# coding=utf-8
# Copyright 2022 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 Donut Swin Transformer model.

This implementation is identical to a regular Swin Transformer, without final layer norm on top of the final hidden
states."""

import collections.abc
import math
from dataclasses import dataclass
from typing import Optional, Union

import torch
from torch import nn

from ...activations import ACT2FN
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_utils import PreTrainedModel
from ...pytorch_utils import find_pruneable_heads_and_indices, meshgrid, prune_linear_layer
from ...utils import ModelOutput, auto_docstring, logging, torch_int
from .configuration_donut_swin import DonutSwinConfig


logger = logging.get_logger(__name__)


@dataclass
@auto_docstring(
    custom_intro="""
    DonutSwin encoder's outputs, with potential hidden states and attentions.
    """
)
# Copied from transformers.models.swin.modeling_swin.SwinEncoderOutput with Swin->DonutSwin
class DonutSwinEncoderOutput(ModelOutput):
    r"""
    reshaped_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 + one for the output of each stage) of
        shape `(batch_size, hidden_size, height, width)`.

        Hidden-states of the model at the output of each layer plus the initial embedding outputs reshaped to
        include the spatial dimensions.
    """

    last_hidden_state: Optional[torch.FloatTensor] = None
    hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None
    attentions: Optional[tuple[torch.FloatTensor, ...]] = None
    reshaped_hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None


@dataclass
@auto_docstring(
    custom_intro="""
    DonutSwin model's outputs that also contains a pooling of the last hidden states.
    """
)
# Copied from transformers.models.swin.modeling_swin.SwinModelOutput with Swin->DonutSwin
class DonutSwinModelOutput(ModelOutput):
    r"""
    pooler_output (`torch.FloatTensor` of shape `(batch_size, hidden_size)`, *optional*, returned when `add_pooling_layer=True` is passed):
        Average pooling of the last layer hidden-state.
    reshaped_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 + one for the output of each stage) of
        shape `(batch_size, hidden_size, height, width)`.

        Hidden-states of the model at the output of each layer plus the initial embedding outputs reshaped to
        include the spatial dimensions.
    """

    last_hidden_state: Optional[torch.FloatTensor] = None
    pooler_output: Optional[torch.FloatTensor] = None
    hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None
    attentions: Optional[tuple[torch.FloatTensor, ...]] = None
    reshaped_hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None


@dataclass
@auto_docstring(
    custom_intro="""
    DonutSwin outputs for image classification.
    """
)
# Copied from transformers.models.swin.modeling_swin.SwinImageClassifierOutput with Swin->DonutSwin
class DonutSwinImageClassifierOutput(ModelOutput):
    r"""
    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
        Classification (or regression if config.num_labels==1) loss.
    logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels)`):
        Classification (or regression if config.num_labels==1) scores (before SoftMax).
    reshaped_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 + one for the output of each stage) of
        shape `(batch_size, hidden_size, height, width)`.

        Hidden-states of the model at the output of each layer plus the initial embedding outputs reshaped to
        include the spatial dimensions.
    """

    loss: Optional[torch.FloatTensor] = None
    logits: Optional[torch.FloatTensor] = None
    hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None
    attentions: Optional[tuple[torch.FloatTensor, ...]] = None
    reshaped_hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None


# Copied from transformers.models.swin.modeling_swin.window_partition
def window_partition(input_feature, window_size):
    """
    Partitions the given input into windows.
    """
    batch_size, height, width, num_channels = input_feature.shape
    input_feature = input_feature.view(
        batch_size, height // window_size, window_size, width // window_size, window_size, num_channels
    )
    windows = input_feature.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, num_channels)
    return windows


# Copied from transformers.models.swin.modeling_swin.window_reverse
def window_reverse(windows, window_size, height, width):
    """
    Merges windows to produce higher resolution features.
    """
    num_channels = windows.shape[-1]
    windows = windows.view(-1, height // window_size, width // window_size, window_size, window_size, num_channels)
    windows = windows.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, height, width, num_channels)
    return windows


# Copied from transformers.models.swin.modeling_swin.SwinEmbeddings with Swin->DonutSwin
class DonutSwinEmbeddings(nn.Module):
    """
    Construct the patch and position embeddings. Optionally, also the mask token.
    """

    def __init__(self, config, use_mask_token=False):
        super().__init__()

        self.patch_embeddings = DonutSwinPatchEmbeddings(config)
        num_patches = self.patch_embeddings.num_patches
        self.patch_grid = self.patch_embeddings.grid_size
        self.mask_token = nn.Parameter(torch.zeros(1, 1, config.embed_dim)) if use_mask_token else None

        if config.use_absolute_embeddings:
            self.position_embeddings = nn.Parameter(torch.zeros(1, num_patches + 1, config.embed_dim))
        else:
            self.position_embeddings = None

        self.norm = nn.LayerNorm(config.embed_dim)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.patch_size = config.patch_size
        self.config = config

    # Copied from transformers.models.vit.modeling_vit.ViTEmbeddings.interpolate_pos_encoding
    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
        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]

        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: Optional[torch.FloatTensor],
        bool_masked_pos: Optional[torch.BoolTensor] = None,
        interpolate_pos_encoding: bool = False,
    ) -> tuple[torch.Tensor]:
        _, num_channels, height, width = pixel_values.shape
        embeddings, output_dimensions = self.patch_embeddings(pixel_values)
        embeddings = self.norm(embeddings)
        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 * (1.0 - mask) + mask_tokens * mask

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

        embeddings = self.dropout(embeddings)

        return embeddings, output_dimensions


# Copied from transformers.models.swin.modeling_swin.SwinPatchEmbeddings with Swin->DonutSwin
class DonutSwinPatchEmbeddings(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):
        super().__init__()
        image_size, patch_size = config.image_size, config.patch_size
        num_channels, hidden_size = config.num_channels, config.embed_dim
        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
        self.grid_size = (image_size[0] // patch_size[0], image_size[1] // patch_size[1])

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

    def maybe_pad(self, pixel_values, height, width):
        if width % self.patch_size[1] != 0:
            pad_values = (0, self.patch_size[1] - width % self.patch_size[1])
            pixel_values = nn.functional.pad(pixel_values, pad_values)
        if height % self.patch_size[0] != 0:
            pad_values = (0, 0, 0, self.patch_size[0] - height % self.patch_size[0])
            pixel_values = nn.functional.pad(pixel_values, pad_values)
        return pixel_values

    def forward(self, pixel_values: Optional[torch.FloatTensor]) -> tuple[torch.Tensor, tuple[int]]:
        _, num_channels, height, width = pixel_values.shape
        # pad the input to be divisible by self.patch_size, if needed
        pixel_values = self.maybe_pad(pixel_values, height, width)
        embeddings = self.projection(pixel_values)
        _, _, height, width = embeddings.shape
        output_dimensions = (height, width)
        embeddings = embeddings.flatten(2).transpose(1, 2)

        return embeddings, output_dimensions


# Copied from transformers.models.swin.modeling_swin.SwinPatchMerging
class DonutSwinPatchMerging(nn.Module):
    """
    Patch Merging Layer.

    Args:
        input_resolution (`tuple[int]`):
            Resolution of input feature.
        dim (`int`):
            Number of input channels.
        norm_layer (`nn.Module`, *optional*, defaults to `nn.LayerNorm`):
            Normalization layer class.
    """

    def __init__(self, input_resolution: tuple[int], dim: int, norm_layer: nn.Module = nn.LayerNorm) -> None:
        super().__init__()
        self.input_resolution = input_resolution
        self.dim = dim
        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
        self.norm = norm_layer(4 * dim)

    def maybe_pad(self, input_feature, height, width):
        should_pad = (height % 2 == 1) or (width % 2 == 1)
        if should_pad:
            pad_values = (0, 0, 0, width % 2, 0, height % 2)
            input_feature = nn.functional.pad(input_feature, pad_values)

        return input_feature

    def forward(self, input_feature: torch.Tensor, input_dimensions: tuple[int, int]) -> torch.Tensor:
        height, width = input_dimensions
        # `dim` is height * width
        batch_size, dim, num_channels = input_feature.shape

        input_feature = input_feature.view(batch_size, height, width, num_channels)
        # pad input to be divisible by width and height, if needed
        input_feature = self.maybe_pad(input_feature, height, width)
        # [batch_size, height/2, width/2, num_channels]
        input_feature_0 = input_feature[:, 0::2, 0::2, :]
        # [batch_size, height/2, width/2, num_channels]
        input_feature_1 = input_feature[:, 1::2, 0::2, :]
        # [batch_size, height/2, width/2, num_channels]
        input_feature_2 = input_feature[:, 0::2, 1::2, :]
        # [batch_size, height/2, width/2, num_channels]
        input_feature_3 = input_feature[:, 1::2, 1::2, :]
        # batch_size height/2 width/2 4*num_channels
        input_feature = torch.cat([input_feature_0, input_feature_1, input_feature_2, input_feature_3], -1)
        input_feature = input_feature.view(batch_size, -1, 4 * num_channels)  # batch_size height/2*width/2 4*C

        input_feature = self.norm(input_feature)
        input_feature = self.reduction(input_feature)

        return input_feature


# 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.swin.modeling_swin.SwinDropPath
class DonutSwinDropPath(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 f"p={self.drop_prob}"


# Copied from transformers.models.swin.modeling_swin.SwinSelfAttention with Swin->DonutSwin
class DonutSwinSelfAttention(nn.Module):
    def __init__(self, config, dim, num_heads, window_size):
        super().__init__()
        if dim % num_heads != 0:
            raise ValueError(
                f"The hidden size ({dim}) is not a multiple of the number of attention heads ({num_heads})"
            )

        self.num_attention_heads = num_heads
        self.attention_head_size = int(dim / num_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size
        self.window_size = (
            window_size if isinstance(window_size, collections.abc.Iterable) else (window_size, window_size)
        )

        self.relative_position_bias_table = nn.Parameter(
            torch.zeros((2 * self.window_size[0] - 1) * (2 * self.window_size[1] - 1), num_heads)
        )

        # get pair-wise relative position index for each token inside the window
        coords_h = torch.arange(self.window_size[0])
        coords_w = torch.arange(self.window_size[1])
        coords = torch.stack(meshgrid([coords_h, coords_w], indexing="ij"))
        coords_flatten = torch.flatten(coords, 1)
        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]
        relative_coords = relative_coords.permute(1, 2, 0).contiguous()
        relative_coords[:, :, 0] += self.window_size[0] - 1
        relative_coords[:, :, 1] += self.window_size[1] - 1
        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
        relative_position_index = relative_coords.sum(-1)
        self.register_buffer("relative_position_index", relative_position_index)

        self.query = nn.Linear(self.all_head_size, self.all_head_size, bias=config.qkv_bias)
        self.key = nn.Linear(self.all_head_size, self.all_head_size, bias=config.qkv_bias)
        self.value = nn.Linear(self.all_head_size, self.all_head_size, bias=config.qkv_bias)

        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = False,
    ) -> tuple[torch.Tensor]:
        batch_size, dim, num_channels = hidden_states.shape
        hidden_shape = (batch_size, dim, -1, self.attention_head_size)

        query_layer = self.query(hidden_states).view(hidden_shape).transpose(1, 2)
        key_layer = self.key(hidden_states).view(hidden_shape).transpose(1, 2)
        value_layer = self.value(hidden_states).view(hidden_shape).transpose(1, 2)

        # Take the dot product between "query" and "key" to get the raw attention scores.
        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))

        attention_scores = attention_scores / math.sqrt(self.attention_head_size)

        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)]
        relative_position_bias = relative_position_bias.view(
            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1
        )

        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()
        attention_scores = attention_scores + relative_position_bias.unsqueeze(0)

        if attention_mask is not None:
            # Apply the attention mask is (precomputed for all layers in DonutSwinModel forward() function)
            mask_shape = attention_mask.shape[0]
            attention_scores = attention_scores.view(
                batch_size // mask_shape, mask_shape, self.num_attention_heads, dim, dim
            )
            attention_scores = attention_scores + attention_mask.unsqueeze(1).unsqueeze(0)
            attention_scores = attention_scores.view(-1, self.num_attention_heads, dim, dim)

        # Normalize the attention scores to probabilities.
        attention_probs = nn.functional.softmax(attention_scores, dim=-1)

        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.
        attention_probs = self.dropout(attention_probs)

        # Mask heads if we want to
        if head_mask is not None:
            attention_probs = attention_probs * head_mask

        context_layer = torch.matmul(attention_probs, value_layer)
        context_layer = context_layer