image-match/python/py/Lib/site-packages/transformers/models/vilt/modeling_vilt.py

# Copyright 2022 NAVER AI Labs and The HuggingFace Inc. team. All rights reserved.
#
# 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 ViLT model."""

import collections.abc
import math
from dataclasses import dataclass

import torch
from torch import nn
from torch.nn import CrossEntropyLoss

from ... import initialization as init
from ...activations import ACT2FN
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPooling,
    MaskedLMOutput,
    ModelOutput,
    SequenceClassifierOutput,
    TokenClassifierOutput,
)
from ...modeling_utils import PreTrainedModel
from ...utils import auto_docstring, logging
from .configuration_vilt import ViltConfig


logger = logging.get_logger(__name__)


@dataclass
@auto_docstring(
    custom_intro="""
    Class for outputs of [`ViltForImagesAndTextClassification`].
    """
)
class ViltForImagesAndTextClassificationOutput(ModelOutput):
    r"""
    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
        Classification (or regression if config.num_labels==1) loss.
    logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels)`):
        Classification (or regression if config.num_labels==1) scores (before SoftMax).
    hidden_states (`list[tuple(torch.FloatTensor)]`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
        List of tuples of `torch.FloatTensor` (one for each image-text pair, each tuple containing the output of
        the embeddings + 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 initial embedding outputs.
    """

    loss: torch.FloatTensor | None = None
    logits: torch.FloatTensor | None = None
    hidden_states: list[tuple[torch.FloatTensor]] | None = None
    attentions: list[tuple[torch.FloatTensor]] | None = None


class ViltEmbeddings(nn.Module):
    """
    Construct the text and patch embeddings.

    Text embeddings are equivalent to BERT embeddings.

    Patch embeddings are equivalent to ViT embeddings.
    """

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

        # text embeddings
        self.text_embeddings = TextEmbeddings(config)
        # patch embeddings
        self.cls_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
        self.patch_embeddings = ViltPatchEmbeddings(config)
        num_patches = self.patch_embeddings.num_patches
        self.position_embeddings = nn.Parameter(torch.zeros(1, num_patches + 1, config.hidden_size))
        # modality type (text/patch) embeddings
        self.token_type_embeddings = nn.Embedding(config.modality_type_vocab_size, config.hidden_size)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.config = config

    def visual_embed(self, pixel_values, pixel_mask, max_image_length=200):
        _, _, ph, pw = self.patch_embeddings.projection.weight.shape

        x = self.patch_embeddings(pixel_values)
        x_mask = pixel_mask[:, None, :, :].float()
        x_mask = nn.functional.interpolate(x_mask, size=(x.shape[2], x.shape[3])).long()
        x_h = x_mask[:, 0].sum(dim=1)[:, 0]
        x_w = x_mask[:, 0].sum(dim=2)[:, 0]

        batch_size, num_channels, height, width = x.shape
        patch_dim = self.config.image_size // self.config.patch_size
        spatial_pos = self.position_embeddings[:, 1:, :].transpose(1, 2).view(1, num_channels, patch_dim, patch_dim)
        pos_embed = torch.cat(
            [
                nn.functional.pad(
                    nn.functional.interpolate(
                        spatial_pos,
                        size=(h, w),
                        mode="bilinear",
                        align_corners=True,
                    ),
                    (0, width - w, 0, height - h),
                )
                for h, w in zip(x_h, x_w)
            ],
            dim=0,
        )

        pos_embed = pos_embed.flatten(2).transpose(1, 2)
        x = x.flatten(2).transpose(1, 2)
        # Set `device` here, otherwise `patch_index` will always be on `CPU` and will fail near the end for torch>=1.13
        patch_index = torch.stack(
            torch.meshgrid(torch.arange(x_mask.shape[-2]), torch.arange(x_mask.shape[-1]), indexing="ij"), dim=-1
        ).to(device=x_mask.device)
        patch_index = patch_index[None, None, :, :, :]
        patch_index = patch_index.expand(x_mask.shape[0], x_mask.shape[1], -1, -1, -1)
        patch_index = patch_index.flatten(1, 3)
        x_mask = x_mask.flatten(1)

        if max_image_length < 0 or max_image_length is None or not isinstance(max_image_length, int):
            # suppose aug is 800 x 1333, then, maximum effective res is 800 x 1333 (if one side gets bigger, the other will be constrained and be shrunk)
            # (800 // self.patch_size) * (1333 // self.patch_size) is the maximum number of patches that single image can get.
            # if self.patch_size = 32, 25 * 41 = 1025
            # if res is 384 x 640, 12 * 20 = 240
            effective_resolution = x_h * x_w
            max_image_length = effective_resolution.max()
        else:
            effective_resolution = x_h * x_w
            max_image_length = min(effective_resolution.max(), max_image_length)

        valid_idx = x_mask.nonzero(as_tuple=False)
        non_valid_idx = (1 - x_mask).nonzero(as_tuple=False)
        unique_rows = valid_idx[:, 0].unique()
        valid_row_idx = [valid_idx[valid_idx[:, 0] == u] for u in unique_rows]
        non_valid_row_idx = [non_valid_idx[non_valid_idx[:, 0] == u] for u in unique_rows]

        valid_nums = [v.size(0) for v in valid_row_idx]
        non_valid_nums = [v.size(0) for v in non_valid_row_idx]
        pad_nums = [max_image_length - v for v in valid_nums]

        select = []
        for i, (v, nv, p) in enumerate(zip(valid_nums, non_valid_nums, pad_nums)):
            if p <= 0:
                valid_choice = torch.multinomial(torch.ones(v).float(), max_image_length)
                select.append(valid_row_idx[i][valid_choice])
            else:
                pad_choice = torch.multinomial(torch.ones(nv).float(), p, replacement=True)
                select.append(torch.cat([valid_row_idx[i], non_valid_row_idx[i][pad_choice]], dim=0))

        select = torch.cat(select, dim=0)
        x = x[select[:, 0], select[:, 1]].view(batch_size, -1, num_channels)
        x_mask = x_mask[select[:, 0], select[:, 1]].view(batch_size, -1)
        # `patch_index` should be on the same device as `select`, which is ensured at definition time.
        patch_index = patch_index[select[:, 0], select[:, 1]].view(batch_size, -1, 2)
        pos_embed = pos_embed[select[:, 0], select[:, 1]].view(batch_size, -1, num_channels)

        cls_tokens = self.cls_token.expand(batch_size, -1, -1)
        x = torch.cat((cls_tokens, x), dim=1)
        pos_embed = torch.cat(
            (self.position_embeddings[:, 0, :][:, None, :].expand(batch_size, -1, -1), pos_embed), dim=1
        )
        x = x + pos_embed
        x = self.dropout(x)

        x_mask = torch.cat([torch.ones(x_mask.shape[0], 1).to(x_mask), x_mask], dim=1)

        return x, x_mask, (patch_index, (height, width))

    def forward(
        self,
        input_ids,
        attention_mask,
        token_type_ids,
        pixel_values,
        pixel_mask,
        inputs_embeds,
        image_embeds,
        image_token_type_idx=1,
    ):
        # PART 1: text embeddings
        text_embeds = self.text_embeddings(
            input_ids=input_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
        )

        # PART 2: patch embeddings (with interpolated position encodings)
        if image_embeds is None:
            image_embeds, image_masks, patch_index = self.visual_embed(
                pixel_values, pixel_mask, max_image_length=self.config.max_image_length
            )
        else:
            image_masks = pixel_mask.flatten(1)

        # PART 3: add modality type embeddings
        # 0 indicates text, 1 indicates image, 2 is optionally used when a second image is provided (NLVR2)
        if image_token_type_idx is None:
            image_token_type_idx = 1
        text_embeds = text_embeds + self.token_type_embeddings(
            torch.zeros_like(attention_mask, dtype=torch.long, device=text_embeds.device)
        )
        image_embeds = image_embeds + self.token_type_embeddings(
            torch.full_like(image_masks, image_token_type_idx, dtype=torch.long, device=text_embeds.device)
        )

        # PART 4: concatenate
        embeddings = torch.cat([text_embeds, image_embeds], dim=1)
        masks = torch.cat([attention_mask, image_masks], dim=1)

        return embeddings, masks


class TextEmbeddings(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.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)

        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
        self.register_buffer(
            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
        )
        self.register_buffer(
            "token_type_ids", torch.zeros(self.position_ids.size(), dtype=torch.long), persistent=False
        )

    def forward(self, input_ids=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]

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

        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
        # issue #5664
        if token_type_ids is None:
            if hasattr(self, "token_type_ids"):
                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
                token_type_ids = buffered_token_type_ids_expanded
            else:
                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)

        if inputs_embeds is None:
            inputs_embeds = self.word_embeddings(input_ids)
        token_type_embeddings = self.token_type_embeddings(token_type_ids)
        embeddings = inputs_embeds + token_type_embeddings

        position_embeddings = self.position_embeddings(position_ids)
        embeddings += position_embeddings

        embeddings = self.LayerNorm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings


class ViltPatchEmbeddings(nn.Module):
    """
    Image to Patch Embedding.
    """

    def __init__(self, config):
        super().__init__()
        image_size, patch_size = config.image_size, config.patch_size
        num_channels, hidden_size = config.num_channels, config.hidden_size

        image_size = image_size if isinstance(image_size, collections.abc.Iterable) else (image_size, image_size)
        patch_size = patch_size if isinstance(patch_size, collections.abc.Iterable) else (patch_size, patch_size)
        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
        self.image_size = image_size
        self.patch_size = patch_size
        self.num_channels = num_channels
        self.num_patches = num_patches

        self.projection = nn.Conv2d(num_channels, hidden_size, kernel_size=patch_size, stride=patch_size)

    def forward(self, pixel_values):
        batch_size, num_channels, height, width = pixel_values.shape
        if num_channels != self.num_channels:
            raise ValueError(
                "Make sure that the channel dimension of the pixel values match with the one set in the configuration."
            )
        target_dtype = self.projection.weight.dtype
        x = self.projection(pixel_values.to(dtype=target_dtype))
        return x


class ViltSelfAttention(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.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, bias=config.qkv_bias)
        self.key = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
        self.value = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)

        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)

    def forward(self, hidden_states, attention_mask=None, output_attentions=False):
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.attention_head_size)
        query_layer = self.query(hidden_states).view(hidden_shape).transpose(1, 2)
        key_layer = self.key(hidden_states).view(hidden_shape).transpose(1, 2)
        value_layer = self.value(hidden_states).view(hidden_shape).transpose(1, 2)

        # Take the dot product between "query" and "key" to get the raw attention scores.
        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
        if attention_mask is not None:
            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
            attention_scores = attention_scores + attention_mask

        # Normalize the attention scores to probabilities.
        attention_probs = nn.Softmax(dim=-1)(attention_scores)

        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.
        attention_probs = self.dropout(attention_probs)

        context_layer = torch.matmul(attention_probs, value_layer)

        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(*new_context_layer_shape)

        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)

        return outputs


# Copied from transformers.models.vit.modeling_vit.ViTSelfOutput with ViT->Vilt
class ViltSelfOutput(nn.Module):
    """
    The residual connection is defined in ViltLayer instead of here (as is the case with other models), due to the
    layernorm applied before each block.
    """

    def __init__(self, config: ViltConfig):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        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)
        return hidden_states


class ViltAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.attention = ViltSelfAttention(config)
        self.output = ViltSelfOutput(config)

    def forward(self, hidden_states, attention_mask=None, output_attentions=False):
        self_outputs = self.attention(hidden_states, attention_mask, output_attentions)

        attention_output = self.output(self_outputs[0], hidden_states)

        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
        return outputs


# Copied from transformers.models.vit.modeling_vit.ViTIntermediate with ViT->Vilt
class ViltIntermediate(nn.Module):
    def __init__(self, config: ViltConfig):
        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.vit.modeling_vit.ViTOutput with ViT->Vilt
class ViltOutput(nn.Module):
    def __init__(self, config: ViltConfig):
        super().__init__()
        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
        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 = hidden_states + input_tensor
        return hidden_states


class ViltLayer(GradientCheckpointingLayer):
    """This corresponds to the Block class in the timm implementation."""

    def __init__(self, config):
        super().__init__()
        self.chunk_size_feed_forward = config.chunk_size_feed_forward
        self.seq_len_dim = 1
        self.attention = ViltAttention(config)
        self.intermediate = ViltIntermediate(config)
        self.output = ViltOutput(config)
        self.layernorm_before = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.layernorm_after = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)

    def forward(self, hidden_states, attention_mask=None, output_attentions=False):
        self_attention_outputs = self.attention(
            self.layernorm_before(hidden_states),  # in ViLT, layernorm is applied before self-attention
            attention_mask,
            output_attentions=output_attentions,
        )
        attention_output = self_attention_outputs[0]
        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights

        # first residual connection
        hidden_states = attention_output + hidden_states.to(attention_output.device)

        # in ViLT, layernorm is also applied after self-attention
        layer_output = self.layernorm_after(hidden_states)
        layer_output = self.intermediate(layer_output)

        # second residual connection is done here
        layer_output = self.output(layer_output, hidden_states)

        outputs = (layer_output,) + outputs

        return outputs


class ViltEncoder(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.layer = nn.ModuleList([ViltLayer(config) for _ in range(config.num_hidden_layers)])
        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        output_attentions=False,
        output_hidden_states=False,
        return_dict=True,
    ):
        all_hidden_states = () if output_hidden_states else None
        all_self_attentions = () if output_attentions else None

        for i, layer_module in enumerate(self.layer):
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            layer_outputs = layer_module(hidden_states, attention_mask, output_attentions)

            hidden_states = layer_outputs[0]

            if output_attentions:
                all_self_attentions = all_self_attentions + (layer_outputs[1],)

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
        return BaseModelOutput(
            last_hidden_state=hidden_states,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
        )


@auto_docstring
class ViltPreTrainedModel(PreTrainedModel):
    config: ViltConfig
    base_model_prefix = "vilt"
    input_modalities = ("image", "text")
    supports_gradient_checkpointing = True
    _no_split_modules = ["ViltEmbeddings", "ViltSelfAttention"]

    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, TextEmbeddings):
            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
            init.zeros_(module.token_type_ids)


@auto_docstring
class ViltModel(ViltPreTrainedModel):
    def __init__(self, config, add_pooling_layer=True):
        r"""
        add_pooling_layer (bool, *optional*, defaults to `True`):
            Whether to add a pooling layer
        """
        super().__init__(config)
        self.config = config

        self.embeddings = ViltEmbeddings(config)
        self.encoder = ViltEncoder(config)

        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.pooler = ViltPooler(config) if add_pooling_layer else None

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

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

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

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        pixel_mask: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        image_embeds: torch.FloatTensor | None = None,
        image_token_type_idx: int | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> BaseModelOutputWithPooling | tuple[torch.FloatTensor]:
        r"""
        image_embeds (`torch.FloatTensor` of shape `(batch_size, num_patches, hidden_size)`, *optional*):
            Optionally, instead of passing `pixel_values`, you can choose to directly pass an embedded representation.
            This is useful if you want more control over how to convert `pixel_values` into patch embeddings.
        image_token_type_idx (`int`, *optional*):
            - The token type ids for images.

        Examples:

        ```python
        >>> from transformers import ViltProcessor, ViltModel
        >>> from PIL import Image
        >>> import httpx
        >>> from io import BytesIO

        >>> # prepare image and text
        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read()))
        >>> text = "hello world"

        >>> processor = ViltProcessor.from_pretrained("dandelin/vilt-b32-mlm")
        >>> model = ViltModel.from_pretrained("dandelin/vilt-b32-mlm")

        >>> inputs = processor(image, text, return_tensors="pt")
        >>> outputs = model(**inputs)
        >>> last_hidden_states = outputs.last_hidden_state
        ```"""
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.return_dict

        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")

        text_batch_size, seq_length = input_shape
        device = input_ids.device if input_ids is not None else inputs_embeds.device

        if attention_mask is None:
            attention_mask = torch.ones(((text_batch_size, seq_length)), device=device)

        if pixel_values is not None and image_embeds is not None:
            raise ValueError("You cannot specify both pixel_values and image_embeds at the same time")
        elif pixel_values is None and image_embeds is None:
            raise ValueError("You have to specify either pixel_values or image_embeds")

        image_batch_size = pixel_values.shape[0] if pixel_values is not None else image_embeds.shape[0]
        if image_batch_size != text_batch_size:
            raise ValueError("The text inputs and image inputs need to have the same batch size")
        if pixel_mask is None:
            pixel_mask = torch.ones((image_batch_size, self.config.image_size, self.config.image_size), device=device)

        embedding_output, attention_mask = self.embeddings(
            input_ids,
            attention_mask,
            token_type_ids,
            pixel_values,
            pixel_mask,
            inputs_embeds,
            image_embeds,
            image_token_type_idx=image_token_type_idx,
        )

        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
        # ourselves in which case we just need to make it broadcastable to all heads.
        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape)

        encoder_outputs = self.encoder(
            embedding_output,
            attention_mask=extended_attention_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        sequence_output = encoder_outputs[0]
        sequence_output = self.layernorm(sequence_output)
        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None

        if not return_dict:
            return (sequence_output, pooled_output) + encoder_outputs[1:]

        return BaseModelOutputWithPooling(
            last_hidden_state=sequence_output,
            pooler_output=pooled_output,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
        )


class ViltPooler(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):
        # 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


@auto_docstring(
    custom_intro="""
    ViLT Model with a language modeling head on top as done during pretraining.
    """
)
class ViltForMaskedLM(ViltPreTrainedModel):
    _tied_weights_keys = {
        "mlm_score.decoder.weight": "vilt.embeddings.text_embeddings.word_embeddings.weight",
    }

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

        self.vilt = ViltModel(config)
        self.mlm_score = ViltMLMHead(config)

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

    def get_output_embeddings(self):
        return self.mlm_score.decoder

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

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        pixel_mask: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        image_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> MaskedLMOutput | tuple[torch.FloatTensor]:
        r"""
        image_embeds (`torch.FloatTensor` of shape `(batch_size, num_patches, hidden_size)`, *optional*):
            Optionally, instead of passing `pixel_values`, you can choose to directly pass an embedded representation.
            This is useful if you want more control over how to convert `pixel_values` into patch embeddings.
        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 ViltProcessor, ViltForMaskedLM
        >>> import httpx
        >>> from io import BytesIO
        >>> from PIL import Image
        >>> import re
        >>> import torch

        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read()))
        >>> text = "a bunch of [MASK] laying on a [MASK]."

        >>> processor = ViltProcessor.from_pretrained("dandelin/vilt-b32-mlm")
        >>> model = ViltForMaskedLM.from_pretrained("dandelin/vilt-b32-mlm")

        >>> # prepare inputs
        >>> encoding = processor(image, text, return_tensors="pt")

        >>> # forward pass
        >>> outputs = model(**encoding)

        >>> tl = len(re.findall("\[MASK\]", text))
        >>> inferred_token = [text]

        >>> # gradually fill in the MASK tokens, one by one
        >>> with torch.no_grad():
        ...     for i in range(tl):
        ...         encoded = processor.tokenizer(inferred_token)
        ...         input_ids = torch.tensor(encoded.input_ids)
        ...         encoded = encoded["input_ids"][0][1:-1]
        ...         outputs = model(input_ids=input_ids, pixel_values=encoding.pixel_values)
        ...         mlm_logits = outputs.logits[0]  # shape (seq_len, vocab_size)
        ...         # only take into account text features (minus CLS and SEP token)
        ...         mlm_logits = mlm_logits[1 : input_ids.shape[1] - 1, :]
        ...         mlm_values, mlm_ids = mlm_logits.softmax(dim=-1).max(dim=-1)
        ...         # only take into account text
        ...         mlm_values[torch.tensor(encoded) != 103] = 0
        ...         select = mlm_values.argmax().item()
        ...         encoded[select] = mlm_ids[select].item()
        ...         inferred_token = [processor.decode(encoded)]

        >>> selected_token = ""
        >>> encoded = processor.tokenizer(inferred_token)
        >>> output = processor.decode(encoded.input_ids[0], skip_special_tokens=True)
        >>> print(output)
        a bunch of cats laying on a couch.
        ```"""
        return_dict = return_dict if return_dict is not None else self.config.return_dict

        outputs = self.vilt(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            pixel_values=pixel_values,
            pixel_mask=pixel_mask,
            inputs_embeds=inputs_embeds,
            image_embeds=image_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output, pooled_output = outputs[:2]
        # split up final hidden states into text and image features
        text_seq_len = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
        text_features, _ = (sequence_output[:, :text_seq_len], sequence_output[:, text_seq_len:])

        mlm_logits = self.mlm_score(text_features)

        masked_lm_loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()  # -100 index = padding token
            # move labels to correct device to enable PP
            labels = labels.to(mlm_logits.device)
            masked_lm_loss = loss_fct(mlm_logits.view(-1, self.config.vocab_size), labels.view(-1))

        if not return_dict:
            output = (mlm_logits,) + outputs[2:]
            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output

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


class ViltPredictionHeadTransform(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):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.transform_act_fn(hidden_states)
        hidden_states = self.LayerNorm(hidden_states)
        return hidden_states


class ViltMLMHead(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.transform = ViltPredictionHeadTransform(config)
        self.decoder = nn.Linear(config.hidden_size, config.vocab_size)

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


@auto_docstring(
    custom_intro="""
    Vilt Model transformer with a classifier head on top (a linear layer on top of the final hidden state of the [CLS]
    token) for visual question answering, e.g. for VQAv2.
    """
)
class ViltForQuestionAnswering(ViltPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.num_labels = config.num_labels
        self.vilt = ViltModel(config)

        # Classifier head
        self.classifier = nn.Sequential(
            nn.Linear(config.hidden_size, config.hidden_size * 2),
            nn.LayerNorm(config.hidden_size * 2),
            nn.GELU(),
            nn.Linear(config.hidden_size * 2, config.num_labels),
        )

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

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        pixel_mask: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        image_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> SequenceClassifierOutput | tuple[torch.FloatTensor]:
        r"""
        image_embeds (`torch.FloatTensor` of shape `(batch_size, num_patches, hidden_size)`, *optional*):
            Optionally, instead of passing `pixel_values`, you can choose to directly pass an embedded representation.
            This is useful if you want more control over how to convert `pixel_values` into patch embeddings.
        labels (`torch.FloatTensor` of shape `(batch_size, num_labels)`, *optional*):
            Labels for computing the visual question answering loss. This tensor must be either a one-hot encoding of
            all answers that are applicable for a given example in the batch, or a soft encoding indicating which
            answers are applicable, where 1.0 is the highest score.

        Examples:

        ```python
        >>> from transformers import ViltProcessor, ViltForQuestionAnswering
        >>> import httpx
        >>> from io import BytesIO
        >>> from PIL import Image

        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read()))
        >>> text = "How many cats are there?"

        >>> processor = ViltProcessor.from_pretrained("dandelin/vilt-b32-finetuned-vqa")
        >>> model = ViltForQuestionAnswering.from_pretrained("dandelin/vilt-b32-finetuned-vqa")

        >>> # prepare inputs
        >>> encoding = processor(image, text, return_tensors="pt")

        >>> # forward pass
        >>> outputs = model(**encoding)
        >>> logits = outputs.logits
        >>> idx = logits.argmax(-1).item()
        >>> print("Predicted answer:", model.config.id2label[idx])
        Predicted answer: 2
        ```"""
        return_dict = return_dict if return_dict is not None else self.config.return_dict

        outputs = self.vilt(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            pixel_values=pixel_values,
            pixel_mask=pixel_mask,
            inputs_embeds=inputs_embeds,
            image_embeds=image_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        pooler_output = outputs.pooler_output if return_dict else outputs[1]

        logits = self.classifier(pooler_output)

        loss = None
        if labels is not None:
            # move labels to correct device to enable PP
            labels = labels.to(logits.device)
            loss = nn.functional.binary_cross_entropy_with_logits(logits, labels) * labels.shape[1]
            # see https://github.com/jnhwkim/ban-vqa/blob/master/train.py#L19

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

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


@auto_docstring(
    custom_intro="""
    Vilt Model transformer with a classifier head on top (a linear layer on top of the final hidden state of the [CLS]
    token) for image-to-text or text-to-image retrieval, e.g. MSCOCO and F30K.
    """
)
class ViltForImageAndTextRetrieval(ViltPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.vilt = ViltModel(config)

        # Classifier head
        self.rank_output = nn.Linear(config.hidden_size, 1)

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

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        pixel_mask: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        image_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> SequenceClassifierOutput | tuple[torch.FloatTensor]:
        r"""
        image_embeds (`torch.FloatTensor` of shape `(batch_size, num_patches, hidden_size)`, *optional*):
            Optionally, instead of passing `pixel_values`, you can choose to directly pass an embedded representation.
            This is useful if you want more control over how to convert `pixel_values` into patch embeddings.
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels are currently not supported.

        Examples:

        ```python
        >>> from transformers import ViltProcessor, ViltForImageAndTextRetrieval
        >>> import httpx
        >>> from io import BytesIO
        >>> from PIL import Image

        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read()))
        >>> texts = ["An image of two cats chilling on a couch", "A football player scoring a goal"]

        >>> processor = ViltProcessor.from_pretrained("dandelin/vilt-b32-finetuned-coco")
        >>> model = ViltForImageAndTextRetrieval.from_pretrained("dandelin/vilt-b32-finetuned-coco")

        >>> # forward pass
        >>> scores = dict()
        >>> for text in texts:
        ...     # prepare inputs
        ...     encoding = processor(image, text, return_tensors="pt")
        ...     outputs = model(**encoding)
        ...     scores[text] = outputs.logits[0, :].item()
        ```"""
        return_dict = return_dict if return_dict is not None else self.config.return_dict

        loss = None
        if labels is not None:
            raise NotImplementedError("Training is not yet supported.")

        outputs = self.vilt(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            pixel_values=pixel_values,
            pixel_mask=pixel_mask,
            inputs_embeds=inputs_embeds,
            image_embeds=image_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        pooler_output = outputs.pooler_output if return_dict else outputs[1]

        logits = self.rank_output(pooler_output)

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

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


@auto_docstring(
    custom_intro="""
    Vilt Model transformer with a classifier head on top for natural language visual reasoning, e.g. NLVR2.
    """
)
class ViltForImagesAndTextClassification(ViltPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.num_labels = config.num_labels
        self.vilt = ViltModel(config)

        # Classifier head
        num_images = config.num_images
        self.classifier = nn.Sequential(
            nn.Linear(config.hidden_size * num_images, config.hidden_size * num_images),
            nn.LayerNorm(config.hidden_size * num_images),
            nn.GELU(),
            nn.Linear(config.hidden_size * num_images, config.num_labels),
        )

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

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        pixel_mask: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        image_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> ViltForImagesAndTextClassificationOutput | tuple[torch.FloatTensor]:
        r"""
        image_embeds (`torch.FloatTensor` of shape `(batch_size, num_patches, hidden_size)`, *optional*):
            Optionally, instead of passing `pixel_values`, you can choose to directly pass an embedded representation.
            This is useful if you want more control over how to convert `pixel_values` into patch embeddings.
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Binary classification labels.

        Examples:

        ```python
        >>> from transformers import ViltProcessor, ViltForImagesAndTextClassification
        >>> import httpx
        >>> from io import BytesIO
        >>> from PIL import Image

        >>> url_1 = "https://lil.nlp.cornell.edu/nlvr/exs/ex0_0.jpg"
        >>> with httpx.stream("GET", url_1) as response:
        ...     image_1 = Image.open(BytesIO(response.read()))

        >>> url_2 = "https://lil.nlp.cornell.edu/nlvr/exs/ex0_1.jpg"
        >>> with httpx.stream("GET", url_2) as response:
        ...     image_2 = Image.open(BytesIO(response.read()))

        >>> text = "The left image contains twice the number of dogs as the right image."

        >>> processor = ViltProcessor.from_pretrained("dandelin/vilt-b32-finetuned-nlvr2")
        >>> model = ViltForImagesAndTextClassification.from_pretrained("dandelin/vilt-b32-finetuned-nlvr2")

        >>> # prepare inputs
        >>> encoding = processor([image_1, image_2], text, return_tensors="pt")

        >>> # forward pass
        >>> outputs = model(input_ids=encoding.input_ids, pixel_values=encoding.pixel_values.unsqueeze(0))
        >>> logits = outputs.logits
        >>> idx = logits.argmax(-1).item()
        >>> print("Predicted answer:", model.config.id2label[idx])
        Predicted answer: True
        ```"""
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.return_dict

        if pixel_values is not None and pixel_values.ndim == 4:
            # add dummy num_images dimension
            pixel_values = pixel_values.unsqueeze(1)

        if image_embeds is not None and image_embeds.ndim == 3:
            # add dummy num_images dimension
            image_embeds = image_embeds.unsqueeze(1)

        num_images = pixel_values.shape[1] if pixel_values is not None else None
        if num_images is None:
            num_images = image_embeds.shape[1] if image_embeds is not None else None
        if num_images != self.config.num_images:
            raise ValueError(
                "Make sure to match the number of images in the model with the number of images in the input."
            )
        pooler_outputs = []
        hidden_states = [] if output_hidden_states else None
        attentions = [] if output_attentions else None
        for i in range(num_images):
            # forward every image through the model
            outputs = self.vilt(
                input_ids,
                attention_mask=attention_mask,
                token_type_ids=token_type_ids,
                pixel_values=pixel_values[:, i, :, :, :] if pixel_values is not None else None,
                pixel_mask=pixel_mask[:, i, :, :] if pixel_mask is not None else None,
                inputs_embeds=inputs_embeds,
                image_embeds=image_embeds[:, i, :, :] if image_embeds is not None else None,
                image_token_type_idx=i + 1,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )
            pooler_output = outputs.pooler_output if return_dict else outputs[1]
            pooler_outputs.append(pooler_output)
            if output_hidden_states:
                hidden_states.append(outputs.hidden_states)
            if output_attentions:
                attentions.append(outputs.attentions)

        pooled_output = torch.cat(pooler_outputs, dim=-1)
        logits = self.classifier(pooled_output)

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            # move labels to correct device to enable PP
            labels = labels.to(logits.device)
            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

        if not return_dict:
            output = (logits, hidden_states, attentions)
            return ((loss,) + output) if loss is not None else output

        return ViltForImagesAndTextClassificationOutput(
            loss=loss,
            logits=logits,
            hidden_states=hidden_states,
            attentions=attentions,
        )


@auto_docstring
class ViltForTokenClassification(ViltPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.num_labels = config.num_labels
        self.vilt = ViltModel(config, add_pooling_layer=False)

        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()

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        pixel_mask: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        image_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> TokenClassifierOutput | tuple[torch.FloatTensor]:
        r"""
        image_embeds (`torch.FloatTensor` of shape `(batch_size, num_patches, hidden_size)`, *optional*):
            Optionally, instead of passing `pixel_values`, you can choose to directly pass an embedded representation.
            This is useful if you want more control over how to convert `pixel_values` into patch embeddings.
        labels (`torch.LongTensor` of shape `(batch_size, text_sequence_length)`, *optional*):
            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
        """

        return_dict = return_dict if return_dict is not None else self.config.return_dict

        outputs = self.vilt(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            pixel_values=pixel_values,
            pixel_mask=pixel_mask,
            inputs_embeds=inputs_embeds,
            image_embeds=image_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output = outputs[0]

        text_input_size = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]

        sequence_output = self.dropout(sequence_output)
        logits = self.classifier(sequence_output[:, :text_input_size])

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            # move labels to correct device to enable PP
            labels = labels.to(logits.device)
            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

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


__all__ = [
    "ViltForImageAndTextRetrieval",
    "ViltForImagesAndTextClassification",
    "ViltForTokenClassification",
    "ViltForMaskedLM",
    "ViltForQuestionAnswering",
    "ViltLayer",
    "ViltModel",
    "ViltPreTrainedModel",
]