visual_modeling.py

# coding=utf-8
# Copyright 2021 The OpenAI Team Authors 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 CLIP model."""

from dataclasses import dataclass
from typing import Any, Optional, Tuple, Union

import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
import torch.nn.functional as F

from transformers.activations import ACT2FN
from transformers.modeling_attn_mask_utils import _create_4d_causal_attention_mask, _prepare_4d_attention_mask
from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, ImageClassifierOutput
from transformers.modeling_utils import PreTrainedModel
from transformers.pytorch_utils import is_torch_greater_or_equal_than_2_2
from transformers.utils import (
    ModelOutput,
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    is_flash_attn_2_available,
    is_flash_attn_greater_or_equal_2_10,
    logging,
    replace_return_docstrings,
    torch_int,
)
try:
    from configuration_clip import CLIPConfig, CLIPTextConfig, CLIPVisionConfig
except ImportError:
    from .configuration_clip import CLIPConfig, CLIPTextConfig, CLIPVisionConfig


if is_flash_attn_2_available():
    from transformers.modeling_flash_attention_utils import _flash_attention_forward


logger = logging.get_logger(__name__)

# General docstring
_CONFIG_FOR_DOC = "CLIPConfig"
_CHECKPOINT_FOR_DOC = "openai/clip-vit-base-patch32"

# Image classification docstring
_IMAGE_CLASS_CHECKPOINT = "openai/clip-vit-base-patch32"
_IMAGE_CLASS_EXPECTED_OUTPUT = "LABEL_0"


# contrastive loss function, adapted from
# https://sachinruk.github.io/blog/2021-03-07-clip.html
def contrastive_loss(logits: torch.Tensor) -> torch.Tensor:
    return nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device))


def clip_loss(similarity: torch.Tensor) -> torch.Tensor:
    caption_loss = contrastive_loss(similarity)
    image_loss = contrastive_loss(similarity.t())
    return (caption_loss + image_loss) / 2.0


def _get_vector_norm(tensor: torch.Tensor) -> torch.Tensor:
    """
    This method is equivalent to tensor.norm(p=2, dim=-1, keepdim=True) and used to make
    model `executorch` exportable. See issue https://github.com/pytorch/executorch/issues/3566
    """
    square_tensor = torch.pow(tensor, 2)
    sum_tensor = torch.sum(square_tensor, dim=-1, keepdim=True)
    normed_tensor = torch.pow(sum_tensor, 0.5)
    return normed_tensor


@dataclass
class CLIPVisionModelOutput(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: torch.FloatTensor = None
    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None


@dataclass
class CLIPTextModelOutput(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: torch.FloatTensor = None
    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None


@dataclass
class CLIPOutput(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 [`CLIPTextModel`].
        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
            The image embeddings obtained by applying the projection layer to the pooled output of [`CLIPVisionModel`].
        text_model_output (`BaseModelOutputWithPooling`):
            The output of the [`CLIPTextModel`].
        vision_model_output (`BaseModelOutputWithPooling`):
            The output of the [`CLIPVisionModel`].
    """

    loss: Optional[torch.FloatTensor] = None
    logits_per_image: torch.FloatTensor = None
    logits_per_text: torch.FloatTensor = None
    text_embeds: torch.FloatTensor = None
    image_embeds: 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 CLIPVisionEmbeddings(nn.Module):
    def __init__(self, config: CLIPVisionConfig):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.image_size = config.image_size
        self.patch_size = config.patch_size

        self.class_embedding = nn.Parameter(torch.randn(self.embed_dim))

        self.patch_embedding = nn.Conv2d(
            in_channels=config.num_channels,
            out_channels=self.embed_dim,
            kernel_size=self.patch_size,
            stride=self.patch_size,
            bias=False,
        )

        self.num_patches = (self.image_size // self.patch_size) ** 2
        self.num_positions = self.num_patches + 1
        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.

        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] - 1
        position_embedding = self.position_embedding.weight.unsqueeze(0)
        num_positions = position_embedding.shape[1] - 1

        # 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)

        class_pos_embed = position_embedding[:, :1]
        patch_pos_embed = position_embedding[:, 1:]

        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 torch.cat((class_pos_embed, patch_pos_embed), dim=1)

    def forward(self, pixel_values: torch.FloatTensor, interpolate_pos_encoding=False) -> torch.Tensor:
        batch_size, _, height, width = pixel_values.shape
        if not interpolate_pos_encoding and (height != self.image_size or width != self.image_size):
            raise ValueError(
                f"Input image size ({height}*{width}) doesn't match model" f" ({self.image_size}*{self.image_size})."
            )
        target_dtype = self.patch_embedding.weight.dtype
        patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]
        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)

        class_embeds = self.class_embedding.expand(batch_size, 1, -1)
        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
        if interpolate_pos_encoding:
            embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width)
        else:
            embeddings = embeddings + self.position_embedding(self.position_ids)
        return embeddings


class CLIPTextEmbeddings(nn.Module):
    def __init__(self, config: CLIPTextConfig):
        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]

        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


class CLIPAttention(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.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 _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()

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

        bsz, tgt_len, embed_dim = hidden_states.size()

        # get query proj
        query_states = self.q_proj(hidden_states) * self.scale
        key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
        value_states = self._shape(self.v_proj(hidden_states), -1, bsz)

        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
        key_states = key_states.view(*proj_shape)
        value_states = value_states.view(*proj_shape)

        src_len = key_states.size(1)
        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))

        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
                f" {attn_weights.size()}"
            )

        # apply the causal_attention_mask first
        if causal_attention_mask is not None:
            if causal_attention_mask.size() != (bsz, 1, tgt_len, src_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is"
                    f" {causal_attention_mask.size()}"
                )
            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + causal_attention_mask
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
                )
            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

        attn_weights = nn.functional.softmax(attn_weights, dim=-1)

        if output_attentions:
            # this operation is a bit akward, but it's required to
            # make sure that attn_weights keeps its gradient.
            # In order to do so, attn_weights have to reshaped
            # twice and have to be reused in the following
            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
        else:
            attn_weights_reshaped = None

        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)

        attn_output = torch.bmm(attn_probs, value_states)

        if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
        attn_output = attn_output.transpose(1, 2)
        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights_reshaped


class CLIPFlashAttention2(CLIPAttention):
    """
    CLIPAttention flash attention module. This module inherits from `CLIPAttention` as the weights of the module stays
    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
    flash attention and deal with padding tokens in case the input contains any of them.
    """

    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()

    # Adapted from transformers.models.llama.modeling_llama.LlamaFlashAttention2.forward
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        causal_attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
        output_attentions = False

        batch_size, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        # Flash attention requires the input to have the shape
        # batch_size x seq_length x head_dim x hidden_dim
        # therefore we just need to keep the original shape
        query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim)
        key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim)
        value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim)

        dropout_rate = self.dropout if self.training else 0.0

        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
        # therefore the input hidden states gets silently casted in float32. Hence, we need
        # cast them back in the correct dtype just to be sure everything works as expected.
        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
        # in fp32.

        input_dtype = query_states.dtype
        if input_dtype == torch.float32:
            if torch.is_autocast_enabled():
                target_dtype = torch.get_autocast_gpu_dtype()
            # Handle the case where the model is quantized
            elif hasattr(self.config, "_pre_quantization_dtype"):
                target_dtype = self.config._pre_quantization_dtype
            else:
                target_dtype = self.q_proj.weight.dtype

            logger.warning_once(
                f"The input hidden states seems to be silently casted in float32, this might be related to"
                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
                f" {target_dtype}."
            )

            query_states = query_states.to(target_dtype)
            key_states = key_states.to(target_dtype)
            value_states = value_states.to(target_dtype)

        attn_output = _flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            dropout=dropout_rate,
            is_causal=causal_attention_mask is not None,
            use_top_left_mask=self._flash_attn_uses_top_left_mask,
        )

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

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights


class CLIPSdpaAttention(CLIPAttention):
    """
    SDPA attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
    `CLIPAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
    SDPA API.
    """

    # Adapted from CLIPAttention.forward
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        causal_attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
        if output_attentions:
            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
            logger.warning_once(
                "CLIPModel is using CLIPSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not "
                "support `output_attentions=True`. Falling back to the manual attention implementation, but specifying "
                "the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can "
                'be removed using the argument `attn_implementation="eager"` when loading the model.'
            )
            return super().forward(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                causal_attention_mask=causal_attention_mask,
                output_attentions=output_attentions,
            )

        # CLIP text model uses both `causal_attention_mask` and `attention_mask`
        if attention_mask is not None and causal_attention_mask is not None:
            attn_mask = attention_mask + causal_attention_mask
        elif causal_attention_mask is not None:
            attn_mask = causal_attention_mask
        else:
            attn_mask = attention_mask

        bsz, tgt_len, embed_dim = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)

        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
        # Reference: https://github.com/pytorch/pytorch/issues/112577.
        if not is_torch_greater_or_equal_than_2_2 and query_states.device.type == "cuda" and attn_mask is not None:
            query_states = query_states.contiguous()
            key_states = key_states.contiguous()
            value_states = value_states.contiguous()

        # CLIP text model uses both `causal_attention_mask` and `attention_mask` sequentially.
        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=attn_mask,
            dropout_p=self.dropout if self.training else 0.0,
            scale=self.scale,
        )

        attn_output = attn_output.transpose(1, 2)
        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, None


CLIP_ATTENTION_CLASSES = {
    "eager": CLIPAttention,
    "sdpa": CLIPSdpaAttention,
    "flash_attention_2": CLIPFlashAttention2,
}


class CLIPMLP(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 CLIPEncoderLayer(nn.Module):
    def __init__(self, config: CLIPConfig):
        super().__init__()
        self.embed_dim = config.hidden_size
        self.self_attn = CLIP_ATTENTION_CLASSES[config._attn_implementation](config)
        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
        self.mlp = CLIPMLP(config)
        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        causal_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 size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
                `(config.encoder_attention_heads,)`.
            output_attentions (`bool`, *optional*):
                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,
            causal_attention_mask=causal_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


class CLIPPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = CLIPConfig
    base_model_prefix = "clip"
    supports_gradient_checkpointing = True
    _supports_sdpa = True
    _supports_flash_attn_2 = True

    def _init_weights(self, module):
        """Initialize the weights"""
        factor = self.config.initializer_factor
        if isinstance(module, CLIPTextEmbeddings):
            module.token_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02)
            module.position_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02)
        elif isinstance(module, CLIPVisionEmbeddings):
            factor = self.config.initializer_factor
            nn.init.normal_(module.class_embedding, mean=0.0, std=module.embed_dim**-0.5 * factor)
            nn.init.normal_(module.patch_embedding.weight, std=module.config.initializer_range * factor)
            nn.init.normal_(module.position_embedding.weight, std=module.config.initializer_range * factor)
        elif isinstance(module, CLIPAttention):
            factor = self.config.initializer_factor
            in_proj_std = (module.embed_dim**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
            out_proj_std = (module.embed_dim**-0.5) * factor
            nn.init.normal_(module.q_proj.weight, std=in_proj_std)
            nn.init.normal_(module.k_proj.weight, std=in_proj_std)
            nn.init.normal_(module.v_proj.weight, std=in_proj_std)
            nn.init.normal_(module.out_proj.weight, std=out_proj_std)
        elif isinstance(module, CLIPMLP):
            factor = self.config.initializer_factor
            in_proj_std = (module.config.hidden_size**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
            fc_std = (2 * module.config.hidden_size) ** -0.5 * factor
            nn.init.normal_(module.fc1.weight, std=fc_std)
            nn.init.normal_(module.fc2.weight, std=in_proj_std)
        elif isinstance(module, CLIPModel):
            nn.init.normal_(
                module.text_projection.weight,
                std=module.text_embed_dim**-0.5 * self.config.initializer_factor,
            )
            nn.init.normal_(
                module.visual_projection.weight,
                std=module.vision_embed_dim**-0.5 * self.config.initializer_factor,
            )

        if isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)
        if isinstance(module, nn.Linear) and module.bias is not None:
            module.bias.data.zero_()


CLIP_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.

    Parameters:
        config ([`CLIPConfig`]): Model configuration class with all the parameters of the model.
            Initializing with a config file does not load the weights associated with the model, only the
            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""

CLIP_TEXT_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
            it.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        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)
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.max_position_embeddings - 1]`.

            [What are position IDs?](../glossary#position-ids)
        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.
"""

CLIP_VISION_INPUTS_DOCSTRING = r"""
    Args:
        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
        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.
        interpolate_pos_encoding (`bool`, *optional*, defaults `False`):
            Whether to interpolate the pre-trained position encodings.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""

CLIP_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
            it.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        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)
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.max_position_embeddings - 1]`.

            [What are position IDs?](../glossary#position-ids)
        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
        return_loss (`bool`, *optional*):
            Whether or not to return the contrastive loss.
        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.
        interpolate_pos_encoding (`bool`, *optional*, defaults `False`):
            Whether to interpolate the pre-trained position encodings.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""


class CLIPEncoder(nn.Module):
    """
    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
    [`CLIPEncoderLayer`].

    Args:
        config: CLIPConfig
    """

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

    def forward(
        self,
        inputs_embeds,
        attention_mask: Optional[torch.Tensor] = None,
        causal_attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, 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)
            causal_attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Causal mask for the text model. 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
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

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

        hidden_states = inputs_embeds
        for idx, encoder_layer in enumerate(self.layers):
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)
            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    encoder_layer.__call__,
                    hidden_states,
                    attention_mask,
                    causal_attention_mask,
                    output_attentions,
                )
            else:
                layer_outputs = encoder_layer(
                    hidden_states,
                    attention_mask,
                    causal_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,)

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

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

        self.embeddings = CLIPVisionEmbeddings(config)
        self.pre_layrnorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
        self.encoder = CLIPEncoder(config)
        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)

    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPVisionConfig)
    def forward(
        self,
        pixel_values: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        interpolate_pos_encoding: Optional[bool] = False,
    ) -> Union[Tuple, BaseModelOutputWithPooling]:
        r"""
        Returns:

        """
        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.use_return_dict

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

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

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

        last_hidden_state = encoder_outputs[0]
        pooled_output = last_hidden_state[:, 0, :]
        pooled_output = self.post_layernorm(pooled_output)

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

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


@add_start_docstrings(
    """The vision model from CLIP without any head or projection on top.""",
    CLIP_START_DOCSTRING,
)
class CLIPVisionModel(CLIPPreTrainedModel):
    config_class = CLIPVisionConfig
    main_input_name = "pixel_values"
    _no_split_modules = ["CLIPEncoderLayer"]

    def __init__(self, config: CLIPVisionConfig):
        super().__init__(config)
        self.vision_model = CLIPVisionTransformer(config)
        # Initialize weights and apply final processing
        self.post_init()

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

    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPVisionConfig)
    def forward(
        self,
        pixel_values: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        interpolate_pos_encoding: bool = False,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutputWithPooling]:
        r"""
        Returns:

        Examples:

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

        >>> model = CLIPVisionModel.from_pretrained("openai/clip-vit-base-patch32")
        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")

        >>> 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 CLS states
        ```"""
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

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


@add_start_docstrings(CLIP_START_DOCSTRING)
class CLIPModel(CLIPPreTrainedModel):
    config_class = CLIPConfig
    _no_split_modules = ["CLIPTextEmbeddings", "CLIPEncoderLayer", "CLIPVisionEmbeddings"]

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

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

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

        text_config = config.text_config
        vision_config = config.vision_config

        self.projection_dim = config.projection_dim
        self.text_embed_dim = text_config.hidden_size
        self.vision_embed_dim = vision_config.hidden_size

        vision_model = CLIPVisionModel._from_config(vision_config)
        self.vision_model = vision_model.vision_model

        self.visual_projection = nn.Linear(self.vision_embed_dim, self.projection_dim, bias=False)
        self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim, bias=False)
        self.logit_scale = nn.Parameter(torch.tensor(self.config.logit_scale_init_value))

        # Initialize weights and apply final processing
        self.post_init()
        self.reference_embedding = None
        self.cossim = nn.CosineSimilarity(dim=-1)
        
    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_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,
        return_dict: Optional[bool] = None,
    ) -> 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 [`CLIPVisionModel`].

        Examples:

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

        >>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
        >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")

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

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

        >>> image_features = model.get_image_features(**inputs)
        ```"""
        # Use CLIP 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
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

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

        output = vision_outputs[0]
        return output
    
    @torch.no_grad()
    def set_reference_embedding(self, x):
        self.reference_embedding = self.get_image_features(x)[:, 0, :]
        
    def encode_image(self, x, n_patches=64):
        image_embeds = self.get_image_features(x)
        
        image_embeds = image_embeds[:, 1:, :]
        b, n, c = image_embeds.shape
        sqrt_n = int(n**0.5)
        image_embeds = image_embeds.permute(0, 2, 1).view(b, c, sqrt_n, sqrt_n)
        stride = int(sqrt_n // (n_patches ** 0.5))
        image_embeds = F.avg_pool2d(image_embeds, kernel_size=(stride, stride), stride=stride)
        image_embeds = image_embeds.view(b, c, -1).permute(0, 2, 1).contiguous()
        
        return image_embeds
    
    def encode_image_w_similarity(self, x, n_patches=64):
        image_embeds = self.get_image_features(x)
        
        # Calculate cosine similarity with reference embedding before processing
        original_embeds = image_embeds[:, 0, :]
        cos = nn.CosineSimilarity(dim=-1)
        similarity = cos(original_embeds, self.reference_embedding)
        
        image_embeds = image_embeds[:, 1:, :]
        b, n, c = image_embeds.shape
        sqrt_n = int(n**0.5)
        image_embeds = image_embeds.permute(0, 2, 1).view(b, c, sqrt_n, sqrt_n)
        stride = int(sqrt_n // (n_patches ** 0.5))
        image_embeds = F.avg_pool2d(image_embeds, kernel_size=(stride, stride), stride=stride)
        image_embeds = image_embeds.view(b, c, -1).permute(0, 2, 1).contiguous()
        
        return image_embeds, similarity