"""
An official Pytorch impl of `Transcending the Limit of Local Window:
Advanced Super-Resolution Transformer with Adaptive Token Dictionary`.

Arxiv: 'https://arxiv.org/abs/2401.08209'
"""

from __future__ import annotations

import math

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

from spandrel.util import store_hyperparameters
from spandrel.util.timm import to_2tuple, trunc_normal_


# Shuffle operation for Categorization and UnCategorization operations.
def index_reverse(index):
    index_r = torch.zeros_like(index)
    ind = torch.arange(0, index.shape[-1]).to(index.device)
    for i in range(index.shape[0]):
        index_r[i, index[i, :]] = ind
    return index_r


def feature_shuffle(x, index):
    dim = index.dim()
    assert (
        x.shape[:dim] == index.shape
    ), f"x ({x.shape}) and index ({index.shape}) shape incompatible"

    for _ in range(x.dim() - index.dim()):
        index = index.unsqueeze(-1)
    index = index.expand(x.shape)

    shuffled_x = torch.gather(x, dim=dim - 1, index=index)
    return shuffled_x


class dwconv(nn.Module):
    def __init__(self, hidden_features, kernel_size=5):
        super().__init__()
        self.depthwise_conv = nn.Sequential(
            nn.Conv2d(
                hidden_features,
                hidden_features,
                kernel_size=kernel_size,
                stride=1,
                padding=(kernel_size - 1) // 2,
                dilation=1,
                groups=hidden_features,
            ),
            nn.GELU(),
        )
        self.hidden_features = hidden_features

    def forward(self, x, x_size):
        x = (
            x.transpose(1, 2)
            .view(x.shape[0], self.hidden_features, x_size[0], x_size[1])
            .contiguous()
        )  # b Ph*Pw c
        x = self.depthwise_conv(x)
        x = x.flatten(2).transpose(1, 2).contiguous()
        return x


class ConvFFN(nn.Module):
    def __init__(
        self,
        in_features,
        hidden_features=None,
        out_features=None,
        kernel_size=5,
        act_layer=nn.GELU,
    ):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.dwconv = dwconv(hidden_features=hidden_features, kernel_size=kernel_size)
        self.fc2 = nn.Linear(hidden_features, out_features)

    def forward(self, x, x_size):
        x = self.fc1(x)
        x = self.act(x)
        x = x + self.dwconv(x, x_size)
        x = self.fc2(x)
        return x


def window_partition(x, window_size):
    """
    Args:
        x: (b, h, w, c)
        window_size (int): window size

    Returns:
        windows: (num_windows*b, window_size, window_size, c)
    """
    b, h, w, c = x.shape
    x = x.view(b, h // window_size, window_size, w // window_size, window_size, c)
    windows = (
        x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, c)
    )
    return windows


def window_reverse(windows, window_size, h, w):
    """
    Args:
        windows: (num_windows*b, window_size, window_size, c)
        window_size (int): Window size
        h (int): Height of image
        w (int): Width of image

    Returns:
        x: (b, h, w, c)
    """
    b = int(windows.shape[0] / (h * w / window_size / window_size))
    x = windows.view(
        b, h // window_size, w // window_size, window_size, window_size, -1
    )
    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(b, h, w, -1)
    return x


class WindowAttention(nn.Module):
    r"""
    Shifted Window-based Multi-head Self-Attention

    Args:
        dim (int): Number of input channels.
        window_size (tuple[int]): The height and width of the window.
        num_heads (int): Number of attention heads.
        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
    """

    def __init__(self, dim, window_size, num_heads, qkv_bias=True):
        super().__init__()
        self.dim = dim
        self.window_size = window_size  # Wh, Ww
        self.num_heads = num_heads
        self.qkv_bias = qkv_bias
        head_dim = dim // num_heads
        self.scale = head_dim**-0.5

        # define a parameter table of relative position bias
        self.relative_position_bias_table = nn.Parameter(  # type: ignore
            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)
        )  # 2*Wh-1 * 2*Ww-1, nH

        self.proj = nn.Linear(dim, dim)

        trunc_normal_(self.relative_position_bias_table, std=0.02)
        self.softmax = nn.Softmax(dim=-1)

    def forward(self, qkv, rpi, mask=None):
        r"""
        Args:
            qkv: Input query, key, and value tokens with shape of (num_windows*b, n, c*3)
            rpi: Relative position index
            mask (0/-inf):  Mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
        """
        b_, n, c3 = qkv.shape
        c = c3 // 3
        qkv = qkv.reshape(b_, n, 3, self.num_heads, c // self.num_heads).permute(
            2, 0, 3, 1, 4
        )
        q, k, v = (
            qkv[0],
            qkv[1],
            qkv[2],
        )  # make torchscript happy (cannot use tensor as tuple)

        q = q * self.scale
        attn = q @ k.transpose(-2, -1)

        relative_position_bias = self.relative_position_bias_table[rpi.view(-1)].view(
            self.window_size[0] * self.window_size[1],
            self.window_size[0] * self.window_size[1],
            -1,
        )  # Wh*Ww,Wh*Ww,nH
        relative_position_bias = relative_position_bias.permute(
            2, 0, 1
        ).contiguous()  # nH, Wh*Ww, Wh*Ww
        attn = attn + relative_position_bias.unsqueeze(0)

        if mask is not None:
            nw = mask.shape[0]
            attn = attn.view(b_ // nw, nw, self.num_heads, n, n) + mask.unsqueeze(
                1
            ).unsqueeze(0)
            attn = attn.view(-1, self.num_heads, n, n)
            attn = self.softmax(attn)
        else:
            attn = self.softmax(attn)

        x = (attn @ v).transpose(1, 2).reshape(b_, n, c)
        x = self.proj(x)
        return x

    def extra_repr(self) -> str:
        return f"dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}, qkv_bias={self.qkv_bias}"


class ATD_CA(nn.Module):
    r"""
    Adaptive Token Dictionary Cross-Attention.

    Args:
        dim (int): Number of input channels.
        input_resolution (tuple[int]): Input resolution.
        num_tokens (int): Number of tokens in external token dictionary. Default: 64
        reducted_dim (int, optional): Reducted dimension number for query and key matrix. Default: 4
        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
    """

    def __init__(
        self, dim, input_resolution, num_tokens=64, reducted_dim=10, qkv_bias=True
    ):
        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.num_tokens = num_tokens
        self.rc = reducted_dim
        self.qkv_bias = qkv_bias

        self.wq = nn.Linear(dim, reducted_dim, bias=qkv_bias)
        self.wk = nn.Linear(dim, reducted_dim, bias=qkv_bias)
        self.wv = nn.Linear(dim, dim, bias=qkv_bias)

        self.scale = nn.Parameter(  # type: ignore
            torch.ones([self.num_tokens]) * 0.5, requires_grad=True
        )
        self.softmax = nn.Softmax(dim=-1)

    def forward(self, x, td, x_size):
        r"""
        Args:
            x: input features with shape of (b, n, c)
            td: token dicitionary with shape of (b, m, c)
            x_size: size of the input x (h, w)
        """
        b, n, c = x.shape
        b, _m, c = td.shape

        # Q: b, n, c
        q = self.wq(x)
        # K: b, m, c
        k = self.wk(td)
        # V: b, m, c
        v = self.wv(td)

        # Q @ K^T
        attn = F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(
            -2, -1
        )  # b, n, n_tk
        scale = torch.clamp(self.scale, 0, 1)
        attn = attn * (1 + scale * np.log(self.num_tokens))
        attn = self.softmax(attn)

        # Attn * V
        x = (attn @ v).reshape(b, n, c)

        return x, attn


class AC_MSA(nn.Module):
    r"""
    Adaptive Category-based Multihead Self-Attention.

    Args:
        dim (int): Number of input channels.
        input_resolution (tuple[int]): Input resolution.
        num_tokens (int): Number of tokens in external dictionary. Default: 64
        num_heads (int): Number of attention heads. Default: 4
        category_size (int): Number of tokens in each group for global sparse attention. Default: 128
        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
    """

    def __init__(
        self,
        dim,
        input_resolution,
        num_tokens=64,
        num_heads=4,
        category_size=128,
        qkv_bias=True,
    ):
        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.num_tokens = num_tokens
        self.num_heads = num_heads
        self.category_size = category_size

        # self.wqkv = nn.Linear(dim, 3 * dim, bias=qkv_bias)
        self.proj = nn.Linear(dim, dim, bias=qkv_bias)

        self.logit_scale = nn.Parameter(  # type: ignore
            torch.log(10 * torch.ones((1, 1))), requires_grad=True
        )
        self.softmax = nn.Softmax(dim=-1)

    def forward(self, qkv, sim, x_size):
        """
        Args:
            x: input features with shape of (b, HW, c)
            mask: similarity map with shape of (b, HW, m)
            x_size: size of the input x
        """

        b, n, c3 = qkv.shape
        c = c3 // 3
        b, n, _m = sim.shape
        gs = min(n, self.category_size)  # group size
        ng = (n + gs - 1) // gs

        # classify features into groups based on similarity map (sim)
        tk_id = torch.argmax(sim, dim=-1, keepdim=False)
        # sort features by type
        _x_sort_values, x_sort_indices = torch.sort(tk_id, dim=-1, stable=False)
        x_sort_indices_reverse = index_reverse(x_sort_indices)
        shuffled_qkv = feature_shuffle(qkv, x_sort_indices)  # b, n, c3
        pad_n = ng * gs - n
        paded_qkv = torch.cat(
            (shuffled_qkv, torch.flip(shuffled_qkv[:, n - pad_n : n, :], dims=[1])),
            dim=1,
        )
        y = paded_qkv.reshape(b, -1, gs, c3)

        qkv = y.reshape(b, ng, gs, 3, self.num_heads, c // self.num_heads).permute(
            3, 0, 1, 4, 2, 5
        )  # 3, b, ng, nh, gs, c//nh
        q, k, v = qkv[0], qkv[1], qkv[2]

        # Q @ K^T
        attn = q @ k.transpose(-2, -1)  # b, ng, nh, gs, gs

        logit_scale = torch.clamp(
            self.logit_scale, max=torch.log(torch.tensor(1.0 / 0.01)).to(qkv.device)
        ).exp()
        attn = attn * logit_scale

        # softmax
        attn = self.softmax(attn)  # b, ng, nh, gs, gs

        # Attn * V
        y = (attn @ v).permute(0, 1, 3, 2, 4).reshape(b, n + pad_n, c)[:, :n, :]

        x = feature_shuffle(y, x_sort_indices_reverse)
        x = self.proj(x)

        return x


class ATDTransformerLayer(nn.Module):
    r"""
    ATD Transformer Layer

    Args:
        dim (int): Number of input channels.
        idx (int): Layer index.
        input_resolution (tuple[int]): Input resolution.
        num_heads (int): Number of attention heads.
        window_size (int): Window size.
        shift_size (int): Shift size for SW-MSA.
        category_size (int): Category size for AC-MSA.
        num_tokens (int): Token number for each token dictionary.
        reducted_dim (int): Reducted dimension number for query and key matrix.
        convffn_kernel_size (int): Convolutional kernel size for ConvFFN.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
        is_last (bool): True if this layer is the last of a ATD Block. Default: False
    """

    def __init__(
        self,
        dim,
        idx,
        input_resolution,
        num_heads,
        window_size,
        shift_size,
        category_size,
        num_tokens,
        reducted_dim,
        convffn_kernel_size,
        mlp_ratio,
        qkv_bias=True,
        act_layer=nn.GELU,
        norm_layer=nn.LayerNorm,
        is_last=False,
    ):
        super().__init__()

        self.dim = dim
        self.input_resolution = input_resolution
        self.num_heads = num_heads
        self.window_size = window_size
        self.shift_size = shift_size
        self.mlp_ratio = mlp_ratio
        self.convffn_kernel_size = convffn_kernel_size
        self.num_tokens = num_tokens
        self.softmax = nn.Softmax(dim=-1)
        self.lrelu = nn.LeakyReLU()
        self.sigmoid = nn.Sigmoid()
        self.reducted_dim = reducted_dim
        self.is_last = is_last

        self.norm1 = norm_layer(dim)
        self.norm2 = norm_layer(dim)
        if not is_last:
            self.norm3 = nn.InstanceNorm1d(num_tokens, affine=True)
            self.sigma = nn.Parameter(torch.zeros([num_tokens, 1]), requires_grad=True)  # type: ignore

        self.wqkv = nn.Linear(dim, 3 * dim, bias=qkv_bias)

        self.attn_win = WindowAttention(
            self.dim,
            window_size=to_2tuple(self.window_size),
            num_heads=num_heads,
            qkv_bias=qkv_bias,
        )
        self.attn_atd = ATD_CA(
            self.dim,
            input_resolution=input_resolution,
            qkv_bias=qkv_bias,
            num_tokens=num_tokens,
            reducted_dim=reducted_dim,
        )
        self.attn_aca = AC_MSA(
            self.dim,
            input_resolution=input_resolution,
            num_tokens=num_tokens,
            num_heads=num_heads,
            category_size=category_size,
            qkv_bias=qkv_bias,
        )

        mlp_hidden_dim = int(dim * mlp_ratio)
        self.convffn = ConvFFN(
            in_features=dim,
            hidden_features=mlp_hidden_dim,
            kernel_size=convffn_kernel_size,
            act_layer=act_layer,
        )

    def forward(self, x, td, x_size, params):
        h, w = x_size
        b, n, c = x.shape
        c3 = 3 * c

        shortcut = x
        x = self.norm1(x)
        qkv = self.wqkv(x)

        # ATD_CA
        x_atd, sim_atd = self.attn_atd(
            x, td, x_size
        )  # x_atd: (b, n, c)  sim_atd: (b, n,)

        # AC_MSA
        x_aca = self.attn_aca(qkv, sim_atd, x_size)

        # SW-MSA
        qkv = qkv.reshape(b, h, w, c3)

        # cyclic shift
        if self.shift_size > 0:
            shifted_qkv = torch.roll(
                qkv, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)
            )
            attn_mask = params["attn_mask"]
        else:
            shifted_qkv = qkv
            attn_mask = None

        # partition windows
        x_windows = window_partition(
            shifted_qkv, self.window_size
        )  # nw*b, window_size, window_size, c
        x_windows = x_windows.view(
            -1, self.window_size * self.window_size, c3
        )  # nw*b, window_size*window_size, c

        # W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
        attn_windows = self.attn_win(x_windows, rpi=params["rpi_sa"], mask=attn_mask)

        # merge windows
        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, c)
        shifted_x = window_reverse(attn_windows, self.window_size, h, w)  # b h' w' c

        # reverse cyclic shift
        if self.shift_size > 0:
            attn_x = torch.roll(
                shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)
            )
        else:
            attn_x = shifted_x
        x_win = attn_x
        x = shortcut + x_win.view(b, n, c) + x_atd + x_aca

        # FFN
        x = x + self.convffn(self.norm2(x), x_size)

        b, N, c = x.shape
        b, n, c = td.shape

        # Adaptive Token Refinement
        if not self.is_last:
            mask_soft = self.softmax(self.norm3(sim_atd.transpose(-1, -2)))
            mask_x = x.reshape(b, N, c)
            s = self.sigmoid(self.sigma)
            td = s * td + (1 - s) * torch.einsum("btn,bnc->btc", mask_soft, mask_x)

        return x, td


class PatchMerging(nn.Module):
    r"""Patch Merging Layer.

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

    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
        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 forward(self, x):
        """
        x: b, h*w, c
        """
        h, w = self.input_resolution
        b, seq_len, c = x.shape
        assert seq_len == h * w, "input feature has wrong size"
        assert h % 2 == 0 and w % 2 == 0, f"x size ({h}*{w}) are not even."

        x = x.view(b, h, w, c)

        x0 = x[:, 0::2, 0::2, :]  # b h/2 w/2 c
        x1 = x[:, 1::2, 0::2, :]  # b h/2 w/2 c
        x2 = x[:, 0::2, 1::2, :]  # b h/2 w/2 c
        x3 = x[:, 1::2, 1::2, :]  # b h/2 w/2 c
        x = torch.cat([x0, x1, x2, x3], -1)  # b h/2 w/2 4*c
        x = x.view(b, -1, 4 * c)  # b h/2*w/2 4*c

        x = self.norm(x)
        x = self.reduction(x)

        return x

    def extra_repr(self) -> str:
        return f"input_resolution={self.input_resolution}, dim={self.dim}"


class BasicBlock(nn.Module):
    """A basic ATD Block for one stage.

    Args:
        dim (int): Number of input channels.
        input_resolution (tuple[int]): Input resolution.
        idx (int): Block index.
        depth (int): Number of blocks.
        num_heads (int): Number of attention heads.
        window_size (int): Local window size.
        category_size (int): Category size for AC-MSA.
        num_tokens (int): Token number for each token dictionary.
        reducted_dim (int): Reducted dimension number for query and key matrix.
        convffn_kernel_size (int): Convolutional kernel size for ConvFFN.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
    """

    def __init__(
        self,
        dim,
        input_resolution,
        idx,
        depth,
        num_heads,
        window_size,
        category_size,
        num_tokens,
        convffn_kernel_size,
        reducted_dim,
        mlp_ratio=4.0,
        qkv_bias=True,
        norm_layer=nn.LayerNorm,
        downsample=None,
    ):
        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.depth = depth
        self.idx = idx

        self.layers = nn.ModuleList()
        for i in range(depth):
            self.layers.append(
                ATDTransformerLayer(
                    dim=dim,
                    idx=i,
                    input_resolution=input_resolution,
                    num_heads=num_heads,
                    window_size=window_size,
                    shift_size=0 if (i % 2 == 0) else window_size // 2,
                    category_size=category_size,
                    num_tokens=num_tokens,
                    convffn_kernel_size=convffn_kernel_size,
                    reducted_dim=reducted_dim,
                    mlp_ratio=mlp_ratio,
                    qkv_bias=qkv_bias,
                    norm_layer=norm_layer,
                    is_last=i == depth - 1,
                )
            )

        # patch merging layer
        if downsample is not None:
            self.downsample = downsample(
                input_resolution, dim=dim, norm_layer=norm_layer
            )
        else:
            self.downsample = None

        # Token Dictionary
        self.td = nn.Parameter(torch.randn([num_tokens, dim]), requires_grad=True)  # type: ignore

    def forward(self, x, x_size, params):
        b, _n, _c = x.shape
        td = self.td.repeat([b, 1, 1])
        for layer in self.layers:
            x, td = layer(x, td, x_size, params)
        if self.downsample is not None:
            x = self.downsample(x)
        return x

    def extra_repr(self) -> str:
        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"


class ATDB(nn.Module):
    """Adaptive Token Dictionary Block (ATDB).

    Args:
        dim (int): Number of input channels.
        input_resolution (tuple[int]): Input resolution.
        depth (int): Number of blocks.
        num_heads (int): Number of attention heads.
        window_size (int): Local window size.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
        img_size: Input image size.
        patch_size: Patch size.
        resi_connection: The convolutional block before residual connection.
    """

    def __init__(
        self,
        dim,
        idx,
        input_resolution,
        depth,
        num_heads,
        window_size,
        category_size,
        num_tokens,
        reducted_dim,
        convffn_kernel_size,
        mlp_ratio,
        qkv_bias=True,
        norm_layer=nn.LayerNorm,
        downsample=None,
        img_size=224,
        patch_size=4,
        resi_connection="1conv",
    ):
        super().__init__()

        self.dim = dim
        self.input_resolution = input_resolution

        self.patch_embed = PatchEmbed(
            img_size=img_size,
            patch_size=patch_size,
            in_chans=0,
            embed_dim=dim,
            norm_layer=None,
        )

        self.patch_unembed = PatchUnEmbed(
            img_size=img_size,
            patch_size=patch_size,
            in_chans=0,
            embed_dim=dim,
            norm_layer=None,
        )

        self.residual_group = BasicBlock(
            dim=dim,
            input_resolution=input_resolution,
            idx=idx,
            depth=depth,
            num_heads=num_heads,
            window_size=window_size,
            num_tokens=num_tokens,
            category_size=category_size,
            reducted_dim=reducted_dim,
            convffn_kernel_size=convffn_kernel_size,
            mlp_ratio=mlp_ratio,
            qkv_bias=qkv_bias,
            norm_layer=norm_layer,
            downsample=downsample,
        )

        if resi_connection == "1conv":
            self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
        elif resi_connection == "3conv":
            # to save parameters and memory
            self.conv = nn.Sequential(
                nn.Conv2d(dim, dim // 4, 3, 1, 1),
                nn.LeakyReLU(negative_slope=0.2, inplace=True),
                nn.Conv2d(dim // 4, dim // 4, 1, 1, 0),
                nn.LeakyReLU(negative_slope=0.2, inplace=True),
                nn.Conv2d(dim // 4, dim, 3, 1, 1),
            )

    def forward(self, x, x_size, params):
        return (
            self.patch_embed(
                self.conv(
                    self.patch_unembed(self.residual_group(x, x_size, params), x_size)
                )
            )
            + x
        )


class PatchEmbed(nn.Module):
    r"""Image to Patch Embedding

    Args:
        img_size (int): Image size.  Default: 224.
        patch_size (int): Patch token size. Default: 4.
        in_chans (int): Number of input image channels. Default: 3.
        embed_dim (int): Number of linear projection output channels. Default: 96.
        norm_layer (nn.Module, optional): Normalization layer. Default: None
    """

    def __init__(
        self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None
    ):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        patches_resolution = [
            img_size[0] // patch_size[0],
            img_size[1] // patch_size[1],
        ]
        self.img_size = img_size
        self.patch_size = patch_size
        self.patches_resolution = patches_resolution
        self.num_patches = patches_resolution[0] * patches_resolution[1]

        self.in_chans = in_chans
        self.embed_dim = embed_dim

        if norm_layer is not None:
            self.norm = norm_layer(embed_dim)
        else:
            self.norm = None

    def forward(self, x):
        x = x.flatten(2).transpose(1, 2)  # b Ph*Pw c
        if self.norm is not None:
            x = self.norm(x)
        return x


class PatchUnEmbed(nn.Module):
    r"""Image to Patch Unembedding

    Args:
        img_size (int): Image size.  Default: 224.
        patch_size (int): Patch token size. Default: 4.
        in_chans (int): Number of input image channels. Default: 3.
        embed_dim (int): Number of linear projection output channels. Default: 96.
        norm_layer (nn.Module, optional): Normalization layer. Default: None
    """

    def __init__(
        self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None
    ):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        patches_resolution = [
            img_size[0] // patch_size[0],
            img_size[1] // patch_size[1],
        ]
        self.img_size = img_size
        self.patch_size = patch_size
        self.patches_resolution = patches_resolution
        self.num_patches = patches_resolution[0] * patches_resolution[1]

        self.in_chans = in_chans
        self.embed_dim = embed_dim

    def forward(self, x, x_size):
        x = x.transpose(1, 2).view(
            x.shape[0], self.embed_dim, x_size[0], x_size[1]
        )  # b Ph*Pw c
        return x


class Upsample(nn.Sequential):
    """Upsample module.

    Args:
        scale (int): Scale factor. Supported scales: 2^n and 3.
        num_feat (int): Channel number of intermediate features.
    """

    def __init__(self, scale, num_feat):
        m = []
        self.scale = scale
        self.num_feat = num_feat
        if (scale & (scale - 1)) == 0:  # scale = 2^n
            for _ in range(int(math.log(scale, 2))):
                m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
                m.append(nn.PixelShuffle(2))
        elif scale == 3:
            m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
            m.append(nn.PixelShuffle(3))
        else:
            raise ValueError(
                f"scale {scale} is not supported. Supported scales: 2^n and 3."
            )
        super().__init__(*m)


class UpsampleOneStep(nn.Sequential):
    """UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle)
       Used in lightweight SR to save parameters.

    Args:
        scale (int): Scale factor. Supported scales: 2^n and 3.
        num_feat (int): Channel number of intermediate features.

    """

    def __init__(self, scale, num_feat, num_out_ch, input_resolution=None):
        self.num_feat = num_feat
        self.input_resolution = input_resolution
        m = []
        m.append(nn.Conv2d(num_feat, (scale**2) * num_out_ch, 3, 1, 1))
        m.append(nn.PixelShuffle(scale))
        super().__init__(*m)


@store_hyperparameters()
class ATD(nn.Module):
    r"""
    ATD

    A PyTorch impl of : `Transcending the Limit of Local Window: Advanced Super-Resolution Transformer with Adaptive Token Dictionary`.

    Args:
        img_size (int | tuple(int)): Input image size. Default 64
        patch_size (int | tuple(int)): Patch size. Default: 1
        in_chans (int): Number of input image channels. Default: 3
        embed_dim (int): Patch embedding dimension. Default: 96
        depths (tuple(int)): Depth of each Swin Transformer layer.
        num_heads (tuple(int)): Number of attention heads in different layers.
        window_size (int): Window size. Default: 7
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 2
        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
        patch_norm (bool): If True, add normalization after patch embedding. Default: True
        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
        upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction
        img_range: Image range. 1. or 255.
        upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None
        resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
    """

    hyperparameters = {}

    def __init__(
        self,
        *,
        img_size=64,
        patch_size=1,
        in_chans=3,
        embed_dim=90,
        depths=(6, 6, 6, 6),
        num_heads=(6, 6, 6, 6),
        window_size=8,
        category_size=256,
        num_tokens=64,
        reducted_dim=4,
        convffn_kernel_size=5,
        mlp_ratio=2.0,
        qkv_bias=True,
        norm_layer=nn.LayerNorm,
        ape=False,
        patch_norm=True,
        upscale=1,
        img_range=1.0,
        upsampler="",
        resi_connection="1conv",
        norm=True,
    ):
        super().__init__()
        num_in_ch = in_chans
        num_out_ch = in_chans
        num_feat = 64
        self.img_range = img_range

        self.no_norm: torch.Tensor | None
        if not norm:
            self.register_buffer("no_norm", torch.zeros(1))
        else:
            self.no_norm = None

        if in_chans == 3:
            rgb_mean = (0.4488, 0.4371, 0.4040)
            self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
        else:
            self.mean = torch.zeros(1, 1, 1, 1)
        self.upscale = upscale
        self.upsampler = upsampler

        # ------------------------- 1, shallow feature extraction ------------------------- #
        self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1)

        # ------------------------- 2, deep feature extraction ------------------------- #
        self.num_layers = len(depths)
        self.embed_dim = embed_dim
        self.ape = ape
        self.patch_norm = patch_norm
        self.num_features = embed_dim
        self.mlp_ratio = mlp_ratio
        self.window_size = window_size

        # split image into non-overlapping patches
        self.patch_embed = PatchEmbed(
            img_size=img_size,
            patch_size=patch_size,
            in_chans=embed_dim,
            embed_dim=embed_dim,
            norm_layer=norm_layer if self.patch_norm else None,
        )
        num_patches = self.patch_embed.num_patches
        patches_resolution = self.patch_embed.patches_resolution
        self.patches_resolution = patches_resolution

        # merge non-overlapping patches into image
        self.patch_unembed = PatchUnEmbed(
            img_size=img_size,
            patch_size=patch_size,
            in_chans=embed_dim,
            embed_dim=embed_dim,
            norm_layer=norm_layer if self.patch_norm else None,
        )

        # absolute position embedding
        if self.ape:
            self.absolute_pos_embed = nn.Parameter(  # type: ignore
                torch.zeros(1, num_patches, embed_dim)
            )
            trunc_normal_(self.absolute_pos_embed, std=0.02)

        # relative position index
        relative_position_index_SA = self.calculate_rpi_sa()
        self.register_buffer("relative_position_index_SA", relative_position_index_SA)

        # build Residual Adaptive Token Dictionary Blocks (ATDB)
        self.layers = nn.ModuleList()
        for i_layer in range(self.num_layers):
            layer = ATDB(
                dim=embed_dim,
                idx=i_layer,
                input_resolution=(patches_resolution[0], patches_resolution[1]),
                depth=depths[i_layer],
                num_heads=num_heads[i_layer],
                window_size=window_size,
                category_size=category_size,
                num_tokens=num_tokens,
                reducted_dim=reducted_dim,
                convffn_kernel_size=convffn_kernel_size,
                mlp_ratio=self.mlp_ratio,
                qkv_bias=qkv_bias,
                norm_layer=norm_layer,
                downsample=None,
                img_size=img_size,
                patch_size=patch_size,
                resi_connection=resi_connection,
            )
            self.layers.append(layer)
        self.norm = norm_layer(self.num_features)

        # build the last conv layer in deep feature extraction
        if resi_connection == "1conv":
            self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
        elif resi_connection == "3conv":
            # to save parameters and memory
            self.conv_after_body = nn.Sequential(
                nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1),
                nn.LeakyReLU(negative_slope=0.2, inplace=True),
                nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0),
                nn.LeakyReLU(negative_slope=0.2, inplace=True),
                nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1),
            )

        # ------------------------- 3, high quality image reconstruction ------------------------- #
        if self.upsampler == "pixelshuffle":
            # for classical SR
            self.conv_before_upsample = nn.Sequential(
                nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True)
            )
            self.upsample = Upsample(upscale, num_feat)
            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
        elif self.upsampler == "pixelshuffledirect":
            # for lightweight SR (to save parameters)
            self.upsample = UpsampleOneStep(
                upscale,
                embed_dim,
                num_out_ch,
                (patches_resolution[0], patches_resolution[1]),
            )
        elif self.upsampler == "nearest+conv":
            # for real-world SR (less artifacts)
            assert self.upscale == 4, "only support x4 now."
            self.conv_before_upsample = nn.Sequential(
                nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True)
            )
            self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
            self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
            self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
            self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
        else:
            # for image denoising and JPEG compression artifact reduction
            self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1)

        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=0.02)
            if isinstance(m, nn.Linear) and m.bias is not None:  # type: ignore
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)

    @torch.jit.ignore  # type: ignore
    def no_weight_decay(self):
        return {"absolute_pos_embed"}

    @torch.jit.ignore  # type: ignore
    def no_weight_decay_keywords(self):
        return {"relative_position_bias_table"}

    def forward_features(self, x, params):
        x_size = (x.shape[2], x.shape[3])
        x = self.patch_embed(x)
        if self.ape:
            x = x + self.absolute_pos_embed

        for layer in self.layers:
            x = layer(x, x_size, params)

        x = self.norm(x)  # b seq_len c
        x = self.patch_unembed(x, x_size)

        return x

    def calculate_rpi_sa(self):
        # calculate relative position index for SW-MSA
        coords_h = torch.arange(self.window_size)
        coords_w = torch.arange(self.window_size)
        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
        relative_coords = (
            coords_flatten[:, :, None] - coords_flatten[:, None, :]
        )  # 2, Wh*Ww, Wh*Ww
        relative_coords = relative_coords.permute(
            1, 2, 0
        ).contiguous()  # Wh*Ww, Wh*Ww, 2
        relative_coords[:, :, 0] += self.window_size - 1  # shift to start from 0
        relative_coords[:, :, 1] += self.window_size - 1
        relative_coords[:, :, 0] *= 2 * self.window_size - 1
        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
        return relative_position_index

    def calculate_mask(self, x_size):
        # calculate attention mask for SW-MSA
        h, w = x_size
        img_mask = torch.zeros((1, h, w, 1))  # 1 h w 1
        h_slices = (
            slice(0, -self.window_size),
            slice(-self.window_size, -(self.window_size // 2)),
            slice(-(self.window_size // 2), None),
        )
        w_slices = (
            slice(0, -self.window_size),
            slice(-self.window_size, -(self.window_size // 2)),
            slice(-(self.window_size // 2), None),
        )
        cnt = 0
        for h in h_slices:
            for w in w_slices:
                img_mask[:, h, w, :] = cnt
                cnt += 1

        mask_windows = window_partition(
            img_mask, self.window_size
        )  # nw, window_size, window_size, 1
        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
        attn_mask = attn_mask.masked_fill(attn_mask != 0, -100.0).masked_fill(
            attn_mask == 0, 0.0
        )

        return attn_mask

    @property
    def is_norm(self):
        return self.no_norm is None

    def forward(self, x):
        # padding
        h_ori, w_ori = x.size()[-2], x.size()[-1]
        mod = self.window_size
        h_pad = ((h_ori + mod - 1) // mod) * mod - h_ori
        w_pad = ((w_ori + mod - 1) // mod) * mod - w_ori
        h, w = h_ori + h_pad, w_ori + w_pad
        x = torch.cat([x, torch.flip(x, [2])], 2)[:, :, :h, :]
        x = torch.cat([x, torch.flip(x, [3])], 3)[:, :, :, :w]

        # rgb norm
        if self.is_norm:
            self.mean = self.mean.type_as(x)
            x = (x - self.mean) * self.img_range

        attn_mask = self.calculate_mask([h, w]).to(x.device)
        params = {"attn_mask": attn_mask, "rpi_sa": self.relative_position_index_SA}

        if self.upsampler == "pixelshuffle":
            # for classical SR
            x = self.conv_first(x)
            x = self.conv_after_body(self.forward_features(x, params)) + x
            x = self.conv_before_upsample(x)
            x = self.conv_last(self.upsample(x))
        elif self.upsampler == "pixelshuffledirect":
            # for lightweight SR
            x = self.conv_first(x)
            x = self.conv_after_body(self.forward_features(x, params)) + x
            x = self.upsample(x)
        elif self.upsampler == "nearest+conv":
            # for real-world SR
            x = self.conv_first(x)
            x = self.conv_after_body(self.forward_features(x, params)) + x
            x = self.conv_before_upsample(x)
            x = self.lrelu(
                self.conv_up1(F.interpolate(x, scale_factor=2, mode="nearest"))
            )
            x = self.lrelu(
                self.conv_up2(F.interpolate(x, scale_factor=2, mode="nearest"))
            )
            x = self.conv_last(self.lrelu(self.conv_hr(x)))
        else:
            # for image denoising and JPEG compression artifact reduction
            x_first = self.conv_first(x)
            res = self.conv_after_body(self.forward_features(x_first, params)) + x_first
            x = x + self.conv_last(res)

        if self.is_norm:
            x = x / self.img_range + self.mean

        # unpadding
        x = x[..., : h_ori * self.upscale, : w_ori * self.upscale]

        return x