MRI_recon/code/Frequency-Diffusion/networks/compare_models/ARTfuse_layer.py

"""
ART layer in "ICLR 2023: ACCURATE IMAGE RESTORATION WITH ATTENTION RETRACTABLE TRANSFORMER".
""" 

import torch
import torch.nn as nn
import torch.nn.functional as F
from timm.models.layers import DropPath, to_2tuple, trunc_normal_
from pdb import set_trace as stx
import numbers

from einops import rearrange
import math

NEG_INF = -1000000


##########################################################################
class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        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.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x


class DynamicPosBias(nn.Module):
    def __init__(self, dim, num_heads):
        super().__init__()
        self.num_heads = num_heads
        self.pos_dim = dim // 4
        self.pos_proj = nn.Linear(2, self.pos_dim)
        self.pos1 = nn.Sequential(
            nn.LayerNorm(self.pos_dim),
            nn.ReLU(inplace=True),
            nn.Linear(self.pos_dim, self.pos_dim),
        )
        self.pos2 = nn.Sequential(
            nn.LayerNorm(self.pos_dim),
            nn.ReLU(inplace=True),
            nn.Linear(self.pos_dim, self.pos_dim)
        )
        self.pos3 = nn.Sequential(
            nn.LayerNorm(self.pos_dim),
            nn.ReLU(inplace=True),
            nn.Linear(self.pos_dim, self.num_heads)
        )

    def forward(self, biases):
        pos = self.pos3(self.pos2(self.pos1(self.pos_proj(biases))))
        return pos

    def flops(self, N):
        flops = N * 2 * self.pos_dim
        flops += N * self.pos_dim * self.pos_dim
        flops += N * self.pos_dim * self.pos_dim
        flops += N * self.pos_dim * self.num_heads
        return flops


#########################################
class Attention(nn.Module):
    r""" Multi-head self attention module with dynamic position bias.

    Args:
        dim (int): Number of input channels.
        num_heads (int): Number of attention heads.
        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
    """

    def __init__(self, dim, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.,
                 position_bias=True):

        super().__init__()
        self.dim = dim
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5
        self.position_bias = position_bias
        if self.position_bias:
            self.pos = DynamicPosBias(self.dim // 4, self.num_heads)

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

        self.softmax = nn.Softmax(dim=-1)

    def forward(self, x, H, W, mask=None):
        """
        Args:
            x: input features with shape of (num_groups*B, N, C)
            mask: (0/-inf) mask with shape of (num_groups, Gh*Gw, Gh*Gw) or None
            H: height of each group
            W: width of each group
        """
        # print("input to the Attention layer:", x.max(), x.min())
        group_size = (H, W)
        B_, N, C = x.shape
        assert H * W == N
        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4).contiguous()
        q, k, v = qkv[0], qkv[1], qkv[2]

        # print("q range:", q.max(), q.min())
        # print("k range:", k.max(), k.min())

        q = q * self.scale
        attn = q @ k.transpose(-2, -1).contiguous()  # (B_, self.num_heads, N, N), N = H*W
        # print("attention matrix:", attn.shape, attn)

        if self.position_bias:
            # generate mother-set
            position_bias_h = torch.arange(1 - group_size[0], group_size[0], device=attn.device)
            position_bias_w = torch.arange(1 - group_size[1], group_size[1], device=attn.device)
            biases = torch.stack(torch.meshgrid([position_bias_h, position_bias_w]))  # 2, 2Gh-1, 2W2-1
            biases = biases.flatten(1).transpose(0, 1).contiguous().float()  # (2h-1)*(2w-1) 2

            # get pair-wise relative position index for each token inside the window
            coords_h = torch.arange(group_size[0], device=attn.device)
            coords_w = torch.arange(group_size[1], device=attn.device)
            coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Gh, Gw
            coords_flatten = torch.flatten(coords, 1)  # 2, Gh*Gw
            relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Gh*Gw, Gh*Gw
            relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Gh*Gw, Gh*Gw, 2
            relative_coords[:, :, 0] += group_size[0] - 1  # shift to start from 0
            relative_coords[:, :, 1] += group_size[1] - 1
            relative_coords[:, :, 0] *= 2 * group_size[1] - 1
            relative_position_index = relative_coords.sum(-1)  # Gh*Gw, Gh*Gw

            pos = self.pos(biases)  # 2Gh-1 * 2Gw-1, heads
            # select position bias
            relative_position_bias = pos[relative_position_index.view(-1)].view(
                group_size[0] * group_size[1], group_size[0] * group_size[1], -1)  # Gh*Gw,Gh*Gw,nH
            relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Gh*Gw, Gh*Gw
            attn = attn + relative_position_bias.unsqueeze(0)


        if mask is not None:
            nP = mask.shape[0]
            attn = attn.view(B_ // nP, nP, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)  # (B, nP, nHead, N, N)
            attn = attn.view(-1, self.num_heads, N, N)
            attn = self.softmax(attn)
        else:
            attn = self.softmax(attn)

        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
        # print("feature before proj-layer:", x.max(), x.min())
        x = self.proj(x)
        # print("feature after proj-layer:", x.max(), x.min())
        x = self.proj_drop(x)
        return x


##########################################################################
class TransformerBlock(nn.Module):
    r""" ART Transformer Block.

    Args:
        dim (int): Number of input channels.
        num_heads (int): Number of attention heads.
        window_size: window size of dense attention
        interval: interval size of sparse attention
        ds_flag (int): use Dense Attention or Sparse Attention, 0 for DAB and 1 for SAB.
        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
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
        drop (float, optional): Dropout rate. Default: 0.0
        attn_drop (float, optional): Attention dropout rate. Default: 0.0
        drop_path (float, optional): Stochastic depth rate. Default: 0.0
        # act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
    """

    def __init__(self,
                 dim,
                 num_heads,
                 window_size=7,
                 interval=8,
                 ds_flag=0,
                 mlp_ratio=4.,
                 qkv_bias=True,
                 qk_scale=None,
                 drop=0.,
                 attn_drop=0.,
                 drop_path=0.,
                 act_layer=nn.GELU,
                 norm_layer=nn.LayerNorm, 
                 pre_norm=True):
        super().__init__()
        self.dim = dim
        self.num_heads = num_heads
        self.window_size = window_size
        self.interval = interval
        self.ds_flag = ds_flag
        self.mlp_ratio = mlp_ratio

        # self.conv = nn.Conv2d(dim, dim, kernel_size=1)

        self.norm1 = norm_layer(dim)

        self.attn = Attention(
            dim, num_heads=num_heads,
            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop,
            position_bias=True)
        
        # self.conv = ConvBlock(self.dim, self.dim, drop)

        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
        # self.bn_norm = nn.BatchNorm2d(dim)
        self.pre_norm = pre_norm

    def forward(self, x, x_size):
        H, W = x_size
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size %d, %d, %d" % (L, H, W)

        if min(H, W) <= self.window_size:
            # if window size is larger than input resolution, we don't partition windows
            self.ds_flag = 0
            self.window_size = min(H, W)

        # x = self.conv(x)
        shortcut = x
        if self.pre_norm:
            # print("using pre_norm")
            x = self.norm1(x)  ## 归一化之后特征范围变得非常小
        x = x.view(B, H, W, C)
        # print("normalized input range:", x.max(), x.min(), x.mean())

        # padding
        size_par = self.interval if self.ds_flag == 1 else self.window_size
        pad_l = pad_t = 0
        pad_r = (size_par - W % size_par) % size_par
        pad_b = (size_par - H % size_par) % size_par
        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
        _, Hd, Wd, _ = x.shape

        mask = torch.zeros((1, Hd, Wd, 1), device=x.device)
        if pad_b > 0:
            mask[:, -pad_b:, :, :] = -1
        if pad_r > 0:
            mask[:, :, -pad_r:, :] = -1

        # print("pad_b and pad_r:", pad_b, pad_r)

        # partition the whole feature map into several groups
        if self.ds_flag == 0:  # Dense Attention
            G = Gh = Gw = self.window_size
            x = x.reshape(B, Hd // G, G, Wd // G, G, C).permute(0, 1, 3, 2, 4, 5).contiguous()
            x = x.reshape(B * Hd * Wd // G ** 2, G ** 2, C)
            nP = Hd * Wd // G ** 2  # number of partitioning groups
            # attn_mask
            if pad_r > 0 or pad_b > 0:
                mask = mask.reshape(1, Hd // G, G, Wd // G, G, 1).permute(0, 1, 3, 2, 4, 5).contiguous()
                mask = mask.reshape(nP, 1, G * G)
                attn_mask = torch.zeros((nP, G * G, G * G), device=x.device)
                attn_mask = attn_mask.masked_fill(mask < 0, NEG_INF)
            else:
                attn_mask = None
        if self.ds_flag == 1:  # Sparse Attention
            I, Gh, Gw = self.interval, Hd // self.interval, Wd // self.interval
            x = x.reshape(B, Gh, I, Gw, I, C).permute(0, 2, 4, 1, 3, 5).contiguous()
            x = x.reshape(B * I * I, Gh * Gw, C)
            nP = I ** 2  # number of partitioning groups
            # attn_mask
            if pad_r > 0 or pad_b > 0:
                mask = mask.reshape(1, Gh, I, Gw, I, 1).permute(0, 2, 4, 1, 3, 5).contiguous()
                mask = mask.reshape(nP, 1, Gh * Gw)
                attn_mask = torch.zeros((nP, Gh * Gw, Gh * Gw), device=x.device)
                attn_mask = attn_mask.masked_fill(mask < 0, NEG_INF)
            else:
                attn_mask = None

        # MSA
        # print("attn mask:", attn_mask)
        x = self.attn(x, Gh, Gw, mask=attn_mask)  # nP*B, Gh*Gw, C

        # merge embeddings
        if self.ds_flag == 0:
            x = x.reshape(B, Hd // G, Wd // G, G, G, C).permute(0, 1, 3, 2, 4,
                                                                5).contiguous()  # B, Hd//G, G, Wd//G, G, C
        else:
            x = x.reshape(B, I, I, Gh, Gw, C).permute(0, 3, 1, 4, 2, 5).contiguous()  # B, Gh, I, Gw, I, C
        x = x.reshape(B, Hd, Wd, C)
        # print("range of feature before layer:", x.max(), x.min(), x.mean())
        # x = self.conv(x.permute(0, 3, 1, 2)).permute(0, 2, 3, 1)

        # x_bn = self.bn_norm(x.permute(0, 3, 1, 2))

        # remove padding
        if pad_r > 0 or pad_b > 0:
            x = x[:, :H, :W, :].contiguous()
        x = x.view(B, H * W, C)

        # FFN
        # print("[ART layer] input and output feature difference:", torch.mean(torch.abs(shortcut - x)))
        # print("range of feature before ART layer:", shortcut.max(), shortcut.min(), shortcut.mean())
        # print("range of feature after ART layer:", x.max(), x.min(), x.mean())
        if self.pre_norm:
            x = shortcut + self.drop_path(x)
            x = x + self.drop_path(self.mlp(self.norm2(x)))
        else:
            # print("using post_norm")
            x = self.norm1(shortcut + self.drop_path(x))
            x = self.norm2(x + self.drop_path(self.mlp(x)))            
            # x = shortcut + self.drop_path(x)
            # x = x + self.drop_path(self.mlp(x))

        # print("range of fused feature:", x.max(), x.min())
        # print("x:", x.shape)

        return x


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



class ConvBlock(nn.Module):
    """
    A Convolutional Block that consists of two convolution layers each followed by
    instance normalization, LeakyReLU activation and dropout.
    """

    def __init__(self, in_chans: int, out_chans: int, drop_prob: float):
        """
        Args:
            in_chans: Number of channels in the input.
            out_chans: Number of channels in the output.
            drop_prob: Dropout probability.
        """
        super().__init__()

        self.in_chans = in_chans
        self.out_chans = out_chans
        self.drop_prob = drop_prob

        self.layers = nn.Sequential(
            nn.Conv2d(in_chans, out_chans, kernel_size=3, padding=1, bias=False),
            nn.InstanceNorm2d(out_chans),
            nn.LeakyReLU(negative_slope=0.2, inplace=True),
            nn.Dropout2d(drop_prob),
            nn.Conv2d(out_chans, out_chans, kernel_size=3, padding=1, bias=False),
            nn.InstanceNorm2d(out_chans),
            nn.LeakyReLU(negative_slope=0.2, inplace=True),
            nn.Dropout2d(drop_prob),
        )

    def forward(self, image: torch.Tensor) -> torch.Tensor:
        return self.layers(image)



##########################################################################
class Cross_TransformerBlock(nn.Module):
    r""" ART Transformer Block.

    Args:
        dim (int): Number of input channels.
        num_heads (int): Number of attention heads.
        window_size: window size of dense attention
        interval: interval size of sparse attention
        ds_flag (int): use Dense Attention or Sparse Attention, 0 for DAB and 1 for SAB.
        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
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
        drop (float, optional): Dropout rate. Default: 0.0
        attn_drop (float, optional): Attention dropout rate. Default: 0.0
        drop_path (float, optional): Stochastic depth rate. Default: 0.0
        # act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
    """

    def __init__(self,
                 dim,
                 num_heads,
                 window_size=7,
                 interval=8,
                 ds_flag=0,
                 mlp_ratio=4.,
                 qkv_bias=True,
                 qk_scale=None,
                 drop=0.,
                 attn_drop=0.,
                 drop_path=0.,
                 act_layer=nn.GELU,
                 norm_layer=nn.LayerNorm):
        super().__init__()
        self.dim = dim
        self.num_heads = num_heads
        self.window_size = window_size
        self.interval = interval
        self.ds_flag = ds_flag
        self.mlp_ratio = mlp_ratio

        self.norm1 = norm_layer(dim)

        self.attn = Attention(
            dim, num_heads=num_heads,
            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop,
            position_bias=True)

        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp_x = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
        self.mlp_y = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

    def forward(self, x, y, x_size):
        """
        x, y: feature maps of two modalities. They are from the same level with same feature size.
        """
        H, W = x_size
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size %d, %d, %d" % (L, H, W)

        if min(H, W) <= self.window_size:
            # if window size is larger than input resolution, we don't partition windows
            self.ds_flag = 0
            self.window_size = min(H, W)

        shortcut_x = x
        x = self.norm1(x)
        x = x.view(B, H, W, C)

        shortcut_y = y
        y = self.norm1(y)
        y = y.view(B, H, W, C)

        # padding
        size_par = self.interval if self.ds_flag == 1 else self.window_size
        pad_l = pad_t = 0
        pad_r = (size_par - W % size_par) % size_par
        pad_b = (size_par - H % size_par) % size_par
        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
        y = F.pad(y, (0, 0, pad_l, pad_r, pad_t, pad_b))
        _, Hd, Wd, _ = x.shape

        mask = torch.zeros((1, Hd, Wd, 1), device=x.device)
        if pad_b > 0:
            mask[:, -pad_b:, :, :] = -1
        if pad_r > 0:
            mask[:, :, -pad_r:, :] = -1

        # print("pad_b and pad_r:", pad_b, pad_r)

        # partition the whole feature map into several groups
        if self.ds_flag == 0:  # Dense Attention
            G = Gh = Gw = self.window_size
            x = x.reshape(B, Hd // G, G, Wd // G, G, C).permute(0, 1, 3, 2, 4, 5).contiguous()
            x = x.reshape(B * Hd * Wd // G ** 2, G ** 2, C)
            y = y.reshape(B, Hd // G, G, Wd // G, G, C).permute(0, 1, 3, 2, 4, 5).contiguous()
            y = y.reshape(B * Hd * Wd // G ** 2, G ** 2, C)
            nP = Hd * Wd // G ** 2  # number of partitioning groups
            # attn_mask
            if pad_r > 0 or pad_b > 0:
                mask = mask.reshape(1, Hd // G, G, Wd // G, G, 1).permute(0, 1, 3, 2, 4, 5).contiguous()
                mask = mask.reshape(nP, 1, G * G)
                attn_mask = torch.zeros((nP, G * G, G * G), device=x.device)
                attn_mask = attn_mask.masked_fill(mask < 0, NEG_INF)
            else:
                attn_mask = None
        if self.ds_flag == 1:  # Sparse Attention
            I, Gh, Gw = self.interval, Hd // self.interval, Wd // self.interval
            x = x.reshape(B, Gh, I, Gw, I, C).permute(0, 2, 4, 1, 3, 5).contiguous()
            x = x.reshape(B * I * I, Gh * Gw, C)
            y = y.reshape(B, Gh, I, Gw, I, C).permute(0, 2, 4, 1, 3, 5).contiguous()
            y = y.reshape(B * I * I, Gh * Gw, C)            
            nP = I ** 2  # number of partitioning groups
            # attn_mask
            if pad_r > 0 or pad_b > 0:
                mask = mask.reshape(1, Gh, I, Gw, I, 1).permute(0, 2, 4, 1, 3, 5).contiguous()
                mask = mask.reshape(nP, 1, Gh * Gw)
                attn_mask = torch.zeros((nP, Gh * Gw, Gh * Gw), device=x.device)
                attn_mask = attn_mask.masked_fill(mask < 0, NEG_INF)
            else:
                attn_mask = None

        # Inside each window, fuse the x and y, then compute self-attention.
        xy = torch.cat((x, y), 1)
        xy = self.attn(xy, Gh, 2*Gw, mask=attn_mask)  # nP*B, Gh*2Gw, C
        # print("fused xy:", xy.shape)

        # merge embeddings
        if self.ds_flag == 0:
            x = xy[:, :xy.shape[1]//2, :].reshape(B, Hd // G, Wd // G, G, G, C).permute(0, 1, 3, 2, 4, 5).contiguous()  # B, Hd//G, G, Wd//G, G, C
            y = xy[:, xy.shape[1]//2:, :].reshape(B, Hd // G, Wd // G, G, G, C).permute(0, 1, 3, 2, 4, 5).contiguous()  # B, Hd//G, G, Wd//G, G, C
        else:
            x = xy[:, :xy.shape[1]//2, :].reshape(B, I, I, Gh, Gw, C).permute(0, 3, 1, 4, 2, 5).contiguous()  # B, Gh, I, Gw, I, C
            y = xy[:, xy.shape[1]//2:, :].reshape(B, I, I, Gh, Gw, C).permute(0, 3, 1, 4, 2, 5).contiguous()  # B, Gh, I, Gw, I, C
        
        x = x.reshape(B, Hd, Wd, C)
        y = y.reshape(B, Hd, Wd, C)

        # remove padding
        if pad_r > 0 or pad_b > 0:
            x = x[:, :H, :W, :].contiguous()
            y = y[:, :H, :W, :].contiguous()
        x = x.view(B, H * W, C)
        y = y.view(B, H * W, C)

        # FFN
        x = shortcut_x + self.drop_path(x)
        x = x + self.drop_path(self.mlp_x(self.norm2(x)))
        # print("x:", x.shape)
        y = shortcut_y + self.drop_path(y)
        y = y + self.drop_path(self.mlp_y(self.norm2(y)))

        return x, y


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






##########################################################################
class Cross_TransformerBlock_v2(nn.Module):
    r""" ART Transformer Block.
        将两个模态的特征沿着channel方向concat, 一起取window. 之后把取出来的两个模态中同一个window的所有patches合并，送到transformer处理。
    Args:
        dim (int): Number of input channels.
        num_heads (int): Number of attention heads.
        window_size: window size of dense attention
        interval: interval size of sparse attention
        ds_flag (int): use Dense Attention or Sparse Attention, 0 for DAB and 1 for SAB.
        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
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
        drop (float, optional): Dropout rate. Default: 0.0
        attn_drop (float, optional): Attention dropout rate. Default: 0.0
        drop_path (float, optional): Stochastic depth rate. Default: 0.0
        # act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
    """

    def __init__(self,
                 dim,
                 num_heads,
                 window_size=7,
                 interval=8,
                 ds_flag=0,
                 mlp_ratio=4.,
                 qkv_bias=True,
                 qk_scale=None,
                 drop=0.,
                 attn_drop=0.,
                 drop_path=0.,
                 act_layer=nn.GELU,
                 norm_layer=nn.LayerNorm):
        super().__init__()
        self.dim = dim
        self.num_heads = num_heads
        self.window_size = window_size
        self.interval = interval
        self.ds_flag = ds_flag
        self.mlp_ratio = mlp_ratio

        self.norm1 = norm_layer(dim)

        self.attn = Attention(
            dim, num_heads=num_heads,
            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop,
            position_bias=True)

        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp_x = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
        self.mlp_y = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

    def forward(self, x, y, x_size):
        """
        x, y: feature maps of two modalities. They are from the same level with same feature size.
        """
        H, W = x_size
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size %d, %d, %d" % (L, H, W)

        if min(H, W) <= self.window_size:
            # if window size is larger than input resolution, we don't partition windows
            self.ds_flag = 0
            self.window_size = min(H, W)

        shortcut_x = x
        x = self.norm1(x)
        x = x.view(B, H, W, C)

        shortcut_y = y
        y = self.norm1(y)
        y = y.view(B, H, W, C)

        # padding
        xy = torch.cat((x, y), -1)
        size_par = self.interval if self.ds_flag == 1 else self.window_size
        pad_l = pad_t = 0
        pad_r = (size_par - W % size_par) % size_par
        pad_b = (size_par - H % size_par) % size_par
        xy = F.pad(xy, (0, 0, pad_l, pad_r, pad_t, pad_b))
        _, Hd, Wd, _ = x.shape

        mask = torch.zeros((1, Hd, Wd, 1), device=x.device)
        if pad_b > 0:
            mask[:, -pad_b:, :, :] = -1
        if pad_r > 0:
            mask[:, :, -pad_r:, :] = -1

        # print("pad_b and pad_r:", pad_b, pad_r)

        # partition the whole feature map into several groups
        if self.ds_flag == 0:  # Dense Attention
            G = Gh = Gw = self.window_size
            xy = xy.reshape(B, Hd // G, G, Wd // G, G, 2*C).permute(0, 1, 3, 2, 4, 5).contiguous()
            xy = xy.reshape(B * Hd * Wd // G ** 2, G ** 2, 2*C)
            nP = Hd * Wd // G ** 2  # number of partitioning groups
            # attn_mask
            if pad_r > 0 or pad_b > 0:
                mask = mask.reshape(1, Hd // G, G, Wd // G, G, 1).permute(0, 1, 3, 2, 4, 5).contiguous()
                mask = mask.reshape(nP, 1, G * G)
                attn_mask = torch.zeros((nP, G * G, G * G), device=x.device)
                attn_mask = attn_mask.masked_fill(mask < 0, NEG_INF)
            else:
                attn_mask = None
        if self.ds_flag == 1:  # Sparse Attention
            I, Gh, Gw = self.interval, Hd // self.interval, Wd // self.interval
            xy = xy.reshape(B, Gh, I, Gw, I, 2*C).permute(0, 2, 4, 1, 3, 5).contiguous()
            xy = xy.reshape(B * I * I, Gh * Gw, 2*C)         
            nP = I ** 2  # number of partitioning groups
            # attn_mask
            if pad_r > 0 or pad_b > 0:
                mask = mask.reshape(1, Gh, I, Gw, I, 1).permute(0, 2, 4, 1, 3, 5).contiguous()
                mask = mask.reshape(nP, 1, Gh * Gw)
                attn_mask = torch.zeros((nP, Gh * Gw, Gh * Gw), device=x.device)
                attn_mask = attn_mask.masked_fill(mask < 0, NEG_INF)
            else:
                attn_mask = None

        # Inside each window, fuse the x and y, then compute self-attention.
        x = xy[:, :, :C]
        y = xy[:, :, C:]
        xy = torch.cat((x, y), 1)
        xy = self.attn(xy, Gh, 2*Gw, mask=attn_mask)  # nP*B, Gh*2Gw, C
        # print("fused xy:", xy.shape)

        # merge embeddings
        if self.ds_flag == 0:
            x = xy[:, :xy.shape[1]//2, :].reshape(B, Hd // G, Wd // G, G, G, C).permute(0, 1, 3, 2, 4, 5).contiguous()  # B, Hd//G, G, Wd//G, G, C
            y = xy[:, xy.shape[1]//2:, :].reshape(B, Hd // G, Wd // G, G, G, C).permute(0, 1, 3, 2, 4, 5).contiguous()  # B, Hd//G, G, Wd//G, G, C
        else:
            x = xy[:, :xy.shape[1]//2, :].reshape(B, I, I, Gh, Gw, C).permute(0, 3, 1, 4, 2, 5).contiguous()  # B, Gh, I, Gw, I, C
            y = xy[:, xy.shape[1]//2:, :].reshape(B, I, I, Gh, Gw, C).permute(0, 3, 1, 4, 2, 5).contiguous()  # B, Gh, I, Gw, I, C
        
        x = x.reshape(B, Hd, Wd, C)
        y = y.reshape(B, Hd, Wd, C)

        # remove padding
        if pad_r > 0 or pad_b > 0:
            x = x[:, :H, :W, :].contiguous()
            y = y[:, :H, :W, :].contiguous()
        x = x.view(B, H * W, C)
        y = y.view(B, H * W, C)

        # FFN
        x = shortcut_x + self.drop_path(x)
        x = x + self.drop_path(self.mlp_x(self.norm2(x)))
        # print("x:", x.shape)
        y = shortcut_y + self.drop_path(y)
        y = y + self.drop_path(self.mlp_y(self.norm2(y)))

        return x, y


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