#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/dinov2/modular_dinov2.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_dinov2.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2023 Meta AI and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from collections.abc import Callable, Iterable

import torch
from torch import nn

from ... import initialization as init
from ...activations import ACT2FN
from ...backbone_utils import BackboneMixin, filter_output_hidden_states
from ...masking_utils import create_bidirectional_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BackboneOutput, BaseModelOutput, BaseModelOutputWithPooling, ImageClassifierOutput
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import TransformersKwargs, auto_docstring, torch_int
from ...utils.generic import can_return_tuple, merge_with_config_defaults
from ...utils.output_capturing import capture_outputs
from .configuration_dinov2 import Dinov2Config


class Dinov2PatchEmbeddings(nn.Module):
    """
    This class turns `pixel_values` of shape `(batch_size, num_channels, height, width)` into the initial
    `hidden_states` (patch embeddings) of shape `(batch_size, seq_length, hidden_size)` to be consumed by a
    Transformer.
    """

    def __init__(self, config: Dinov2Config):
        super().__init__()
        image_size = config.image_size
        patch_size = config.patch_size
        image_size = image_size if isinstance(image_size, Iterable) else (image_size, image_size)
        patch_size = patch_size if isinstance(patch_size, Iterable) else (patch_size, patch_size)

        self.num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
        self.image_size = image_size
        self.patch_size = patch_size
        self.num_channels = config.num_channels
        self.projection = nn.Conv2d(config.num_channels, config.hidden_size, kernel_size=patch_size, stride=patch_size)

    def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
        num_channels = pixel_values.shape[1]
        if num_channels != self.num_channels:
            raise ValueError(
                "Make sure that the channel dimension of the pixel values match with the one set in the configuration."
                f" Expected {self.num_channels} but got {num_channels}."
            )
        pixel_values = pixel_values.to(self.projection.weight.dtype)
        return self.projection(pixel_values).flatten(2).transpose(1, 2)


class Dinov2Embeddings(nn.Module):
    """Construct the CLS token, mask token, position and patch embeddings."""

    def __init__(self, config: Dinov2Config) -> None:
        super().__init__()

        self.cls_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
        self.mask_token = nn.Parameter(torch.zeros(1, config.hidden_size)) if config.use_mask_token else None
        self.patch_embeddings = Dinov2PatchEmbeddings(config)
        self.patch_size = config.patch_size
        self.position_embeddings = nn.Parameter(
            torch.randn(1, self.patch_embeddings.num_patches + 1, config.hidden_size)
        )
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.config = config

    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 interpolation at torch.float32 precision.

        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
        num_positions = self.position_embeddings.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_embeddings

        class_pos_embed = self.position_embeddings[:, :1]
        patch_pos_embed = self.position_embeddings[:, 1:]
        dim = embeddings.shape[-1]
        patch_height, patch_width = self.patch_embeddings.patch_size
        new_height = height // patch_height
        new_width = width // patch_width

        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)
        # interpolate in float32; bicubic resampling in half precision shifts the embeddings
        target_dtype = patch_pos_embed.dtype
        patch_pos_embed = nn.functional.interpolate(
            patch_pos_embed.to(torch.float32),
            size=(new_height, new_width),
            mode="bicubic",
            align_corners=False,
        ).to(dtype=target_dtype)
        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.Tensor,
        bool_masked_pos: torch.BoolTensor | None = None,
    ) -> torch.Tensor:
        _, _, height, width = pixel_values.shape
        embeddings = self.patch_embeddings(pixel_values)
        batch_size, seq_len, _ = embeddings.size()

        if bool_masked_pos is not None:
            mask_tokens = self.mask_token.expand(batch_size, seq_len, -1)
            # replace the masked visual tokens by mask_tokens
            mask = bool_masked_pos.unsqueeze(-1).type_as(mask_tokens)
            embeddings = embeddings * (1 - mask) + mask_tokens * mask

        cls_tokens = self.cls_token.expand(batch_size, -1, -1)
        embeddings = torch.cat((cls_tokens, embeddings), dim=1)

        if self.position_embeddings is not None:
            embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width)

        embeddings = self.dropout(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 | None = None,
    dropout: float = 0.0,
    **kwargs: Unpack[TransformersKwargs],
):
    if scaling is None:
        scaling = query.size(-1) ** -0.5

    # Take the dot product between "query" and "key" to get the raw attention scores.
    attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling

    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1)
    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 Dinov2Attention(nn.Module):
    def __init__(self, config: Dinov2Config):
        super().__init__()
        self.config = config
        self.num_attention_heads = config.num_attention_heads
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.attention_dropout = config.attention_probs_dropout_prob
        self.scaling = self.head_dim**-0.5
        self.is_causal = False

        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.qkv_bias)
        self.k_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.qkv_bias)
        self.v_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.qkv_bias)
        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=True)

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

        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        value_states = 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,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)

        return attn_output, attn_weights


class Dinov2LayerScale(nn.Module):
    def __init__(self, config) -> None:
        super().__init__()
        self.lambda1 = nn.Parameter(config.layerscale_value * torch.ones(config.hidden_size))

    def forward(self, hidden_state: torch.Tensor) -> torch.Tensor:
        return hidden_state * self.lambda1


class Dinov2MLP(nn.Module):
    def __init__(self, config) -> None:
        super().__init__()
        self.config = config
        self.activation_fn = ACT2FN[config.hidden_act]
        # the hidden size comes from mlp_ratio; the config has no intermediate_size
        self.fc1 = nn.Linear(config.hidden_size, int(config.hidden_size * config.mlp_ratio))
        self.fc2 = nn.Linear(int(config.hidden_size * config.mlp_ratio), 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 Dinov2SwiGLUFFN(nn.Module):
    def __init__(self, config) -> None:
        super().__init__()
        hidden_features = int(config.hidden_size * config.mlp_ratio)
        # SwiGLU keeps two thirds of the MLP hidden size, rounded up to a multiple of 8
        hidden_features = (int(hidden_features * 2 / 3) + 7) // 8 * 8
        self.gate_proj = nn.Linear(config.hidden_size, hidden_features, bias=True)
        self.up_proj = nn.Linear(config.hidden_size, hidden_features, bias=True)
        self.down_proj = nn.Linear(hidden_features, config.hidden_size, bias=True)
        self.act_fn = ACT2FN["silu"]

    def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj


class Dinov2DropPath(nn.Module):
    """Stochastic depth (DropPath) per sample, for residual blocks.

    Identity when ``drop_prob`` is 0 or outside training. See `Deep Networks with Stochastic Depth
    <https://arxiv.org/abs/1603.09382>`_.
    """

    def __init__(self, drop_prob: float = 0.0) -> None:
        super().__init__()
        self.drop_prob = drop_prob

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        if self.drop_prob == 0.0 or not self.training:
            return hidden_states
        keep_prob = 1 - self.drop_prob
        shape = (hidden_states.shape[0],) + (1,) * (hidden_states.ndim - 1)
        random_tensor = torch.rand(shape, dtype=hidden_states.dtype, device=hidden_states.device)
        random_tensor = torch.floor(random_tensor + keep_prob)
        return hidden_states.div(keep_prob) * random_tensor

    def extra_repr(self) -> str:
        return f"p={self.drop_prob}"


class Dinov2Layer(GradientCheckpointingLayer):
    """This corresponds to the Block class in the original implementation."""

    def __init__(self, config: Dinov2Config) -> None:
        super().__init__()
        self.norm1 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.attention = Dinov2Attention(config)
        self.layer_scale1 = Dinov2LayerScale(config)
        self.drop_path = Dinov2DropPath(config.drop_path_rate) if config.drop_path_rate > 0.0 else nn.Identity()
        self.norm2 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.mlp = Dinov2SwiGLUFFN(config) if config.use_swiglu_ffn else Dinov2MLP(config)
        self.layer_scale2 = Dinov2LayerScale(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        # Self Attention
        residual = hidden_states
        hidden_states = self.norm1(hidden_states)
        hidden_states, _ = self.attention(hidden_states, attention_mask=attention_mask, **kwargs)
        hidden_states = self.layer_scale1(hidden_states)
        hidden_states = self.drop_path(hidden_states) + residual

        # Fully Connected
        residual = hidden_states
        hidden_states = self.norm2(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = self.layer_scale2(hidden_states)
        hidden_states = self.drop_path(hidden_states) + residual
        return hidden_states


@auto_docstring
class Dinov2PreTrainedModel(PreTrainedModel):
    config: Dinov2Config
    base_model_prefix = "dinov2"
    main_input_name = "pixel_values"
    input_modalities = ("image",)
    supports_gradient_checkpointing = True
    _no_split_modules = ["Dinov2Embeddings", "Dinov2Layer"]
    _supports_sdpa = True
    _supports_flash_attn = True
    _supports_flex_attn = True
    _supports_attention_backend = True
    _can_compile_fullgraph = True
    _can_record_outputs = {
        "hidden_states": Dinov2Layer,
        "attentions": Dinov2Attention,
    }
    _input_embed_layer = "patch_embeddings"

    @torch.no_grad()
    def _init_weights(self, module) -> None:
        """Initialize the weights"""
        super()._init_weights(module)
        if isinstance(module, Dinov2Embeddings):
            if module.position_embeddings is not None:
                init.trunc_normal_(module.position_embeddings, mean=0.0, std=self.config.initializer_range)
            init.trunc_normal_(module.cls_token, mean=0.0, std=self.config.initializer_range)
            if module.mask_token is not None:
                init.zeros_(module.mask_token)
        if isinstance(module, (nn.Linear, nn.Conv2d)):
            init.trunc_normal_(module.weight, mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                init.zeros_(module.bias)
        if isinstance(module, Dinov2LayerScale):
            init.constant_(module.lambda1, self.config.layerscale_value)


class Dinov2Encoder(Dinov2PreTrainedModel):
    def __init__(self, config: Dinov2Config):
        super().__init__(config)
        self.layer = nn.ModuleList([Dinov2Layer(config) for _ in range(config.num_hidden_layers)])
        self.post_init()

    @merge_with_config_defaults
    @capture_outputs(tie_last_hidden_states=False)
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutput:
        for layer_module in self.layer:
            hidden_states = layer_module(hidden_states, attention_mask, **kwargs)
        return BaseModelOutput(last_hidden_state=hidden_states)


@auto_docstring
class Dinov2Model(Dinov2PreTrainedModel):
    def __init__(self, config: Dinov2Config):
        super().__init__(config)
        self.config = config
        self.embeddings = Dinov2Embeddings(config)
        self.encoder = Dinov2Encoder(config)
        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.post_init()

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        pixel_values: torch.Tensor | None = None,
        bool_masked_pos: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPooling:
        r"""
        bool_masked_pos (`torch.BoolTensor` of shape `(batch_size, sequence_length)`):
            Boolean masked positions. Indicates which patches are masked (1) and which aren't (0). Only relevant for
            pre-training.
        """
        embedding_output = self.embeddings(pixel_values, bool_masked_pos=bool_masked_pos)
        attention_mask = create_bidirectional_mask(
            config=self.config,
            inputs_embeds=embedding_output,
            attention_mask=attention_mask,
        )
        encoder_outputs: BaseModelOutput = self.encoder(embedding_output, attention_mask=attention_mask, **kwargs)
        sequence_output = self.layernorm(encoder_outputs.last_hidden_state)
        pooled_output = sequence_output[:, 0, :]
        return BaseModelOutputWithPooling(
            last_hidden_state=sequence_output,
            pooler_output=pooled_output,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
        )


@auto_docstring(
    custom_intro="""
    Dinov2 Model transformer with an image classification head on top (a linear layer on top of the final hidden state
    of the [CLS] token) e.g. for ImageNet.
    """
)
class Dinov2ForImageClassification(Dinov2PreTrainedModel):
    def __init__(self, config: Dinov2Config) -> None:
        super().__init__(config)

        self.num_labels = config.num_labels
        self.dinov2 = Dinov2Model(config)
        self.classifier = (
            nn.Linear(config.hidden_size * 2, 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: torch.Tensor | None = None,
        labels: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> ImageClassifierOutput:
        outputs: BaseModelOutputWithPooling = self.dinov2(pixel_values, attention_mask=attention_mask, **kwargs)

        sequence_output = outputs.last_hidden_state
        cls_token = sequence_output[:, 0]
        patch_tokens = sequence_output[:, 1:]
        linear_input = torch.cat([cls_token, patch_tokens.mean(dim=1)], dim=1)
        logits = self.classifier(linear_input)

        loss = None
        if labels is not None:
            loss = self.loss_function(labels, logits, self.config, **kwargs)

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


@auto_docstring(
    custom_intro="""
    Dinov2 backbone, to be used with frameworks like DETR and MaskFormer.
    """
)
class Dinov2Backbone(BackboneMixin, Dinov2PreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_features = [config.hidden_size for _ in range(config.num_hidden_layers + 1)]
        self.embeddings = Dinov2Embeddings(config)
        self.encoder = Dinov2Encoder(config)
        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.post_init()

    @can_return_tuple
    @filter_output_hidden_states
    @auto_docstring
    def forward(
        self,
        pixel_values: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BackboneOutput:
        r"""
        Examples:

        ```python
        >>> from transformers import AutoImageProcessor, AutoBackbone
        >>> import torch
        >>> from PIL import Image
        >>> from huggingface_hub.utils import httpx
        >>> from io import BytesIO

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

        >>> processor = AutoImageProcessor.from_pretrained("facebook/dinov2-base")
        >>> model = AutoBackbone.from_pretrained(
        ...     "facebook/dinov2-base", out_features=["stage2", "stage5", "stage8", "stage11"]
        ... )

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

        >>> outputs = model(**inputs)
        >>> feature_maps = outputs.feature_maps
        >>> list(feature_maps[-1].shape)
        [1, 768, 16, 16]
        ```"""
        embedding_output = self.embeddings(pixel_values)
        attention_mask = create_bidirectional_mask(
            config=self.config,
            inputs_embeds=embedding_output,
            attention_mask=attention_mask,
        )
        output: BaseModelOutput = self.encoder(embedding_output, attention_mask=attention_mask, **kwargs)
        hidden_states = output.hidden_states

        feature_maps = ()
        for stage, hidden_state in zip(self.stage_names, hidden_states):
            if stage in self.out_features:
                if self.config.apply_layernorm:
                    hidden_state = self.layernorm(hidden_state)
                if self.config.reshape_hidden_states:
                    hidden_state = hidden_state[:, 1:]
                    # this was actually a bug in the original implementation that we copied here,
                    # cause normally the order is height, width
                    batch_size, _, height, width = pixel_values.shape
                    patch_height, patch_width = self.embeddings.patch_embeddings.patch_size
                    hidden_state = hidden_state.reshape(batch_size, height // patch_height, width // patch_width, -1)
                    hidden_state = hidden_state.permute(0, 3, 1, 2).contiguous()
                feature_maps += (hidden_state,)

        return BackboneOutput(
            feature_maps=feature_maps,
            hidden_states=hidden_states,
            attentions=output.attentions,
        )


__all__ = ["Dinov2ForImageClassification", "Dinov2Model", "Dinov2PreTrainedModel", "Dinov2Backbone"]