modeling_vitdet.py

# coding=utf-8
# 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.
""" PyTorch ViTDet backbone."""


import collections.abc
import math
from typing import Dict, List, Optional, Tuple, Union

import torch
import torch.utils.checkpoint
from torch import nn

from ...activations import ACT2FN
from ...modeling_outputs import BackboneOutput, BaseModelOutput
from ...modeling_utils import PreTrainedModel
from ...utils import (
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    logging,
    replace_return_docstrings,
)
from ...utils.backbone_utils import BackboneMixin
from .configuration_vitdet import VitDetConfig


logger = logging.get_logger(__name__)

# General docstring
_CONFIG_FOR_DOC = "VitDetConfig"


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

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

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

        if config.use_absolute_position_embeddings:
            # Initialize absolute positional embedding with pretrain image size.
            num_positions = num_patches + 1
            self.position_embeddings = nn.Parameter(torch.zeros(1, num_positions, config.hidden_size))
        else:
            self.position_embeddings = None

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

    def get_absolute_positions(self, abs_pos_embeddings, has_cls_token, height, width):
        """
        Calculate absolute positional embeddings. If needed, resize embeddings and remove cls_token dimension for the
        original embeddings.

        Args:
            abs_pos_embeddings (`torch.Tensor`):
                Absolute positional embeddings with (1, num_position, num_channels).
            has_cls_token (`bool`):
                If true, has 1 embedding in abs_pos_embeddings for cls token.
            height (`int`):
                Height of input image tokens.
            width (`int`):
                Width of input image tokens.

        Returns:
            Absolute positional embeddings after processing with shape (1, height, width, num_channels)
        """
        if has_cls_token:
            abs_pos_embeddings = abs_pos_embeddings[:, 1:]
        num_position = abs_pos_embeddings.shape[1]
        size = int(math.sqrt(num_position))  # This is a constant and can be recorded as such in the ONNX export.
        if size * size != num_position:
            raise ValueError("Absolute position embeddings must be a square number.")

        if torch.jit.is_tracing() or (size != height or size != width):
            # nn.functional.interpolate is a noop in case size == height and size == width - we need to always capture this path with jit.trace.
            new_abs_pos_embeddings = nn.functional.interpolate(
                abs_pos_embeddings.reshape(1, size, size, -1).permute(0, 3, 1, 2),
                size=(height, width),
                mode="bicubic",
                align_corners=False,
            )

            return new_abs_pos_embeddings.permute(0, 2, 3, 1)
        else:
            return abs_pos_embeddings.reshape(1, height, width, -1)

    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}."
            )
        embeddings = self.projection(pixel_values)

        if self.position_embeddings is not None:
            # (batch_size, num_channels, height, width) -> (batch_size, height, width, num_channels)
            embeddings = embeddings.permute(0, 2, 3, 1)
            # add position embeddings
            embeddings = embeddings + self.get_absolute_positions(
                self.position_embeddings, True, embeddings.shape[1], embeddings.shape[2]
            )
            # (batch_size, height, width, num_channels) -> (batch_size, num_channels, height, width)
            embeddings = embeddings.permute(0, 3, 1, 2)

        return embeddings


@torch.jit.script_if_tracing  # nn.functional.interpolate's `size` needs to be dynamic.
def get_rel_pos(q_size, k_size, rel_pos):
    """
    Get relative positional embeddings according to the relative positions of query and key sizes.

    Args:
        q_size (`int`):
            Size of query q.
        k_size (`int`):
            Size of key k.
        rel_pos (`torch.Tensor`):
            Relative position embeddings (num_embeddings, num_channels).

    Returns:
        Extracted positional embeddings according to relative positions.
    """
    max_rel_dist = int(2 * max(q_size, k_size) - 1)
    # Interpolate rel pos if needed.
    if rel_pos.shape[0] != max_rel_dist:
        # Interpolate rel position embeddings.
        rel_pos_resized = nn.functional.interpolate(
            rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
            size=max_rel_dist,
            mode="linear",
        )
        rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
    else:
        rel_pos_resized = rel_pos

    # Scale the coords with short length if shapes for q and k are different.
    q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
    k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
    relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)

    return rel_pos_resized[relative_coords.long()]


def add_decomposed_relative_positions(attn, queries, rel_pos_h, rel_pos_w, q_size, k_size):
    """
    Calculate decomposed Relative Positional Embeddings as introduced in
    [MViT2](https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py).

    Args:
        attn (`torch.Tensor`):
            Attention map.
        queries (`torch.Tensor`):
            Query q in the attention layer with shape (batch_size, queries_height * queries_width, num_channels).
        rel_pos_h (`torch.Tensor`):
            Relative position embeddings (Lh, num_channels) for height axis.
        rel_pos_w (`torch.Tensor`):
            Relative position embeddings (Lw, num_channels) for width axis.
        q_size (`Tuple[int]`):
            Spatial sequence size of query q with (queries_height, queries_width).
        k_size (`Tuple[int]`]):
            Spatial sequence size of key k with (keys_height, keys_width).

    Returns:
        attn (Tensor): attention map with added relative positional embeddings.
    """
    queries_height, queries_width = q_size
    keys_height, keys_width = k_size
    relative_height = get_rel_pos(queries_height, keys_height, rel_pos_h)
    relative_width = get_rel_pos(queries_width, keys_width, rel_pos_w)

    batch_size, _, dim = queries.shape
    r_q = queries.reshape(batch_size, queries_height, queries_width, dim)
    relative_height = torch.einsum("bhwc,hkc->bhwk", r_q, relative_height)
    relative_weight = torch.einsum("bhwc,wkc->bhwk", r_q, relative_width)

    attn = (
        attn.view(batch_size, queries_height, queries_width, keys_height, keys_width)
        + relative_height[:, :, :, :, None]
        + relative_weight[:, :, :, None, :]
    ).view(batch_size, queries_height * queries_width, keys_height * keys_width)

    return attn


class VitDetAttention(nn.Module):
    """Multi-head Attention block with relative position embeddings."""

    def __init__(self, config, input_size=None):
        """
        Args:
            config (`VitDetConfig`):
                Model configuration.
            input_size (`Tuple[int]`, *optional*):
                Input resolution, only required in case relative position embeddings are added.
        """
        super().__init__()

        dim = config.hidden_size
        num_heads = config.num_attention_heads

        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = head_dim**-0.5

        self.qkv = nn.Linear(dim, dim * 3, bias=config.qkv_bias)
        self.proj = nn.Linear(dim, dim)

        self.use_relative_position_embeddings = config.use_relative_position_embeddings
        if self.use_relative_position_embeddings:
            # initialize relative positional embeddings
            self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, head_dim))
            self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, head_dim))

    def forward(self, hidden_state, output_attentions=False):
        batch_size, height, width, _ = hidden_state.shape
        # qkv with shape (3, batch_size, num_heads, height * width, num_channels)
        qkv = self.qkv(hidden_state).reshape(batch_size, height * width, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
        # queries, keys and values have shape (batch_size * num_heads, height * width, num_channels)
        queries, keys, values = qkv.reshape(3, batch_size * self.num_heads, height * width, -1).unbind(0)

        attention_scores = (queries * self.scale) @ keys.transpose(-2, -1)

        if self.use_relative_position_embeddings:
            attention_scores = add_decomposed_relative_positions(
                attention_scores, queries, self.rel_pos_h, self.rel_pos_w, (height, width), (height, width)
            )

        attention_probs = attention_scores.softmax(dim=-1)

        hidden_state = attention_probs @ values
        hidden_state = hidden_state.view(batch_size, self.num_heads, height, width, -1)
        hidden_state = hidden_state.permute(0, 2, 3, 1, 4)
        hidden_state = hidden_state.reshape(batch_size, height, width, -1)
        hidden_state = self.proj(hidden_state)

        if output_attentions:
            attention_probs = attention_probs.reshape(
                batch_size, self.num_heads, attention_probs.shape[-2], attention_probs.shape[-1]
            )
            outputs = (hidden_state, attention_probs)
        else:
            outputs = (hidden_state,)

        return outputs


# Copied from transformers.models.beit.modeling_beit.drop_path
def drop_path(input: torch.Tensor, drop_prob: float = 0.0, training: bool = False) -> torch.Tensor:
    """
    Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

    Comment by Ross Wightman: This is the same as the DropConnect impl I created for EfficientNet, etc networks,
    however, the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for changing the
    layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use 'survival rate' as the
    argument.
    """
    if drop_prob == 0.0 or not training:
        return input
    keep_prob = 1 - drop_prob
    shape = (input.shape[0],) + (1,) * (input.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
    random_tensor = keep_prob + torch.rand(shape, dtype=input.dtype, device=input.device)
    random_tensor.floor_()  # binarize
    output = input.div(keep_prob) * random_tensor
    return output


# Copied from transformers.models.beit.modeling_beit.BeitDropPath
class VitDetDropPath(nn.Module):
    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""

    def __init__(self, drop_prob: Optional[float] = None) -> None:
        super().__init__()
        self.drop_prob = drop_prob

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        return drop_path(hidden_states, self.drop_prob, self.training)

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


class VitDetLayerNorm(nn.Module):
    """
    A LayerNorm variant, popularized by Transformers, that performs point-wise mean and variance normalization over the
    channel dimension for inputs that have shape (batch_size, channels, height, width).
    https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119
    """

    def __init__(self, normalized_shape, eps=1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(normalized_shape))
        self.bias = nn.Parameter(torch.zeros(normalized_shape))
        self.eps = eps
        self.normalized_shape = (normalized_shape,)

    def forward(self, x):
        u = x.mean(1, keepdim=True)
        s = (x - u).pow(2).mean(1, keepdim=True)
        x = (x - u) / torch.sqrt(s + self.eps)
        x = self.weight[:, None, None] * x + self.bias[:, None, None]
        return x


class VitDetResBottleneckBlock(nn.Module):
    """
    The standard bottleneck residual block without the last activation layer. It contains 3 conv layers with kernels
    1x1, 3x3, 1x1.
    """

    def __init__(self, config, in_channels, out_channels, bottleneck_channels):
        """
        Args:
            config (`VitDetConfig`):
                Model configuration.
            in_channels (`int`):
                Number of input channels.
            out_channels (`int`):
                Number of output channels.
            bottleneck_channels (`int`):
                Number of output channels for the 3x3 "bottleneck" conv layers.
        """
        super().__init__()
        self.conv1 = nn.Conv2d(in_channels, bottleneck_channels, 1, bias=False)
        self.norm1 = VitDetLayerNorm(bottleneck_channels)
        self.act1 = ACT2FN[config.hidden_act]

        self.conv2 = nn.Conv2d(bottleneck_channels, bottleneck_channels, 3, padding=1, bias=False)
        self.norm2 = VitDetLayerNorm(bottleneck_channels)
        self.act2 = ACT2FN[config.hidden_act]

        self.conv3 = nn.Conv2d(bottleneck_channels, out_channels, 1, bias=False)
        self.norm3 = VitDetLayerNorm(out_channels)

    def forward(self, x):
        out = x
        for layer in self.children():
            out = layer(out)

        out = x + out
        return out


class VitDetMlp(nn.Module):
    def __init__(self, config, in_features: int, hidden_features: int) -> None:
        super().__init__()
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = ACT2FN[config.hidden_act]
        self.fc2 = nn.Linear(hidden_features, in_features)
        self.drop = nn.Dropout(config.dropout_prob)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)

        return x


def window_partition(hidden_state, window_size):
    """
    Partition into non-overlapping windows with padding if needed.

    Args:
        hidden_state (`torch.Tensor`):
            Input tokens with [batch_size, height, width, num_channels].
        window_size (`int`):
            Window size.

    Returns:
        `tuple(torch.FloatTensor)` comprising various elements:
        - windows: windows after partition with [batch_size * num_windows, window_size, window_size, num_channels].
        - (padded_height, padded_width): padded height and width before partition
    """
    batch_size, height, width, num_channels = hidden_state.shape

    pad_height = (window_size - height % window_size) % window_size
    pad_width = (window_size - width % window_size) % window_size

    # Noop in case pad_width == 0 and pad_height == 0.
    hidden_state = nn.functional.pad(hidden_state, (0, 0, 0, pad_width, 0, pad_height))

    padded_height, padded_width = height + pad_height, width + pad_width

    hidden_state = hidden_state.view(
        batch_size, padded_height // window_size, window_size, padded_width // window_size, window_size, num_channels
    )
    windows = hidden_state.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, num_channels)
    return windows, (padded_height, padded_width)


def window_unpartition(windows, window_size, pad_height_width, height_width):
    """
    Window unpartition into original sequences and removing padding.

    Args:
        windows (`torch.Tensor`):
            Input tokens with [batch_size * num_windows, window_size, window_size, num_channels].
        window_size (`int`):
            Window size.
        pad_height_width (`Tuple[int]`):
            Padded height and width (padded_height, padded_width).
        height_width (`Tuple[int]`):
            Original height and width before padding.

    Returns:
        hidden_state: unpartitioned sequences with [batch_size, height, width, num_channels].
    """
    padded_height, padded_width = pad_height_width
    height, width = height_width
    batch_size = windows.shape[0] // (padded_height * padded_width // window_size // window_size)
    hidden_state = windows.view(
        batch_size, padded_height // window_size, padded_width // window_size, window_size, window_size, -1
    )
    hidden_state = hidden_state.permute(0, 1, 3, 2, 4, 5).contiguous()
    hidden_state = hidden_state.view(batch_size, padded_height, padded_width, -1)

    # We always have height <= padded_height and width <= padded_width
    hidden_state = hidden_state[:, :height, :width, :].contiguous()
    return hidden_state


class VitDetLayer(nn.Module):
    """This corresponds to the Block class in the original implementation."""

    def __init__(
        self, config: VitDetConfig, drop_path_rate: float = 0, window_size: int = 0, use_residual_block: bool = False
    ) -> None:
        super().__init__()

        dim = config.hidden_size
        input_size = (config.image_size // config.patch_size, config.image_size // config.patch_size)

        self.norm1 = nn.LayerNorm(dim, eps=config.layer_norm_eps)
        self.attention = VitDetAttention(
            config, input_size=input_size if window_size == 0 else (window_size, window_size)
        )

        self.drop_path = VitDetDropPath(drop_path_rate) if drop_path_rate > 0.0 else nn.Identity()
        self.norm2 = nn.LayerNorm(dim, eps=config.layer_norm_eps)
        self.mlp = VitDetMlp(config=config, in_features=dim, hidden_features=int(dim * config.mlp_ratio))

        self.window_size = window_size

        self.use_residual_block = use_residual_block
        if self.use_residual_block:
            # Use a residual block with bottleneck channel as dim // 2
            self.residual = VitDetResBottleneckBlock(
                config=config,
                in_channels=dim,
                out_channels=dim,
                bottleneck_channels=dim // 2,
            )

    def forward(
        self,
        hidden_states: torch.Tensor,
        head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
        hidden_states = hidden_states.permute(0, 2, 3, 1)

        shortcut = hidden_states

        hidden_states = self.norm1(hidden_states)

        # Window partition
        if self.window_size > 0:
            height, width = hidden_states.shape[1], hidden_states.shape[2]
            hidden_states, pad_height_width = window_partition(hidden_states, self.window_size)

        self_attention_outputs = self.attention(
            hidden_states,
            output_attentions=output_attentions,
        )
        hidden_states = self_attention_outputs[0]
        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights

        # Reverse window partition
        if self.window_size > 0:
            hidden_states = window_unpartition(hidden_states, self.window_size, pad_height_width, (height, width))

        # first residual connection
        hidden_states = shortcut + self.drop_path(hidden_states)

        hidden_states = hidden_states + self.drop_path(self.mlp(self.norm2(hidden_states)))

        hidden_states = hidden_states.permute(0, 3, 1, 2)

        if self.use_residual_block:
            hidden_states = self.residual(hidden_states)

        outputs = (hidden_states,) + outputs

        return outputs


class VitDetEncoder(nn.Module):
    def __init__(self, config: VitDetConfig) -> None:
        super().__init__()
        self.config = config
        depth = config.num_hidden_layers

        # stochastic depth decay rule
        drop_path_rate = [x.item() for x in torch.linspace(0, config.drop_path_rate, depth)]

        layers = []
        for i in range(depth):
            layers.append(
                VitDetLayer(
                    config,
                    drop_path_rate=drop_path_rate[i],
                    window_size=config.window_size if i in config.window_block_indices else 0,
                    use_residual_block=i in config.residual_block_indices,
                )
            )

        self.layer = nn.ModuleList(layers)
        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.Tensor,
        head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
        output_hidden_states: bool = False,
        return_dict: bool = True,
    ) -> Union[tuple, BaseModelOutput]:
        all_hidden_states = () if output_hidden_states else None
        all_self_attentions = () if output_attentions else None

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

            layer_head_mask = head_mask[i] if head_mask is not None else None

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    layer_module.__call__,
                    hidden_states,
                    layer_head_mask,
                    output_attentions,
                )
            else:
                layer_outputs = layer_module(hidden_states, layer_head_mask, output_attentions)

            hidden_states = layer_outputs[0]

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

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

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


def caffe2_msra_fill(module: nn.Module) -> None:
    """
    Initialize `module.weight` using the "MSRAFill" implemented in Caffe2. Also initializes `module.bias` to 0.

    Source: https://detectron2.readthedocs.io/en/latest/_modules/fvcore/nn/weight_init.html.

    Args:
        module (torch.nn.Module): module to initialize.
    """
    nn.init.kaiming_normal_(module.weight, mode="fan_out", nonlinearity="relu")
    if module.bias is not None:
        nn.init.constant_(module.bias, 0)


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

    config_class = VitDetConfig
    base_model_prefix = "vitdet"
    main_input_name = "pixel_values"
    supports_gradient_checkpointing = True
    _no_split_modules = []

    def _init_weights(self, module: Union[nn.Linear, nn.Conv2d, nn.LayerNorm]) -> None:
        """Initialize the weights"""
        if isinstance(module, (nn.Linear, nn.Conv2d)):
            # Upcast the input in `fp32` and cast it back to desired `dtype` to avoid
            # `trunc_normal_cpu` not implemented in `half` issues
            module.weight.data = nn.init.trunc_normal_(
                module.weight.data.to(torch.float32), mean=0.0, std=self.config.initializer_range
            ).to(module.weight.dtype)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)

        elif isinstance(module, VitDetEmbeddings):
            module.position_embeddings.data = nn.init.trunc_normal_(
                module.position_embeddings.data.to(torch.float32),
                mean=0.0,
                std=self.config.initializer_range,
            ).to(module.position_embeddings.dtype)

        elif isinstance(module, VitDetAttention) and self.config.use_relative_position_embeddings:
            module.rel_pos_h.data = nn.init.trunc_normal_(
                module.rel_pos_h.data.to(torch.float32),
                mean=0.0,
                std=self.config.initializer_range,
            )
            module.rel_pos_w.data = nn.init.trunc_normal_(
                module.rel_pos_w.data.to(torch.float32),
                mean=0.0,
                std=self.config.initializer_range,
            )

        elif isinstance(module, VitDetResBottleneckBlock):
            for layer in [module.conv1, module.conv2, module.conv3]:
                caffe2_msra_fill(layer)
            for layer in [module.norm1, module.norm2]:
                layer.weight.data.fill_(1.0)
                layer.bias.data.zero_()
            # zero init last norm layer.
            module.norm3.weight.data.zero_()
            module.norm3.bias.data.zero_()


VITDET_START_DOCSTRING = r"""
    This model is 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 ([`VitDetConfig`]): 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.
"""

VITDET_INPUTS_DOCSTRING = r"""
    Args:
        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
            Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See [`ViTImageProcessor.__call__`]
            for details.

        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        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.
"""


@add_start_docstrings(
    "The bare VitDet Transformer model outputting raw hidden-states without any specific head on top.",
    VITDET_START_DOCSTRING,
)
class VitDetModel(VitDetPreTrainedModel):
    def __init__(self, config: VitDetConfig):
        super().__init__(config)
        self.config = config

        self.embeddings = VitDetEmbeddings(config)
        self.encoder = VitDetEncoder(config)

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

    def get_input_embeddings(self) -> VitDetEmbeddings:
        return self.embeddings.projection

    def _prune_heads(self, heads_to_prune: Dict[int, List[int]]) -> None:
        """
        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
        class PreTrainedModel
        """
        for layer, heads in heads_to_prune.items():
            self.encoder.layer[layer].attention.prune_heads(heads)

    @add_start_docstrings_to_model_forward(VITDET_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=BaseModelOutput, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        pixel_values: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutput]:
        """
        Returns:

        Examples:

        ```python
        >>> from transformers import VitDetConfig, VitDetModel
        >>> import torch

        >>> config = VitDetConfig()
        >>> model = VitDetModel(config)

        >>> pixel_values = torch.randn(1, 3, 224, 224)

        >>> with torch.no_grad():
        ...     outputs = model(pixel_values)

        >>> last_hidden_states = outputs.last_hidden_state
        >>> list(last_hidden_states.shape)
        [1, 768, 14, 14]
        ```"""
        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")

        # Prepare head mask if needed
        # 1.0 in head_mask indicate we keep the head
        # attention_probs has shape bsz x n_heads x N x N
        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)

        embedding_output = self.embeddings(pixel_values)

        encoder_outputs = self.encoder(
            embedding_output,
            head_mask=head_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        sequence_output = encoder_outputs[0]

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

        return BaseModelOutput(
            last_hidden_state=sequence_output,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
        )


@add_start_docstrings(
    """
    ViTDet backbone, to be used with frameworks like Mask R-CNN.
    """,
    VITDET_START_DOCSTRING,
)
class VitDetBackbone(VitDetPreTrainedModel, BackboneMixin):
    def __init__(self, config):
        super().__init__(config)
        super()._init_backbone(config)

        self.embeddings = VitDetEmbeddings(config)
        self.encoder = VitDetEncoder(config)
        self.num_features = [config.hidden_size for _ in range(config.num_hidden_layers + 1)]

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

    def get_input_embeddings(self) -> VitDetEmbeddings:
        return self.embeddings.projection

    @add_start_docstrings_to_model_forward(VITDET_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=BackboneOutput, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        pixel_values: torch.Tensor,
        output_hidden_states: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> BackboneOutput:
        """
        Returns:

        Examples:

        ```python
        >>> from transformers import VitDetConfig, VitDetBackbone
        >>> import torch

        >>> config = VitDetConfig()
        >>> model = VitDetBackbone(config)

        >>> pixel_values = torch.randn(1, 3, 224, 224)

        >>> with torch.no_grad():
        ...     outputs = model(pixel_values)

        >>> feature_maps = outputs.feature_maps
        >>> list(feature_maps[-1].shape)
        [1, 768, 14, 14]
        ```"""
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions

        embedding_output = self.embeddings(pixel_values)

        outputs = self.encoder(
            embedding_output,
            output_hidden_states=True,
            output_attentions=output_attentions,
            return_dict=return_dict,
        )

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

        feature_maps = ()
        for stage, hidden_state in zip(self.stage_names, hidden_states):
            if stage in self.out_features:
                feature_maps += (hidden_state,)

        if not return_dict:
            if output_hidden_states:
                output = (feature_maps,) + outputs[1:]
            else:
                output = (feature_maps,) + outputs[2:]
            return output

        return BackboneOutput(
            feature_maps=feature_maps,
            hidden_states=outputs.hidden_states if output_hidden_states else None,
            attentions=outputs.attentions,
        )