transformers/src/transformers/models/siglip/modeling_siglip.py

# coding=utf-8
# Copyright 2024 Google AI and The HuggingFace 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 Siglip model."""

import math
import warnings
from dataclasses import dataclass
from typing import Any, Callable, Optional, Tuple, Union

import numpy as np
import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from torch.nn.init import _calculate_fan_in_and_fan_out

from ...activations import ACT2FN
from ...modeling_attn_mask_utils import _prepare_4d_attention_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, ImageClassifierOutput
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...utils import ModelOutput, auto_docstring, can_return_tuple, logging, torch_int
from .configuration_siglip import SiglipConfig, SiglipTextConfig, SiglipVisionConfig


logger = logging.get_logger(__name__)


def _trunc_normal_(tensor, mean, std, a, b):
    # Cut & paste from PyTorch official master until it's in a few official releases - RW
    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
    def norm_cdf(x):
        # Computes standard normal cumulative distribution function
        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0

    if (mean < a - 2 * std) or (mean > b + 2 * std):
        warnings.warn(
            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
            "The distribution of values may be incorrect.",
            stacklevel=2,
        )

    # Values are generated by using a truncated uniform distribution and
    # then using the inverse CDF for the normal distribution.
    # Get upper and lower cdf values
    l = norm_cdf((a - mean) / std)
    u = norm_cdf((b - mean) / std)

    # Uniformly fill tensor with values from [l, u], then translate to
    # [2l-1, 2u-1].
    tensor.uniform_(2 * l - 1, 2 * u - 1)

    # Use inverse cdf transform for normal distribution to get truncated
    # standard normal
    tensor.erfinv_()

    # Transform to proper mean, std
    tensor.mul_(std * math.sqrt(2.0))
    tensor.add_(mean)

    # Clamp to ensure it's in the proper range
    tensor.clamp_(min=a, max=b)


def trunc_normal_tf_(
    tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
) -> torch.Tensor:
    """Fills the input Tensor with values drawn from a truncated
    normal distribution. The values are effectively drawn from the
    normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)`
    with values outside :math:`[a, b]` redrawn until they are within
    the bounds. The method used for generating the random values works
    best when :math:`a \\leq \text{mean} \\leq b`.

    NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
    bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
    and the result is subsequently scaled and shifted by the mean and std args.

    Args:
        tensor: an n-dimensional `torch.Tensor`
        mean: the mean of the normal distribution
        std: the standard deviation of the normal distribution
        a: the minimum cutoff value
        b: the maximum cutoff value
    """
    with torch.no_grad():
        _trunc_normal_(tensor, 0, 1.0, a, b)
        tensor.mul_(std).add_(mean)


def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
    if mode == "fan_in":
        denom = fan_in
    elif mode == "fan_out":
        denom = fan_out
    elif mode == "fan_avg":
        denom = (fan_in + fan_out) / 2

    variance = scale / denom

    if distribution == "truncated_normal":
        # constant is stddev of standard normal truncated to (-2, 2)
        trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
    elif distribution == "normal":
        with torch.no_grad():
            tensor.normal_(std=math.sqrt(variance))
    elif distribution == "uniform":
        bound = math.sqrt(3 * variance)
        with torch.no_grad():
            tensor.uniform_(-bound, bound)
    else:
        raise ValueError(f"invalid distribution {distribution}")


def lecun_normal_(tensor):
    variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")


def default_flax_embed_init(tensor):
    variance_scaling_(tensor, mode="fan_in", distribution="normal")


@dataclass
# Copied from transformers.models.clip.modeling_clip.CLIPVisionModelOutput with CLIP->Siglip
class SiglipVisionModelOutput(ModelOutput):
    """
    Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.

    Args:
        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
            The image embeddings obtained by applying the projection layer to the pooler_output.
        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
            Sequence of hidden-states at the output of the last layer of the model.
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
    """

    image_embeds: Optional[torch.FloatTensor] = None
    last_hidden_state: Optional[torch.FloatTensor] = None
    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None


@dataclass
# Copied from transformers.models.clip.modeling_clip.CLIPTextModelOutput with CLIP->Siglip
class SiglipTextModelOutput(ModelOutput):
    """
    Base class for text model's outputs that also contains a pooling of the last hidden states.

    Args:
        text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
            The text embeddings obtained by applying the projection layer to the pooler_output.
        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
            Sequence of hidden-states at the output of the last layer of the model.
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
    """

    text_embeds: Optional[torch.FloatTensor] = None
    last_hidden_state: Optional[torch.FloatTensor] = None
    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None


@dataclass
# Copied from transformers.models.clip.modeling_clip.CLIPOutput with CLIP->Siglip
class SiglipOutput(ModelOutput):
    """
    Args:
        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
            Contrastive loss for image-text similarity.
        logits_per_image (`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`):
            The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
            similarity scores.
        logits_per_text (`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`):
            The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
            similarity scores.
        text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
            The text embeddings obtained by applying the projection layer to the pooled output of [`SiglipTextModel`].
        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
            The image embeddings obtained by applying the projection layer to the pooled output of [`SiglipVisionModel`].
        text_model_output (`BaseModelOutputWithPooling`):
            The output of the [`SiglipTextModel`].
        vision_model_output (`BaseModelOutputWithPooling`):
            The output of the [`SiglipVisionModel`].
    """

    loss: Optional[torch.FloatTensor] = None
    logits_per_image: Optional[torch.FloatTensor] = None
    logits_per_text: Optional[torch.FloatTensor] = None
    text_embeds: Optional[torch.FloatTensor] = None
    image_embeds: Optional[torch.FloatTensor] = None
    text_model_output: BaseModelOutputWithPooling = None
    vision_model_output: BaseModelOutputWithPooling = None

    def to_tuple(self) -> Tuple[Any]:
        return tuple(
            self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
            for k in self.keys()
        )


class SiglipVisionEmbeddings(nn.Module):
    def __init__(self, config: SiglipVisionConfig):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.image_size = config.image_size
        self.patch_size = config.patch_size

        self.patch_embedding = nn.Conv2d(
            in_channels=config.num_channels,
            out_channels=self.embed_dim,
            kernel_size=self.patch_size,
            stride=self.patch_size,
            padding="valid",
        )

        self.num_patches = (self.image_size // self.patch_size) ** 2
        self.num_positions = self.num_patches
        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
        self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)

    def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width: int) -> torch.Tensor:
        """
        This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution
        images. This method is also adapted to support torch.jit tracing and no class embeddings.

        Adapted from:
        - https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174-L194, and
        - https://github.com/facebookresearch/dinov2/blob/e1277af2ba9496fbadf7aec6eba56e8d882d1e35/dinov2/models/vision_transformer.py#L179-L211
        """

        num_patches = embeddings.shape[1]
        num_positions = self.position_embedding.weight.shape[0]

        # always interpolate when tracing to ensure the exported model works for dynamic input shapes
        if not torch.jit.is_tracing() and num_patches == num_positions and height == width:
            return self.position_embedding(self.position_ids)

        patch_pos_embed = self.position_embedding.weight.unsqueeze(0)

        dim = embeddings.shape[-1]

        new_height = height // self.patch_size
        new_width = width // self.patch_size

        sqrt_num_positions = torch_int(num_positions**0.5)
        patch_pos_embed = patch_pos_embed.reshape(1, sqrt_num_positions, sqrt_num_positions, dim)
        patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)

        patch_pos_embed = nn.functional.interpolate(
            patch_pos_embed,
            size=(new_height, new_width),
            mode="bicubic",
            align_corners=False,
        )

        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
        return patch_pos_embed

    def forward(self, pixel_values: torch.FloatTensor, interpolate_pos_encoding=False) -> torch.Tensor:
        _, _, height, width = pixel_values.shape
        target_dtype = self.patch_embedding.weight.dtype
        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]
        embeddings = patch_embeds.flatten(2).transpose(1, 2)

        if interpolate_pos_encoding:
            embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width)
        else:
            embeddings = embeddings + self.position_embedding(self.position_ids)
        return embeddings


# Copied from transformers.models.clip.modeling_clip.CLIPTextEmbeddings with CLIP->Siglip
class SiglipTextEmbeddings(nn.Module):
    def __init__(self, config: SiglipTextConfig):
        super().__init__()
        embed_dim = config.hidden_size

        self.token_embedding = nn.Embedding(config.vocab_size, embed_dim)
        self.position_embedding = nn.Embedding(config.max_position_embeddings, embed_dim)

        # 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
        )

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
    ) -> torch.Tensor:
        seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2]
        max_position_embedding = self.position_embedding.weight.shape[0]

        if seq_length > max_position_embedding:
            raise ValueError(
                f"Sequence length must be less than max_position_embeddings (got `sequence length`: "
                f"{seq_length} and max_position_embeddings: {max_position_embedding}"
            )

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

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

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

        return embeddings


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    scaling: float,
    dropout: float = 0.0,
    **kwargs,
):
    attn_weights = torch.matmul(query, key.transpose(-1, -2)) * 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


class SiglipAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.embed_dim // self.num_heads
        if self.head_dim * self.num_heads != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
                f" {self.num_heads})."
            )
        self.scale = self.head_dim**-0.5
        self.dropout = config.attention_dropout
        self.is_causal = False

        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
        """Input shape: Batch x Time x Channel"""

        batch_size, seq_length, embed_dim = hidden_states.shape

        queries = self.q_proj(hidden_states)
        keys = self.k_proj(hidden_states)
        values = self.v_proj(hidden_states)

        queries = queries.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
        keys = keys.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
        values = values.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            if self.config._attn_implementation == "sdpa" and output_attentions:
                logger.warning_once(
                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
                )
            else:
                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        attn_output, attn_weights = attention_interface(
            self,
            queries,
            keys,
            values,
            attention_mask,
            is_causal=self.is_causal,
            scaling=self.scale,
            dropout=0.0 if not self.training else self.dropout,
        )

        attn_output = attn_output.reshape(batch_size, seq_length, embed_dim).contiguous()
        attn_output = self.out_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights


# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Siglip
class SiglipMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.activation_fn = ACT2FN[config.hidden_act]
        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.fc1(hidden_states)
        hidden_states = self.activation_fn(hidden_states)
        hidden_states = self.fc2(hidden_states)
        return hidden_states


class SiglipEncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Union[SiglipVisionConfig, SiglipTextConfig]):
        super().__init__()
        self.embed_dim = config.hidden_size
        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
        self.self_attn = SiglipAttention(config)
        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
        self.mlp = SiglipMLP(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        output_attentions: Optional[bool] = False,
    ) -> Tuple[torch.FloatTensor]:
        """
        Args:
            hidden_states (`torch.FloatTensor`):
                Input to the layer of shape `(batch, seq_len, embed_dim)`.
            attention_mask (`torch.FloatTensor`):
                Attention mask of shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values.
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states

        hidden_states = self.layer_norm1(hidden_states)
        hidden_states, attn_weights = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            output_attentions=output_attentions,
        )
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.layer_norm2(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (attn_weights,)

        return outputs


@auto_docstring
class SiglipPreTrainedModel(PreTrainedModel):
    config_class = SiglipConfig
    base_model_prefix = "siglip"
    supports_gradient_checkpointing = True

    _no_split_modules = [
        "SiglipTextEmbeddings",
        "SiglipEncoderLayer",
        "SiglipVisionEmbeddings",
        "SiglipEncoderLayer",
        "SiglipMultiheadAttentionPoolingHead",
    ]
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _supports_attention_backend = True

    def _init_weights(self, module):
        """Initialize the weights"""
        if isinstance(module, SiglipVisionEmbeddings):
            width = (
                self.config.vision_config.hidden_size
                if isinstance(self.config, SiglipConfig)
                else self.config.hidden_size
            )
            nn.init.normal_(module.position_embedding.weight, std=1 / np.sqrt(width))
        elif isinstance(module, nn.Embedding):
            default_flax_embed_init(module.weight)
        elif isinstance(module, SiglipAttention):
            nn.init.xavier_uniform_(module.q_proj.weight)
            nn.init.xavier_uniform_(module.k_proj.weight)
            nn.init.xavier_uniform_(module.v_proj.weight)
            nn.init.xavier_uniform_(module.out_proj.weight)
            nn.init.zeros_(module.q_proj.bias)
            nn.init.zeros_(module.k_proj.bias)
            nn.init.zeros_(module.v_proj.bias)
            nn.init.zeros_(module.out_proj.bias)
        elif isinstance(module, SiglipMLP):
            nn.init.xavier_uniform_(module.fc1.weight)
            nn.init.xavier_uniform_(module.fc2.weight)
            nn.init.normal_(module.fc1.bias, std=1e-6)
            nn.init.normal_(module.fc2.bias, std=1e-6)
        elif isinstance(module, SiglipMultiheadAttentionPoolingHead):
            nn.init.xavier_uniform_(module.probe.data)
            nn.init.xavier_uniform_(module.attention.in_proj_weight.data)
            nn.init.zeros_(module.attention.in_proj_bias.data)
        elif isinstance(module, SiglipModel):
            logit_scale_init = torch.log(torch.tensor(1.0))
            module.logit_scale.data.fill_(logit_scale_init)
            module.logit_bias.data.zero_()
        elif isinstance(module, SiglipForImageClassification):
            nn.init.normal_(
                module.classifier.weight,
                std=self.config.vision_config.hidden_size**-0.5 * self.config.initializer_factor,
            )
        elif isinstance(module, (nn.Linear, nn.Conv2d)):
            lecun_normal_(module.weight)
            if module.bias is not None:
                nn.init.zeros_(module.bias)
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)


# Copied from transformers.models.altclip.modeling_altclip.AltCLIPEncoder with AltCLIP->Siglip
class SiglipEncoder(nn.Module):
    """
    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
    [`SiglipEncoderLayer`].

    Args:
        config: SiglipConfig
    """

    def __init__(self, config: SiglipConfig):
        super().__init__()
        self.config = config
        self.layers = nn.ModuleList([SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)])
        self.gradient_checkpointing = False

    # Ignore copy
    @can_return_tuple
    def forward(
        self,
        inputs_embeds,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
    ) -> BaseModelOutput:
        r"""
        Args:
            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
                than the model's internal embedding lookup matrix.
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                [What are attention masks?](../glossary#attention-mask)
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        """
        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
        )

        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None

        hidden_states = inputs_embeds
        for encoder_layer in self.layers:
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)

            layer_outputs = encoder_layer(
                hidden_states,
                attention_mask,
                output_attentions=output_attentions,
            )

            hidden_states = layer_outputs[0]

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

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

        return BaseModelOutput(
            last_hidden_state=hidden_states,
            hidden_states=encoder_states,
            attentions=all_attentions,
        )


class SiglipTextTransformer(nn.Module):
    def __init__(self, config: SiglipTextConfig):
        super().__init__()
        self.config = config
        embed_dim = config.hidden_size
        self.embeddings = SiglipTextEmbeddings(config)
        self.encoder = SiglipEncoder(config)
        self.final_layer_norm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)

        self.head = nn.Linear(embed_dim, config.projection_size)
        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
    ) -> BaseModelOutputWithPooling:
        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
        )

        if input_ids is None:
            raise ValueError("You have to specify input_ids")

        input_shape = input_ids.size()
        input_ids = input_ids.view(-1, input_shape[-1])

        hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids)

        # note: SigLIP's text model does not use a causal mask, unlike the original CLIP model.
        # expand attention_mask
        if attention_mask is not None and not self._use_flash_attention_2:
            # [batch_size, seq_len] -> [batch_size, 1, tgt_seq_len, src_seq_len]
            attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_states.dtype)

        encoder_outputs: BaseModelOutput = self.encoder(
            inputs_embeds=hidden_states,
            attention_mask=attention_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )

        last_hidden_state = encoder_outputs.last_hidden_state
        last_hidden_state = self.final_layer_norm(last_hidden_state)

        # Assuming "sticky" EOS tokenization, last token is always EOS.
        pooled_output = last_hidden_state[:, -1, :]
        pooled_output = self.head(pooled_output)

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


@auto_docstring(
    custom_intro="""
    The text model from SigLIP without any head or projection on top.
    """
)
class SiglipTextModel(SiglipPreTrainedModel):
    config_class = SiglipTextConfig

    def __init__(self, config: SiglipTextConfig):
        super().__init__(config)
        self.text_model = SiglipTextTransformer(config)
        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self) -> nn.Module:
        return self.text_model.embeddings.token_embedding

    def set_input_embeddings(self, value):
        self.text_model.embeddings.token_embedding = value

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
    ) -> BaseModelOutputWithPooling:
        r"""
        Examples:

        ```python
        >>> from transformers import AutoTokenizer, SiglipTextModel

        >>> model = SiglipTextModel.from_pretrained("google/siglip-base-patch16-224")
        >>> tokenizer = AutoTokenizer.from_pretrained("google/siglip-base-patch16-224")

        >>> # important: make sure to set padding="max_length" as that's how the model was trained
        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding="max_length", return_tensors="pt")

        >>> outputs = model(**inputs)
        >>> last_hidden_state = outputs.last_hidden_state
        >>> pooled_output = outputs.pooler_output  # pooled (EOS token) states
        ```"""

        return self.text_model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )


class SiglipVisionTransformer(nn.Module):
    def __init__(self, config: SiglipVisionConfig):
        super().__init__()
        self.config = config
        embed_dim = config.hidden_size

        self.embeddings = SiglipVisionEmbeddings(config)
        self.encoder = SiglipEncoder(config)
        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
        self.use_head = True if not hasattr(config, "vision_use_head") else config.vision_use_head
        if self.use_head:
            self.head = SiglipMultiheadAttentionPoolingHead(config)

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        pixel_values,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        interpolate_pos_encoding: Optional[bool] = False,
    ) -> BaseModelOutputWithPooling:
        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
        )

        hidden_states = self.embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding)

        encoder_outputs: BaseModelOutput = self.encoder(
            inputs_embeds=hidden_states,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )

        last_hidden_state = encoder_outputs.last_hidden_state
        last_hidden_state = self.post_layernorm(last_hidden_state)

        pooler_output = self.head(last_hidden_state) if self.use_head else None

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


class SiglipMultiheadAttentionPoolingHead(nn.Module):
    """Multihead Attention Pooling."""

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

        self.probe = nn.Parameter(torch.randn(1, 1, config.hidden_size))
        self.attention = torch.nn.MultiheadAttention(config.hidden_size, config.num_attention_heads, batch_first=True)
        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.mlp = SiglipMLP(config)

    def forward(self, hidden_state):
        batch_size = hidden_state.shape[0]
        probe = self.probe.repeat(batch_size, 1, 1)

        hidden_state = self.attention(probe, hidden_state, hidden_state)[0]

        residual = hidden_state
        hidden_state = self.layernorm(hidden_state)
        hidden_state = residual + self.mlp(hidden_state)

        return hidden_state[:, 0]


@auto_docstring(
    custom_intro="""
    The vision model from SigLIP without any head or projection on top.
    """
)
class SiglipVisionModel(SiglipPreTrainedModel):
    config_class = SiglipVisionConfig
    main_input_name = "pixel_values"

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

        self.vision_model = SiglipVisionTransformer(config)

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

    def get_input_embeddings(self) -> nn.Module:
        return self.vision_model.embeddings.patch_embedding

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        pixel_values,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        interpolate_pos_encoding: bool = False,
    ) -> BaseModelOutputWithPooling:
        r"""
        Examples:

        ```python
        >>> from PIL import Image
        >>> import requests
        >>> from transformers import AutoProcessor, SiglipVisionModel

        >>> model = SiglipVisionModel.from_pretrained("google/siglip-base-patch16-224")
        >>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")

        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        >>> image = Image.open(requests.get(url, stream=True).raw)

        >>> inputs = processor(images=image, return_tensors="pt")

        >>> outputs = model(**inputs)
        >>> last_hidden_state = outputs.last_hidden_state
        >>> pooled_output = outputs.pooler_output  # pooled features
        ```"""

        return self.vision_model(
            pixel_values=pixel_values,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            interpolate_pos_encoding=interpolate_pos_encoding,
        )


@auto_docstring
class SiglipModel(SiglipPreTrainedModel):
    config_class = SiglipConfig

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

        if not isinstance(config.text_config, SiglipTextConfig):
            raise TypeError(
                "config.text_config is expected to be of type SiglipTextConfig but is of type"
                f" {type(config.text_config)}."
            )

        if not isinstance(config.vision_config, SiglipVisionConfig):
            raise TypeError(
                "config.vision_config is expected to be of type SiglipVisionConfig but is of type"
                f" {type(config.vision_config)}."
            )

        text_config = config.text_config
        vision_config = config.vision_config

        # First, initialize the text and vision models with proper attention implementation
        text_model = SiglipTextModel._from_config(text_config)
        vision_model = SiglipVisionModel._from_config(vision_config)

        # Second, get the text and vision submodules (for backward compatibility)
        self.text_model = text_model.text_model
        self.vision_model = vision_model.vision_model

        self.logit_scale = nn.Parameter(torch.randn(1))
        self.logit_bias = nn.Parameter(torch.randn(1))

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

    @auto_docstring
    def get_text_features(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
    ) -> torch.FloatTensor:
        r"""
        Returns:
            text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
            applying the projection layer to the pooled output of [`SiglipTextModel`].

        Examples:

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

        >>> model = AutoModel.from_pretrained("google/siglip-base-patch16-224")
        >>> tokenizer = AutoTokenizer.from_pretrained("google/siglip-base-patch16-224")

        >>> # important: make sure to set padding="max_length" as that's how the model was trained
        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding="max_length", return_tensors="pt")
        >>> with torch.no_grad():
        ...     text_features = model.get_text_features(**inputs)
        ```"""
        # Use SigLIP model's config for some fields (if specified) instead of those of vision & text components.
        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
        )

        text_outputs: BaseModelOutputWithPooling = self.text_model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )

        pooled_output = text_outputs.pooler_output

        return pooled_output

    @auto_docstring
    def get_image_features(
        self,
        pixel_values: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        interpolate_pos_encoding: bool = False,
    ) -> torch.FloatTensor:
        r"""
        Returns:
            image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by
            applying the projection layer to the pooled output of [`SiglipVisionModel`].

        Examples:

        ```python
        >>> from PIL import Image
        >>> import requests
        >>> from transformers import AutoProcessor, AutoModel
        >>> import torch

        >>> model = AutoModel.from_pretrained("google/siglip-base-patch16-224")
        >>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")

        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        >>> image = Image.open(requests.get(url, stream=True).raw)

        >>> inputs = processor(images=image, return_tensors="pt")

        >>> with torch.no_grad():
        ...     image_features = model.get_image_features(**inputs)
        ```"""
        # Use SiglipModel's config for some fields (if specified) instead of those of vision & text components.
        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
        )

        vision_outputs: BaseModelOutputWithPooling = self.vision_model(
            pixel_values=pixel_values,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            interpolate_pos_encoding=interpolate_pos_encoding,
        )

        pooled_output = vision_outputs.pooler_output

        return pooled_output

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        pixel_values: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        return_loss: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        interpolate_pos_encoding: bool = False,
    ) -> SiglipOutput:
        r"""
        return_loss (`bool`, *optional*):
            Whether or not to return the contrastive loss.

        Examples:

        ```python
        >>> from PIL import Image
        >>> import requests
        >>> from transformers import AutoProcessor, AutoModel
        >>> import torch

        >>> model = AutoModel.from_pretrained("google/siglip-base-patch16-224")
        >>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")

        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        >>> image = Image.open(requests.get(url, stream=True).raw)

        >>> texts = ["a photo of 2 cats", "a photo of 2 dogs"]
        >>> # important: we pass `padding=max_length` since the model was trained with this
        >>> inputs = processor(text=texts, images=image, padding="max_length", return_tensors="pt")

        >>> with torch.no_grad():
        ...     outputs = model(**inputs)

        >>> logits_per_image = outputs.logits_per_image
        >>> probs = torch.sigmoid(logits_per_image) # these are the probabilities
        >>> print(f"{probs[0][0]:.1%} that image 0 is '{texts[0]}'")
        31.9% that image 0 is 'a photo of 2 cats'
        ```"""
        # Use SigLIP model's config for some fields (if specified) instead of those of vision & text components.
        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
        )

        vision_outputs: BaseModelOutputWithPooling = self.vision_model(
            pixel_values=pixel_values,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            interpolate_pos_encoding=interpolate_pos_encoding,
        )

        text_outputs: BaseModelOutputWithPooling = self.text_model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )

        image_embeds = vision_outputs.pooler_output
        text_embeds = text_outputs.pooler_output

        # normalized features
        image_embeds = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True)
        text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)

        # cosine similarity as logits
        logits_per_text = torch.matmul(text_embeds, image_embeds.t().to(text_embeds.device))

        logit_scale, logit_bias = self.logit_scale.to(text_embeds.device), self.logit_bias.to(text_embeds.device)
        logits_per_text = logits_per_text * logit_scale.exp() + logit_bias

        logits_per_image = logits_per_text.t()

        loss = None
        if return_loss:
            # Adapted from https://github.com/google-research/big_vision/blob/01edb81a4716f93a48be43b3a4af14e29cdb3a7f/big_vision/trainers/proj/image_text/siglip.py#L287
            eye = torch.eye(logits_per_text.size(0), device=logits_per_text.device)
            m1_diag1 = -torch.ones_like(logits_per_text) + 2 * eye
            loglik = torch.nn.functional.logsigmoid(m1_diag1 * logits_per_text)
            nll = -torch.sum(loglik, dim=-1)
            loss = nll.mean()

        return SiglipOutput(
            loss=loss,
            logits_per_image=logits_per_image,
            logits_per_text=logits_per_text,
            text_embeds=text_embeds,
            image_embeds=image_embeds,
            text_model_output=text_outputs,
            vision_model_output=vision_outputs,
        )


@auto_docstring(
    custom_intro="""
    SigLIP vision encoder with an image classification head on top (a linear layer on top of the pooled final hidden states of
    the patch tokens) e.g. for ImageNet.
    """
)
class SiglipForImageClassification(SiglipPreTrainedModel):
    main_input_name = "pixel_values"

    def __init__(self, config: SiglipConfig) -> None:
        super().__init__(config)

        self.num_labels = config.num_labels

        # Create the vision model with proper attention
        # and take only vision_model submodule (for backward compatibility)
        vision_model = SiglipVisionModel._from_config(config.vision_config)
        self.vision_model = vision_model.vision_model

        # Classifier head
        self.classifier = (
            nn.Linear(config.vision_config.hidden_size, config.num_labels) if config.num_labels > 0 else nn.Identity()
        )

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        pixel_values: Optional[torch.Tensor] = None,
        labels: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        interpolate_pos_encoding: bool = False,
    ) -> ImageClassifierOutput:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the image 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 AutoImageProcessor, SiglipForImageClassification
        >>> import torch
        >>> from PIL import Image
        >>> import requests

        >>> torch.manual_seed(3)  # doctest: +IGNORE_RESULT
        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        >>> image = Image.open(requests.get(url, stream=True).raw)

        >>> # note: we are loading a `SiglipModel` from the hub here,
        >>> # so the head will be randomly initialized, hence the predictions will be random if seed is not set above.
        >>> image_processor = AutoImageProcessor.from_pretrained("google/siglip-base-patch16-224")
        >>> model = SiglipForImageClassification.from_pretrained("google/siglip-base-patch16-224")

        >>> inputs = image_processor(images=image, return_tensors="pt")
        >>> outputs = model(**inputs)
        >>> logits = outputs.logits
        >>> # model predicts one of the two classes
        >>> predicted_class_idx = logits.argmax(-1).item()
        >>> print("Predicted class:", model.config.id2label[predicted_class_idx])
        Predicted class: LABEL_1
        ```"""
        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
        )

        outputs: BaseModelOutputWithPooling = self.vision_model(
            pixel_values,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            interpolate_pos_encoding=interpolate_pos_encoding,
        )

        sequence_output = outputs.last_hidden_state

        # average pool the patch tokens
        sequence_output = torch.mean(sequence_output, dim=1)
        # apply classifier
        logits = self.classifier(sequence_output)

        loss = None
        if labels is not None:
            # move labels to correct device to enable model parallelism
            labels = labels.to(logits.device)
            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 ImageClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


__all__ = [
    "SiglipModel",
    "SiglipPreTrainedModel",
    "SiglipTextModel",
    "SiglipVisionModel",
    "SiglipForImageClassification",
]