.key(hidden_states).view(hidden_shape).transpose(1, 2)
        value_states = self.value(hidden_states).view(hidden_shape).transpose(1, 2)

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        return attn_output, attn_weights


# Copied from transformers.models.bert.modeling_bert.BertSelfOutput with Bert->LayoutLM
class LayoutLMSelfOutput(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states


# Copied from transformers.models.align.modeling_align.AlignTextAttention with AlignText->LayoutLM
class LayoutLMAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.self = LayoutLMSelfAttention(config)
        self.output = LayoutLMSelfOutput(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.FloatTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states, _ = self.self(
            hidden_states,
            attention_mask=attention_mask,
            **kwargs,
        )
        hidden_states = self.output(hidden_states, residual)
        return hidden_states


# Copied from transformers.models.bert.modeling_bert.BertIntermediate
class LayoutLMIntermediate(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
        if isinstance(config.hidden_act, str):
            self.intermediate_act_fn = ACT2FN[config.hidden_act]
        else:
            self.intermediate_act_fn = config.hidden_act

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.dense(hidden_states)
        hidden_states = self.intermediate_act_fn(hidden_states)
        return hidden_states


# Copied from transformers.models.bert.modeling_bert.BertOutput with Bert->LayoutLM
class LayoutLMOutput(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states


# Copied from transformers.models.align.modeling_align.AlignTextLayer with AlignText->LayoutLM
class LayoutLMLayer(GradientCheckpointingLayer):
    def __init__(self, config):
        super().__init__()
        self.chunk_size_feed_forward = config.chunk_size_feed_forward
        self.seq_len_dim = 1
        self.attention = LayoutLMAttention(config)
        self.intermediate = LayoutLMIntermediate(config)
        self.output = LayoutLMOutput(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.FloatTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        hidden_states = self.attention(
            hidden_states,
            attention_mask=attention_mask,
            **kwargs,
        )

        hidden_states = apply_chunking_to_forward(
            self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, hidden_states
        )

        return hidden_states

    def feed_forward_chunk(self, attention_output):
        intermediate_output = self.intermediate(attention_output)
        layer_output = self.output(intermediate_output, attention_output)
        return layer_output


# Copied from transformers.models.align.modeling_align.AlignTextEncoder with AlignText->LayoutLM
class LayoutLMEncoder(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.layer = nn.ModuleList([LayoutLMLayer(config) for i in range(config.num_hidden_layers)])
        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.FloatTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutput:
        for layer_module in self.layer:
            hidden_states = layer_module(
                hidden_states,
                attention_mask,
                **kwargs,
            )

        return BaseModelOutput(
            last_hidden_state=hidden_states,
        )


# Copied from transformers.models.bert.modeling_bert.BertPooler
class LayoutLMPooler(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.activation = nn.Tanh()

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        # We "pool" the model by simply taking the hidden state corresponding
        # to the first token.
        first_token_tensor = hidden_states[:, 0]
        pooled_output = self.dense(first_token_tensor)
        pooled_output = self.activation(pooled_output)
        return pooled_output


# Copied from transformers.models.bert.modeling_bert.BertPredictionHeadTransform with Bert->LayoutLM
class LayoutLMPredictionHeadTransform(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        if isinstance(config.hidden_act, str):
            self.transform_act_fn = ACT2FN[config.hidden_act]
        else:
            self.transform_act_fn = config.hidden_act
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.dense(hidden_states)
        hidden_states = self.transform_act_fn(hidden_states)
        hidden_states = self.LayerNorm(hidden_states)
        return hidden_states


# Copied from transformers.models.bert.modeling_bert.BertLMPredictionHead with Bert->LayoutLM
class LayoutLMLMPredictionHead(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.transform = LayoutLMPredictionHeadTransform(config)

        # The output weights are the same as the input embeddings, but there is
        # an output-only bias for each token.
        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=True)
        self.bias = nn.Parameter(torch.zeros(config.vocab_size))

    def forward(self, hidden_states):
        hidden_states = self.transform(hidden_states)
        hidden_states = self.decoder(hidden_states)
        return hidden_states


# Copied from transformers.models.bert.modeling_bert.BertOnlyMLMHead with Bert->LayoutLM
class LayoutLMOnlyMLMHead(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.predictions = LayoutLMLMPredictionHead(config)

    def forward(self, sequence_output: torch.Tensor) -> torch.Tensor:
        prediction_scores = self.predictions(sequence_output)
        return prediction_scores


@auto_docstring
class LayoutLMPreTrainedModel(PreTrainedModel):
    config: LayoutLMConfig
    base_model_prefix = "layoutlm"
    supports_gradient_checkpointing = True
    _can_record_outputs = {
        "hidden_states": LayoutLMLayer,
        "attentions": LayoutLMSelfAttention,
    }

    @torch.no_grad()
    def _init_weights(self, module):
        """Initialize the weights"""
        super()._init_weights(module)
        if isinstance(module, LayoutLMLMPredictionHead):
            init.zeros_(module.bias)
        elif isinstance(module, LayoutLMEmbeddings):
            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))


@auto_docstring
class LayoutLMModel(LayoutLMPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.config = config

        self.embeddings = LayoutLMEmbeddings(config)
        self.encoder = LayoutLMEncoder(config)
        self.pooler = LayoutLMPooler(config)

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.embeddings.word_embeddings

    def set_input_embeddings(self, value):
        self.embeddings.word_embeddings = value

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        bbox: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        bbox (`torch.LongTensor` of shape `(batch_size, sequence_length, 4)`, *optional*):
            Bounding boxes of each input sequence tokens. Selected in the range `[0,
            config.max_2d_position_embeddings-1]`. Each bounding box should be a normalized version in (x0, y0, x1, y1)
            format, where (x0, y0) corresponds to the position of the upper left corner in the bounding box, and (x1,
            y1) represents the position of the lower right corner. See [Overview](#Overview) for normalization.

        Examples:

        ```python
        >>> from transformers import AutoTokenizer, LayoutLMModel
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/layoutlm-base-uncased")
        >>> model = LayoutLMModel.from_pretrained("microsoft/layoutlm-base-uncased")

        >>> words = ["Hello", "world"]
        >>> normalized_word_boxes = [637, 773, 693, 782], [698, 773, 733, 782]

        >>> token_boxes = []
        >>> for word, box in zip(words, normalized_word_boxes):
        ...     word_tokens = tokenizer.tokenize(word)
        ...     token_boxes.extend([box] * len(word_tokens))
        >>> # add bounding boxes of cls + sep tokens
        >>> token_boxes = [[0, 0, 0, 0]] + token_boxes + [[1000, 1000, 1000, 1000]]

        >>> encoding = tokenizer(" ".join(words), return_tensors="pt")
        >>> input_ids = encoding["input_ids"]
        >>> attention_mask = encoding["attention_mask"]
        >>> token_type_ids = encoding["token_type_ids"]
        >>> bbox = torch.tensor([token_boxes])

        >>> outputs = model(
        ...     input_ids=input_ids, bbox=bbox, attention_mask=attention_mask, token_type_ids=token_type_ids
        ... )

        >>> last_hidden_states = outputs.last_hidden_state
        ```"""
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
            input_shape = input_ids.size()
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        device = input_ids.device if input_ids is not None else inputs_embeds.device

        if attention_mask is None:
            attention_mask = torch.ones(input_shape, device=device)
        if token_type_ids is None:
            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)

        if bbox is None:
            bbox = torch.zeros(input_shape + (4,), dtype=torch.long, device=device)

        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)

        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)
        extended_attention_mask = (1.0 - extended_attention_mask) * torch.finfo(self.dtype).min

        embedding_output = self.embeddings(
            input_ids=input_ids,
            bbox=bbox,
            position_ids=position_ids,
            token_type_ids=token_type_ids,
            inputs_embeds=inputs_embeds,
        )
        encoder_outputs = self.encoder(
            embedding_output,
            extended_attention_mask,
            **kwargs,
        )
        sequence_output = encoder_outputs[0]
        pooled_output = self.pooler(sequence_output)

        return BaseModelOutputWithPooling(
            last_hidden_state=sequence_output,
            pooler_output=pooled_output,
        )


@auto_docstring
class LayoutLMForMaskedLM(LayoutLMPreTrainedModel):
    _tied_weights_keys = {
        "cls.predictions.decoder.bias": "cls.predictions.bias",
        "cls.predictions.decoder.weight": "layoutlm.embeddings.word_embeddings.weight",
    }

    def __init__(self, config):
        super().__init__(config)

        self.layoutlm = LayoutLMModel(config)
        self.cls = LayoutLMOnlyMLMHead(config)

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.layoutlm.embeddings.word_embeddings

    def get_output_embeddings(self):
        return self.cls.predictions.decoder

    def set_output_embeddings(self, new_embeddings):
        self.cls.predictions.decoder = new_embeddings
        self.cls.predictions.bias = new_embeddings.bias

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        bbox: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | MaskedLMOutput:
        r"""
        bbox (`torch.LongTensor` of shape `(batch_size, sequence_length, 4)`, *optional*):
            Bounding boxes of each input sequence tokens. Selected in the range `[0,
            config.max_2d_position_embeddings-1]`. Each bounding box should be a normalized version in (x0, y0, x1, y1)
            format, where (x0, y0) corresponds to the position of the upper left corner in the bounding box, and (x1,
            y1) represents the position of the lower right corner. See [Overview](#Overview) for normalization.
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
            config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
            loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`

        Examples:

        ```python
        >>> from transformers import AutoTokenizer, LayoutLMForMaskedLM
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/layoutlm-base-uncased")
        >>> model = LayoutLMForMaskedLM.from_pretrained("microsoft/layoutlm-base-uncased")

        >>> words = ["Hello", "[MASK]"]
        >>> normalized_word_boxes = [637, 773, 693, 782], [698, 773, 733, 782]

        >>> token_boxes = []
        >>> for word, box in zip(words, normalized_word_boxes):
        ...     word_tokens = tokenizer.tokenize(word)
        ...     token_boxes.extend([box] * len(word_tokens))
        >>> # add bounding boxes of cls + sep tokens
        >>> token_boxes = [[0, 0, 0, 0]] + token_boxes + [[1000, 1000, 1000, 1000]]

        >>> encoding = tokenizer(" ".join(words), return_tensors="pt")
        >>> input_ids = encoding["input_ids"]
        >>> attention_mask = encoding["attention_mask"]
        >>> token_type_ids = encoding["token_type_ids"]
        >>> bbox = torch.tensor([token_boxes])

        >>> labels = tokenizer("Hello world", return_tensors="pt")["input_ids"]

        >>> outputs = model(
        ...     input_ids=input_ids,
        ...     bbox=bbox,
        ...     attention_mask=attention_mask,
        ...     token_type_ids=token_type_ids,
        ...     labels=labels,
        ... )

        >>> loss = outputs.loss
        ```"""
        outputs = self.layoutlm(
            input_ids,
            bbox,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            **kwargs,
        )

        sequence_output = outputs[0]
        prediction_scores = self.cls(sequence_output)

        masked_lm_loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            masked_lm_loss = loss_fct(
                prediction_scores.view(-1, self.config.vocab_size),
                labels.view(-1),
            )

        return MaskedLMOutput(
            loss=masked_lm_loss,
            logits=prediction_scores,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@auto_docstring(
    custom_intro="""
    LayoutLM Model with a sequence classification head on top (a linear layer on top of the pooled output) e.g. for
    document image classification tasks such as the [RVL-CDIP](https://www.cs.cmu.edu/~aharley/rvl-cdip/) dataset.
    """
)
class LayoutLMForSequenceClassification(LayoutLMPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.layoutlm = LayoutLMModel(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.layoutlm.embeddings.word_embeddings

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        bbox: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | SequenceClassifierOutput:
        r"""
        bbox (`torch.LongTensor` of shape `(batch_size, sequence_length, 4)`, *optional*):
            Bounding boxes of each input sequence tokens. Selected in the range `[0,
            config.max_2d_position_embeddings-1]`. Each bounding box should be a normalized version in (x0, y0, x1, y1)
            format, where (x0, y0) corresponds to the position of the upper left corner in the bounding box, and (x1,
            y1) represents the position of the lower right corner. See [Overview](#Overview) for normalization.
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).

        Examples:

        ```python
        >>> from transformers import AutoTokenizer, LayoutLMForSequenceClassification
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/layoutlm-base-uncased")
        >>> model = LayoutLMForSequenceClassification.from_pretrained("microsoft/layoutlm-base-uncased")

        >>> words = ["Hello", "world"]
        >>> normalized_word_boxes = [637, 773, 693, 782], [698, 773, 733, 782]

        >>> token_boxes = []
        >>> for word, box in zip(words, normalized_word_boxes):
        ...     word_tokens = tokenizer.tokenize(word)
        ...     token_boxes.extend([box] * len(word_tokens))
        >>> # add bounding boxes of cls + sep tokens
        >>> token_boxes = [[0, 0, 0, 0]] + token_boxes + [[1000, 1000, 1000, 1000]]

        >>> encoding = tokenizer(" ".join(words), return_tensors="pt")
        >>> input_ids = encoding["input_ids"]
        >>> attention_mask = encoding["attention_mask"]
        >>> token_type_ids = encoding["token_type_ids"]
        >>> bbox = torch.tensor([token_boxes])
        >>> sequence_label = torch.tensor([1])

        >>> outputs = model(
        ...     input_ids=input_ids,
        ...     bbox=bbox,
        ...     attention_mask=attention_mask,
        ...     token_type_ids=token_type_ids,
        ...     labels=sequence_label,
        ... )

        >>> loss = outputs.loss
        >>> logits = outputs.logits
        ```"""
        outputs = self.layoutlm(
            input_ids=input_ids,
            bbox=bbox,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            **kwargs,
        )

        pooled_output = outputs[1]

        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)

        loss = None
        if labels is not None:
            if self.config.problem_type is None:
                if self.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

            if self.config.problem_type == "regression":
                loss_fct = MSELoss()
                if self.num_labels == 1:
                    loss = loss_fct(logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(logits, labels)
            elif self.config.problem_type == "single_label_classification":
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(logits, labels)

        return SequenceClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@auto_docstring(
    custom_intro="""
    LayoutLM Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
    sequence labeling (information extraction) tasks such as the [FUNSD](https://guillaumejaume.github.io/FUNSD/)
    dataset and the [SROIE](https://rrc.cvc.uab.es/?ch=13) dataset.
    """
)
class LayoutLMForTokenClassification(LayoutLMPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.layoutlm = LayoutLMModel(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.layoutlm.embeddings.word_embeddings

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        bbox: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | TokenClassifierOutput:
        r"""
        bbox (`torch.LongTensor` of shape `(batch_size, sequence_length, 4)`, *optional*):
            Bounding boxes of each input sequence tokens. Selected in the range `[0,
            config.max_2d_position_embeddings-1]`. Each bounding box should be a normalized version in (x0, y0, x1, y1)
            format, where (x0, y0) corresponds to the position of the upper left corner in the bounding box, and (x1,
            y1) represents the position of the lower right corner. See [Overview](#Overview) for normalization.
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.

        Examples:

        ```python
        >>> from transformers import AutoTokenizer, LayoutLMForTokenClassification
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/layoutlm-base-uncased")
        >>> model = LayoutLMForTokenClassification.from_pretrained("microsoft/layoutlm-base-uncased")

        >>> words = ["Hello", "world"]
        >>> normalized_word_boxes = [637, 773, 693, 782], [698, 773, 733, 782]

        >>> token_boxes = []
        >>> for word, box in zip(words, normalized_word_boxes):
        ...     word_tokens = tokenizer.tokenize(word)
        ...     token_boxes.extend([box] * len(word_tokens))
        >>> # add bounding boxes of cls + sep tokens
        >>> token_boxes = [[0, 0, 0, 0]] + token_boxes + [[1000, 1000, 1000, 1000]]

        >>> encoding = tokenizer(" ".join(words), return_tensors="pt")
        >>> input_ids = encoding["input_ids"]
        >>> attention_mask = encoding["attention_mask"]
        >>> token_type_ids = encoding["token_type_ids"]
        >>> bbox = torch.tensor([token_boxes])
        >>> token_labels = torch.tensor([1, 1, 0, 0]).unsqueeze(0)  # batch size of 1

        >>> outputs = model(
        ...     input_ids=input_ids,
        ...     bbox=bbox,
        ...     attention_mask=attention_mask,
        ...     token_type_ids=token_type_ids,
        ...     labels=token_labels,
        ... )

        >>> loss = outputs.loss
        >>> logits = outputs.logits
        ```"""
        outputs = self.layoutlm(
            input_ids=input_ids,
            bbox=bbox,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            **kwargs,
        )

        sequence_output = outputs[0]

        sequence_output = self.dropout(sequence_output)
        logits = self.classifier(sequence_output)

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

        return TokenClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@auto_docstring
class LayoutLMForQuestionAnswering(LayoutLMPreTrainedModel):
    def __init__(self, config, has_visual_segment_embedding=True):
        r"""
        has_visual_segment_embedding (`bool`, *optional*, defaults to `True`):
            Whether or not to add visual segment embeddings.
        """
        super().__init__(config)
        self.num_labels = config.num_labels

        self.layoutlm = LayoutLMModel(config)
        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.layoutlm.embeddings.word_embeddings

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        bbox: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        start_positions: torch.LongTensor | None = None,
        end_positions: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | QuestionAnsweringModelOutput:
        r"""
        bbox (`torch.LongTensor` of shape `(batch_size, sequence_length, 4)`, *optional*):
            Bounding boxes of each input sequence tokens. Selected in the range `[0,
            config.max_2d_position_embeddings-1]`. Each bounding box should be a normalized version in (x0, y0, x1, y1)
            format, where (x0, y0) corresponds to the position of the upper left corner in the bounding box, and (x1,
            y1) represents the position of the lower right corner. See [Overview](#Overview) for normalization.

        Example:

        In the example below, we prepare a question + context pair for the LayoutLM model. It will give us a prediction
        of what it thinks the answer is (the span of the answer within the texts parsed from the image).

        ```python
        >>> from transformers import AutoTokenizer, LayoutLMForQuestionAnswering
        >>> from datasets import load_dataset
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("impira/layoutlm-document-qa", add_prefix_space=True)
        >>> model = LayoutLMForQuestionAnswering.from_pretrained("impira/layoutlm-document-qa", revision="1e3ebac")

        >>> dataset = load_dataset("nielsr/funsd", split="train")
        >>> example = dataset[0]
        >>> question = "what's his name?"
        >>> words = example["words"]
        >>> boxes = example["bboxes"]

        >>> encoding = tokenizer(
        ...     question.split(), words, is_split_into_words=True, return_token_type_ids=True, return_tensors="pt"
        ... )
        >>> bbox = []
        >>> for i, s, w in zip(encoding.input_ids[0], encoding.sequence_ids(0), encoding.word_ids(0)):
        ...     if s == 1:
        ...         bbox.append(boxes[w])
        ...     elif i == tokenizer.sep_token_id:
        ...         bbox.append([1000] * 4)
        ...     else:
        ...         bbox.append([0] * 4)
        >>> encoding["bbox"] = torch.tensor([bbox])

        >>> word_ids = encoding.word_ids(0)
        >>> outputs = model(**encoding)
        >>> loss = outputs.loss
        >>> start_scores = outputs.start_logits
        >>> end_scores = outputs.end_logits
        >>> start, end = word_ids[start_scores.argmax(-1)], word_ids[end_scores.argmax(-1)]
        >>> print(" ".join(words[start : end + 1]))
        M. Hamann P. Harper, P. Martinez
        ```"""

        outputs = self.layoutlm(
            input_ids=input_ids,
            bbox=bbox,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            **kwargs,
        )

        sequence_output = outputs[0]

        logits = self.qa_outputs(sequence_output)
        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1).contiguous()
        end_logits = end_logits.squeeze(-1).contiguous()

        total_loss = None
        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, split add a dimension
            if len(start_positions.size()) > 1:
                start_positions = start_positions.squeeze(-1)
            if len(end_positions.size()) > 1:
                end_positions = end_positions.squeeze(-1)
            # sometimes the start/end positions are outside our model inputs, we ignore these terms
            ignored_index = start_logits.size(1)
            start_positions = start_positions.clamp(0, ignored_index)
            end_positions = end_positions.clamp(0, ignored_index)

            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
            start_loss = loss_fct(start_logits, start_positions)
            end_loss = loss_fct(end_logits, end_positions)
            total_loss = (start_loss + end_loss) / 2

        return QuestionAnsweringModelOutput(
            loss=total_loss,
            start_logits=start_logits,
            end_logits=end_logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


__all__ = [
    "LayoutLMForMaskedLM",
    "LayoutLMForSequenceClassification",
    "LayoutLMForTokenClassification",
    "LayoutLMForQuestionAnswering",
    "LayoutLMModel",
    "LayoutLMPreTrainedModel",
]