image-match/python/py/Lib/site-packages/transformers/models/siglip/modeling_siglip.py

# 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."""

from collections.abc import Callable
from dataclasses import dataclass
from typing import Any

import numpy as np
import torch
from torch import nn

from ... import initialization as init
from ...activations import ACT2FN
from ...masking_utils import create_bidirectional_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, ImageClassifierOutput
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import (
    ModelOutput,
    TransformersKwargs,
    auto_docstring,
    can_return_tuple,
    torch_int,
)
from ...utils.generic import merge_with_config_defaults
from ...utils.output_capturing import capture_outputs
from .configuration_siglip import SiglipConfig, SiglipTextConfig, SiglipVisionConfig


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

    image_embeds: torch.FloatTensor | None = None
    last_hidden_state: torch.FloatTensor | None = None
    hidden_states: tuple[torch.FloatTensor, ...] | None = None
    attentions: tuple[torch.FloatTensor, ...] | None = None


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

    text_embeds: torch.FloatTensor | None = None
    last_hidden_state: torch.FloatTensor | None = None
    hidden_states: tuple[torch.FloatTensor, ...] | None = None
    attentions: tuple[torch.FloatTensor, ...] | None = None


@dataclass
@auto_docstring
# Copied from transformers.models.clip.modeling_clip.CLIPOutput with CLIP->Siglip
class SiglipOutput(ModelOutput):
    r"""
    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: torch.FloatTensor | None = None
    logits_per_image: torch.FloatTensor | None = None
    logits_per_text: torch.FloatTensor | None = None
    text_embeds: torch.FloatTensor | None = None
    image_embeds: torch.FloatTensor | None = None
    text_model_output: BaseModelOutputWithPooling = None
    vision_model_output: BaseModelOutputWithPooling = None

    def to_tuple(self) -> tuple[Any]:
        return tuple(v.to_tuple() if isinstance(v, ModelOutput) else v for v in self.values())


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: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = 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: torch.Tensor | None,
    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: torch.Tensor | None = None,
        **kwargs,
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        """Input shape: Batch x Time x Channel"""

        input_shape = hidden_states.shape[:-1]

        hidden_shape = (*input_shape, -1, self.head_dim)
        queries = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        keys = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        values = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

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

        attn_output, attn_weights = attention_interface(
            self,
            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(*input_shape, -1).contiguous()
        attn_output = self.out_proj(attn_output)

        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: 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)

    @auto_docstring
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.FloatTensor:
        residual = hidden_states

        hidden_states = self.layer_norm1(hidden_states)
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            **kwargs,
        )
        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

        return hidden_states


@auto_docstring
class SiglipPreTrainedModel(PreTrainedModel):
    config: SiglipConfig
    base_model_prefix = "siglip"
    input_modalities = ("image", "text")
    supports_gradient_checkpointing = True

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

    _can_record_outputs = {
        "hidden_states": SiglipEncoderLayer,
        "attentions": SiglipAttention,
    }

    @torch.no_grad()
    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
            )
            init.normal_(module.position_embedding.weight, std=1 / np.sqrt(width))
            if hasattr(module, "position_ids"):
                init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
        elif isinstance(module, nn.Embedding):
            init.default_flax_embed_init_(module.weight)
        elif isinstance(module, SiglipAttention):
            init.xavier_uniform_(module.q_proj.weight)
            init.xavier_uniform_(module.k_proj.weight)
            init.xavier_uniform_(module.v_proj.weight)
            init.xavier_uniform_(module.out_proj.weight)
            init.zeros_(module.q_proj.bias)
            init.zeros_(module.k_proj.bias)
            init.zeros_(module.v_proj.bias)
            init.zeros_(module.out_proj.bias)
        elif isinstance(module, SiglipMLP):
            init.xavier_uniform_(module.fc1.weight)
            init.xavier_uniform_(module.fc2.weight)
            init.normal_(module.fc1.bias, std=1e-6)
            init.normal_(module.fc2.bias, std=1e-6)
        elif isinstance(module, SiglipMultiheadAttentionPoolingHead):
            init.xavier_uniform_(module.probe)
            init.xavier_uniform_(module.attention.in_proj_weight)
            init.zeros_(module.attention.in_proj_bias)
        elif isinstance(module, SiglipModel):
            init.zeros_(module.logit_scale)
            init.zeros_(module.logit_bias)
        elif isinstance(module, SiglipForImageClassification):
            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)):
            init.lecun_normal_(module.weight)
            if module.bias is not None:
                init.zeros_(module.bias)
        elif isinstance(module, nn.LayerNorm):
            init.zeros_(module.bias)
            init.ones_(module.weight)
        elif isinstance(module, SiglipTextEmbeddings):
            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))


# 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
    @auto_docstring
    def forward(
        self,
        inputs_embeds,
        attention_mask: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutput:
        hidden_states = inputs_embeds
        for encoder_layer in self.layers:
            hidden_states = encoder_layer(
                hidden_states,
                attention_mask,
                **kwargs,
            )

        return BaseModelOutput(last_hidden_state=hidden_states)


class SiglipTextTransformer(SiglipPreTrainedModel):
    _input_embed_layer = "token_embedding"

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPooling:
        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.
        attention_mask = create_bidirectional_mask(
            config=self.config,
            inputs_embeds=hidden_states,
            attention_mask=attention_mask,
        )

        encoder_outputs: BaseModelOutput = self.encoder(
            inputs_embeds=hidden_states,
            attention_mask=attention_mask,
            **kwargs,
        )

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

        # The model uses the last token's hidden state, which may be padding.
        pooled_output = last_hidden_state[:, -1, :]
        pooled_output = self.head(pooled_output)

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


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

    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

    @merge_with_config_defaults
    @capture_outputs(tie_last_hidden_states=False)
    @auto_docstring
    def forward(
        self,
        input_ids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> 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,
            **kwargs,
        )


class SiglipVisionTransformer(SiglipPreTrainedModel):
    _input_embed_layer = "patch_embedding"

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

        self.post_init()

    @auto_docstring
    def forward(
        self,
        pixel_values,
        interpolate_pos_encoding: bool | None = False,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPooling:
        hidden_states = self.embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding)

        encoder_outputs: BaseModelOutput = self.encoder(
            inputs_embeds=hidden_states,
            **kwargs,
        )

        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,
        )


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: SiglipVisionConfig
    main_input_name = "pixel_values"
    input_modalities = ("image",)

    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

    @merge_with_config_defaults
    @capture_outputs(tie_last_hidden_states=False)
    @auto_docstring
    def forward(
        self,
        pixel_values,
        interpolate_pos_encoding: bool = False,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPooling:
        r"""
        Examples:

        ```python
        >>> from PIL import Image
        >>> import httpx
        >>> from io import BytesIO
        >>> 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"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read()))

        >>> 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,
            interpolate_pos_encoding=interpolate_pos_encoding,
            **kwargs,
        )


@auto_docstring
class SiglipModel(SiglipPreTrainedModel):
    config: 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()

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

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

    @can_return_tuple
    @auto_docstring
    def get_text_features(
        self,
        input_ids: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        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)
        ```"""
        return self.text_model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            **kwargs,
        )

    @can_return_tuple
    @auto_docstring
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        interpolate_pos_encoding: bool = False,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        Examples:

        ```python
        >>> import torch
        >>> from transformers import AutoProcessor, AutoModel
        >>> from transformers.image_utils import load_image

        >>> 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 = load_image(url)

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

        >>> with torch.no_grad():
        ...     image_features = model.get_image_features(**inputs)
        ```"""
        return self.vision_model(
            pixel_values=pixel_values,
            interpolate_pos_encoding=interpolate_pos_encoding,
            **kwargs,
        )

    # NOTE: SiglipModel uses Pretrained backbones, so we don't need to add `capture_outputs` here
    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        return_loss: bool | None = None,
        interpolate_pos_encoding: bool = False,
        **kwargs: Unpack[TransformersKwargs],
    ) -> SiglipOutput:
        r"""
        return_loss (`bool`, *optional*):
            Whether or not to return the contrastive loss.

        Examples:

        ```python
        >>> from PIL import Image
        >>> import httpx
        >>> from io import BytesIO
        >>> 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"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read()))

        >>> 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'
        ```"""
        vision_outputs: BaseModelOutputWithPooling = self.vision_model(
            pixel_values=pixel_values,
            interpolate_pos_encoding=interpolate_pos_encoding,
            **kwargs,
        )

        text_outputs: BaseModelOutputWithPooling = self.text_model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            **kwargs,
        )

        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"
    input_modalities = ("image",)

    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()

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

    def set_input_embeddings(self, value: nn.Module):
        self.vision_model.embeddings.patch_embedding = value

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(
        self,
        pixel_values: torch.Tensor | None = None,
        labels: torch.Tensor | None = None,
        interpolate_pos_encoding: bool = False,
        **kwargs: Unpack[TransformersKwargs],
    ) -> 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 httpx
        >>> from io import BytesIO

        >>> torch.manual_seed(3)  # doctest: +IGNORE_RESULT
        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read()))

        >>> # 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
        ```"""
        outputs: BaseModelOutputWithPooling = self.vision_model(
            pixel_values,
            interpolate_pos_encoding=interpolate_pos_encoding,
            **kwargs,
        )

        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:
            loss = self.loss_function(labels, logits, self.config)

        return ImageClassifierOutput(
            loss=loss,
            logits=logits,
        )


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