model.py

import ot
import torch
import torch.nn as nn
import torch.utils.checkpoint as checkpoint
from fightingcv_attention.conv.CondConv import CondConv
from timm.models.layers import DropPath, to_2tuple, trunc_normal_
from einops import rearrange
import torchvision.models as models


class Mlp(nn.Module):
    """
    MLP of the transformer layer
    input : x(B,N,C)
    Args:
        in_features: the dim of the input features
        hidden_features: the hidden dim of the MLP
        out_features: the dim of the output features
        act_layers: the act function of the MLP, default GELU
        drop: dropout ratio
    returns: (B,N,C)
    """

    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 ConvFFN(nn.Module):
    """CNN for the transformer FFN module

    input: x(B,C,H,W)
    Args:
        in_features: the dim of the input features
        hidden_features: the hidden dim of the MLP
        out_features: the dim of the output features
        act_layers: the act function of the MLP, default GELU
        drop: dropout ratio
    returns: (B,C,H,W)
    """

    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.conv = nn.Conv2d(in_features, in_features, 3, 1, 3, groups=in_features)
        self.norm = nn.LayerNorm(in_features)
        self.fc1 = nn.Conv2d(in_features, hidden_features, 1, 1, 0)
        self.act = act_layer()
        self.fc2 = nn.Conv2d(hidden_features, in_features, 1, 1, 0)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.conv(x)
        x = x.permute(0, 2, 3, 1).contiguous()  # B H W C
        x = self.norm(x)
        x = x.permute(0, 3, 1, 2).contiguous()
        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(x, window_size):
    """
    window partion function
    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""" Window based multi-head self attention (W-MSA) module with relative position bias.
    It supports both of shifted and non-shifted window.

    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
        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, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.,
                 drop_key=False):

        super().__init__()
        self.dim = dim
        self.window_size = window_size  # Wh, Ww
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5
        self.drop_key = drop_key

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

        # 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(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[0] - 1  # shift to start from 0
        relative_coords[:, :, 1] += self.window_size[1] - 1
        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
        self.register_buffer("relative_position_index", relative_position_index)

        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)

        trunc_normal_(self.relative_position_bias_table, std=.02)
        self.softmax = nn.Softmax(dim=-1)

    def forward(self, x, mask=None):
        """
        Args:
            x: input features with shape of (num_windows*B, N, C)
            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
        """
        B_, N, C = x.shape
        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2,
                                                                                         0, 3, 1,
                                                                                         4).contiguous()  # (3,B*num_windows,heads,N,head_dim)
        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))  # q:(B*windows, heads, N, C), att:(B*windows, heads, N, N)

        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.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)  # positional enbedding

        # Drop key
        if self.drop_key:
            m_r = torch.ones_like(attn) * 0.1
            attn = attn + torch.bernoulli(m_r) * -1e12

        if mask is not None:
            nW = mask.shape[0]  # mask(nW,N,N)
            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(
                0)  # att(bs,nW,heads,N,N); mask(1,nW,1,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)  # x:(B*windows, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class Channel_Adapter(nn.Module):
    """The Channel Adapter module
    args:
        num_channels: the channel number of the feature map
        resolution: the resolution of the feature map
    """

    def __init__(self, num_channels, resolution):
        super().__init__()
        self.proj = nn.Sequential(
            nn.Conv2d(num_channels, 4 * num_channels, 1),
            nn.GELU(),
            nn.Conv2d(4 * num_channels, num_channels, 1)
        )
        self.pool = nn.AvgPool2d(resolution)
        self.mlp = nn.Sequential(
            nn.Linear(num_channels, 4 * num_channels),
            nn.Linear(4 * num_channels, num_channels),
            nn.Sigmoid()
        )
        self.a = num_channels
        self.b = resolution

    def forward(self, x):
        B, C, H, W = x.shape
        x_proj = self.proj(x)
        avg_x = self.pool(x_proj).reshape(B, C)  # (B,C)
        attn = self.mlp(avg_x)
        attn = attn.reshape(B, C, 1, 1) * x_proj
        x = x + attn
        return x


class PEG_Conv(nn.Module):
    """Conditional position encode generator, using depth-wise convolution

    Args:
        input: (B,C,H,W), based on patch
        dim: the dim of the heads' token
    """

    def __init__(self, dim):
        super().__init__()
        self.conv = nn.Conv2d(dim, dim, kernel_size=3, stride=1, padding=1, groups=dim)

    def forward(self, x):
        pe = self.conv(x)
        return pe


class ConditinoalAttention(nn.Module):
    r""" Global attention with conditional positional encoding.

    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
        drop_key (bool, optional): Using dropkey or not. Default: False
    """

    def __init__(self, dim, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0., drop_key=False):
        super().__init__()
        self.dim = dim
        self.head_dim = dim // num_heads
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5
        self.drop_key = drop_key

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)  # 生成 QKV
        self.attn_drop = nn.Dropout(attn_drop)  # 注意力的 dropout
        self.proj = nn.Linear(dim, dim)  # 输出映射矩阵
        self.proj_drop = nn.Dropout(proj_drop)  # 输出的 dropout

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

    def forward(self, x, mask=None):
        """
        Args:
            x: input features with shape of (B, N, C)
        """
        B_, N, C = x.shape
        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2,
                                                                                         0, 3, 1,
                                                                                         4).contiguous()  # (3,B*num_windows,heads,N,head_dim)
        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))  # q:(B, heads, N, C), att:(B, heads, N, N)

        # Drop key
        if self.drop_key:
            m_r = torch.ones_like(attn) * 0.1
            attn = attn + torch.bernoulli(m_r) * -1e12

        attn = self.softmax(attn)

        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)  # x:(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


def PatchReverse(x, patch_size):
    """Transform the tokens back to image

    input:
        x: (B, N, C), N = (ph pw): Token representation.
    returns:
        x: (B,C,H,W)

    """
    _, N, _ = x.shape
    p1 = p2 = int(N ** 0.5)  # 有多少个 patch
    x = rearrange(x, 'b (p1 p2) (ph pw c) -> b c (p1 ph) (p2 pw)', p1=p1, p2=p2, ph=patch_size,
                  pw=patch_size).contiguous()
    return x


def PatchDivide(x, patch_size):
    """Transform the image to tokens.

    input:
        x: (B,C,H,W)
    returns:
        x: (B, N, C)
    """
    x = rearrange(x, 'b c (p1 h) (p2 w) -> b (p1 p2) (h w c)', h=patch_size, w=patch_size).contiguous()
    return x


class DownSampler(nn.Module):
    """Down sample the feature map.
    """

    def __init__(self, in_channel, out_channel):
        super().__init__()
        self.conv = nn.Conv2d(in_channels=in_channel, out_channels=out_channel, kernel_size=4, stride=2, padding=1)

    def forward(self, x):
        x = self.conv(x)
        return x


class UpSampler(nn.Module):
    """Up sample the feature map."""

    def __init__(self, in_channel, out_channel):
        super().__init__()
        self.conv = nn.ConvTranspose2d(in_channels=in_channel, out_channels=out_channel, kernel_size=2, stride=2)

    def forward(self, x):
        x = self.conv(x)
        return x


class SwinTransformerBlock(nn.Module):
    r""" Swin Transformer Block.

    Args:
        dim (int): Number of input channels.
        input_resolution (tuple[int]): Input resulotion, based on patch.
        num_heads (int): Number of attention heads.
        window_size (int): Window size.
        shift_size (int): Shift size for SW-MSA.
        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.LeakyReLU
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
        ffn_type (string): FFN module,using Convolution or MLP, Default: Mlp
        fused_window_process (bool, optional): If True, use one kernel to fused window shift & window partition for acceleration, similar for the reversed part. Default: False
        drop_key (bool, optional): Using dropkey or not. Default: False

        输入的 x 为 (B,N,C) 表示，在过程中进行窗口划分
    """

    def __init__(self, dim, input_resolution, num_heads, window_size=8, shift_size=0,  # 你妈的，这里写错了，弄得窗口变得没有滑动了
                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
                 act_layer=nn.LeakyReLU, norm_layer=nn.LayerNorm, ffn_type='Mlp',  # 消融：修改为 LeakyReLU
                 fused_window_process=False, drop_key=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.ffn_type = ffn_type
        if min(self.input_resolution) <= self.window_size:
            # if window size is larger than input resolution, we don't partition windows
            self.shift_size = 0
            self.window_size = min(self.input_resolution)
        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"

        self.norm1 = norm_layer(dim)
        self.attn = WindowAttention(
            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, drop_key=drop_key)

        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)
        if self.ffn_type == 'Mlp':
            self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
        else:
            self.mlp = ConvFFN(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

        if self.shift_size > 0:
            # calculate attention mask for SW-MSA
            H, W = self.input_resolution
            img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
            h_slices = (slice(0, -self.window_size),
                        slice(-self.window_size, -self.shift_size),
                        slice(-self.shift_size, None))
            w_slices = (slice(0, -self.window_size),
                        slice(-self.window_size, -self.shift_size),
                        slice(-self.shift_size, 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, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
        else:
            attn_mask = None

        self.register_buffer("attn_mask", attn_mask, persistent=False)  # 注册为 buffer
        self.fused_window_process = fused_window_process  # 是否执行 merge 操作

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

        shortcut = x
        x = self.norm1(x)
        x = x.view(B, H, W, C)  # turn to 2-D reprensentation

        # cyclic shift
        if self.shift_size > 0:
            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
            # partition windows
            x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
        else:
            shifted_x = x
            # partition windows
            x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C

        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C

        # W-MSA/SW-MSA
        attn_windows = self.attn(x_windows, mask=self.attn_mask)  # nW*B, window_size*window_size, C

        # merge windows
        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)  # (B*windows,Wh,Ww,C)

        # reverse cyclic shift 恢复循环移位操作
        if self.shift_size > 0:
            shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
        else:
            shifted_x = window_reverse(attn_windows, self.window_size, H, W)
            x = shifted_x
        x = x.view(B, H * W, C)

        x = shortcut + self.drop_path(x)

        # FFN
        if self.ffn_type == 'Mlp':
            x = x + self.drop_path(self.mlp(self.norm2(x)))
        else:
            shortcut2 = x
            x = rearrange(x, 'bs (h w) c -> bs c h w', bs=B, h=H, w=W).contiguous()  # B C H W
            x = self.mlp(x)
            x = rearrange(x, 'bs c h w -> bs (h w) c').contiguous()  # B N C
            x = self.drop_path(x)
            x = shortcut2 + x
        return x


class Global_Enhanced_BottleNeck_Block(nn.Module):
    r""" Global Enhanced BottleNeck Block, using global attention to model global information.

    Args:
        dim (int): Number of input channels.
        input_resolution (tuple[int]): Input resulotion. 这个 resolution 是 patch 的分辨率
        num_heads (int): Number of attention heads.
        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
        ffn_type (string): FFN module,using Convolution or MLP, Default: Mlp
        drop_key (bool, optional): Using dropkey or not. Default: False

        输入的 x 为 (B,N,C) 表示，不再划分窗口
    """

    def __init__(self, dim, input_resolution, num_heads,
                 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, ffn_type='Mlp', drop_key=False):
        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.num_heads = num_heads
        self.mlp_ratio = mlp_ratio
        self.ffn_type = ffn_type

        self.norm1 = norm_layer(dim)
        self.attn = ConditinoalAttention(
            dim, num_heads=num_heads,
            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, drop_key=drop_key)

        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)
        if self.ffn_type == 'Mlp':
            self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
        else:
            self.mlp = ConvFFN(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

    def forward(self, x):
        H, W = self.input_resolution
        B, L, C = x.shape

        shortcut = x
        x = self.norm1(x)

        # W-MSA/SW-MSA
        x = self.attn(x)  # B, N, C

        x = shortcut + self.drop_path(x)

        # FFN
        if self.ffn_type == 'Mlp':
            x = x + self.drop_path(self.mlp(self.norm2(x)))
        else:
            # 使用 ConvNext block 替换
            shortcut2 = x
            x = rearrange(x, 'bs (h w) c -> bs c h w', bs=B, h=H, w=W).contiguous()  # B C H W
            x = self.mlp(x)
            x = rearrange(x, 'bs c h w -> bs (h w) c').contiguous()  # B N C
            x = self.drop_path(x)
            x = shortcut2 + x
        return x


class GEB(nn.Module):
    r""" Global Enhanced BottleNeck.

    Args:
        dim (int): Number of input channels.
        input_resolution (tuple[int]): Input resulotion. 这个 resolution 是 patch 的分辨率
        num_heads (int): Number of attention heads.
        patch_size (int): Patch size o
        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
        ffn_type (string): FFN module,using Convolution or MLP, Default: Mlp
        drop_key (bool, optional): Using dropkey or not. Default: False

        输入的 x 为 (B,N,C) 表示，不再划分窗口
    """

    def __init__(self, dim, input_resolution, num_heads, patch_size, depth,
                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
                 norm_layer=nn.LayerNorm, use_checkpoint=False, drop_key=False, mode=None):
        super().__init__()
        self.mode = mode
        self.use_checkpoint = use_checkpoint
        self.patch_size = patch_size
        self.blocks = nn.ModuleList([
            Global_Enhanced_BottleNeck_Block(dim=dim, input_resolution=to_2tuple(input_resolution),
                                             num_heads=num_heads,
                                             mlp_ratio=mlp_ratio,
                                             qkv_bias=qkv_bias, qk_scale=qk_scale,
                                             drop=drop, attn_drop=attn_drop,
                                             drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
                                             norm_layer=norm_layer, drop_key=drop_key)
            for i in range(depth)])
        self.pe_generator = PEG_Conv(dim)

    def forward(self, pool3, latent):
        map = OT_MAPPING()
        x = map(pool3, latent, self.mode)  # x (8, 128, 16, 16)
        for param in map.parameters():
            param.requires_grad = False
        x = PatchDivide(x, self.patch_size)
        B_, N, C = x.shape
        x = rearrange(x, 'b (h w) c -> b c h w ', h=int(N ** 0.5), w=int(N ** 0.5)).contiguous()  # 转换为图像表示 (8,1024,8,8)
        pe = self.pe_generator(x)  # 条件位置编码
        x = x + pe  # 消融: PEG
        x = rearrange(x, ' b c h w -> b (h w) c').contiguous()  # 转换成序列表示
        for blk in self.blocks:
            if self.use_checkpoint:
                x = checkpoint.checkpoint(blk, x)
            else:
                x = blk(x)
        x = PatchReverse(x, self.patch_size)
        return x


class Swin_Transformer(nn.Module):
    """ A basic Swin Transformer layer for one stage.

    input:
        x: (B, N, C)

    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
        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 | tuple[float], optional): Stochastic depth rate. Default: 0.0
        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
        fused_window_process (bool, optional): If True, use one kernel to fused window shift & window partition for acceleration, similar for the reversed part. Default: False
        drop_key (bool, optional): Using dropkey or not. Default: False
    """

    def __init__(self, dim, input_resolution, depth, num_heads, window_size,
                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., norm_layer=nn.LayerNorm, use_checkpoint=False,
                 fused_window_process=False, drop_key=False):

        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.depth = depth
        self.use_checkpoint = use_checkpoint

        # build blocks
        # 构建一个 Swin Transformer Block
        self.blocks = nn.ModuleList([
            SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
                                 num_heads=num_heads, window_size=window_size,
                                 shift_size=0 if (i % 2 == 0) else window_size // 2,  # 消融：滑动窗口
                                 mlp_ratio=mlp_ratio,
                                 qkv_bias=qkv_bias, qk_scale=qk_scale,
                                 drop=drop, attn_drop=attn_drop,
                                 drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
                                 norm_layer=norm_layer,
                                 fused_window_process=fused_window_process, drop_key=drop_key)
            for i in range(depth)])

    def forward(self, x):
        for blk in self.blocks:
            if self.use_checkpoint:
                x = checkpoint.checkpoint(blk, x)
            else:
                x = blk(x)
        return x


class Channel_Adaptive_Transformer_Block(nn.Module):
    """Channel Adaptive Transformer Block (CATB).

    input:
        x: (B, C, H, W), feature map

    Args:
        dim (int): Number of input channels.
        input_resolution (int): Input resolution. based on patch
        num_heads (int): Number of attention heads.
        window_size (int): Local window size.
        patch_size (int): the size of the patch
        depth (int): Number of Swin Transformer blocks, default 2.
        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 | tuple[float], optional): Stochastic depth rate. Default: 0.0
        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
        fused_window_process (bool, optional): If True, use one kernel to fused window shift & window partition for acceleration, similar for the reversed part. Default: False
        drop_key (bool, optional): Using dropkey or not. Default: False
    """

    def __init__(self, dim, input_resolution, num_heads, window_size, patch_size,
                 depth=2, mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., norm_layer=nn.LayerNorm, use_checkpoint=False,
                 fused_window_process=False, drop_key=False):
        super().__init__()
        self.transformer_block = Swin_Transformer(dim, to_2tuple(input_resolution), depth, num_heads, window_size,
                                                  mlp_ratio, qkv_bias,
                                                  qk_scale, drop, attn_drop, drop_path, norm_layer, use_checkpoint,
                                                  fused_window_process, drop_key=drop_key)
        self.CA = Channel_Adapter(dim // (patch_size ** 2), input_resolution * patch_size)
        self.patch_size = patch_size
        self.window_size = window_size
        self.input_resolution = input_resolution

        self.a = dim
        self.b = patch_size

    def forward(self, x):
        x = self.CA(x)  # 消融: CA
        x = PatchDivide(x, self.patch_size)
        x = self.transformer_block(x)
        x = PatchReverse(x, self.patch_size)
        return x


class CATB_Layer(nn.Module):
    """Channel Adaptive Transformer Block Layer, compose with numbers of CATBs.

    input:
        x: (B, C, H, W), feature map

    Args:
        dim (int): Number of input channels.
        input_resolution (int): Input resolution. based on patch
        num_heads (int): Number of attention heads.
        window_size (int): Local window size.
        patch_size (int): the size of the patch
        depth (int): Number of CATB, default 1.
        depth_tr (int): Number of Swin Transformer blocks in each CATB, default 2.
        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 | tuple[float], optional): Stochastic depth rate. Default: 0.0
        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
        fused_window_process (bool, optional): If True, use one kernel to fused window shift & window partition for acceleration, similar for the reversed part. Default: False
        drop_key (bool, optional): Using dropkey or not. Default: False
    """

    def __init__(self, dim, input_resolution, num_heads, window_size, patch_size, depth=1,
                 depth_tr=2, mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., norm_layer=nn.LayerNorm, use_checkpoint=False,
                 fused_window_process=False, drop_key=False):

        super().__init__()
        self.use_checkpoint = use_checkpoint
        self.blocks = nn.ModuleList([
            Channel_Adaptive_Transformer_Block(dim, input_resolution, num_heads, window_size, patch_size, depth_tr,
                                               mlp_ratio, qkv_bias,
                                               qk_scale, drop, attn_drop, drop_path, norm_layer, use_checkpoint,
                                               fused_window_process, drop_key) for i in range(depth)])

    def forward(self, x):
        for blk in self.blocks:
            if self.use_checkpoint:
                x = checkpoint.checkpoint(blk, x)
            else:
                x = blk(x)
        return x


class OT_MAPPING(nn.Module):
    """ ot map"""

    def __init__(self):
        super().__init__()
        self.ot_emd = ot.da.EMDTransport()

    def forward(self, pool3, latent, mode):  # bottle (8,128,16,16)
        maps = []
        N, C, H, W = latent.shape
        latent = latent.reshape(N, C, H * W)
        white_noise = torch.randn(N, C, H * W).to('cuda:0')  # source (128,256)
        bottle = torch.zeros_like(latent)

        for i in range(N):
            self.ot_emd.fit(Xs=white_noise[i], Xt=latent[i])
            maps.append(self.ot_emd)

        for index, map in enumerate(maps):
            bottle[index] = map.transform(Xs=white_noise[index])
            if mode == 'reveal':
                bottle[index] += white_noise[index]

        bottle = bottle.reshape(N, C, H, W)
        return bottle


class StegFormer(nn.Module):
    """StegFormer

    Args:
        img_resolution (int): Resolution of images
        input_dim(int): Dim of the input
        output_dim (int): Dim of the finnal output
        cnn_emb_dim (int): Embedding dim using convolution
        output_act (nn.Module): The act function in the end of StegFormer, Default: None
        patch_size(int): the size of the patch
        num_heads (list): Number of attention heads.
        window_size (int): Local window size
        depth (list): Number of CATB in each CATB Layer
        depth_tr (list): Number of the Swin Transformer Block in each CATB
        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 | tuple[float], optional): Stochastic depth rate. Default: 0.0
        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
        drop_key (bool, optional): Using dropkey or not. Default: False
    """

    def __init__(self, img_resolution, input_dim=3, output_dim=3, cnn_emb_dim=16, output_act=None, patch_size=2,
                 num_heads=[1, 2, 4, 8, 16, 16, 8, 4, 2], window_size=8,
                 depth=[1, 1, 1, 1, 2, 1, 1, 1, 1], depth_tr=[2, 2, 2, 2, 2, 2, 2, 2], mlp_ratio=4., qkv_bias=True,
                 qk_scale=None, drop=0., attn_drop=0., mode=None,
                 drop_path=0., norm_layer=nn.LayerNorm, use_checkpoint=False, drop_key=False, pretrain_encoder=None):
        super().__init__()
        self.dim = cnn_emb_dim
        self.token_dim = (patch_size ** 2) * cnn_emb_dim
        self.embedding = nn.Conv2d(input_dim, self.dim, 3, 1, 1)
        self.patch_size = patch_size
        self.patch_resolution = img_resolution // patch_size
        self.mode = mode
        if output_act:
            self.output_act = output_act()
        else:
            self.output_act = None

        # encoder
        self.encoderlayer_0 = CATB_Layer(self.token_dim, self.patch_resolution, num_heads[0],
                                         window_size, patch_size, depth[0], depth_tr[0], mlp_ratio, qkv_bias, qk_scale,
                                         drop, attn_drop, drop_path, norm_layer, use_checkpoint, drop_key=drop_key)
        self.downsampler_0 = DownSampler(self.dim, self.dim * 2)

        self.encoderlayer_1 = CATB_Layer(self.token_dim * 2, self.patch_resolution // (2 ** 1),
                                         num_heads[1], window_size, patch_size, depth[1], depth_tr[1], mlp_ratio,
                                         qkv_bias, qk_scale, drop, attn_drop, drop_path, norm_layer, use_checkpoint,
                                         drop_key=drop_key)
        self.downsampler_1 = DownSampler(self.dim * 2, self.dim * 4)

        self.encoderlayer_2 = CATB_Layer(self.token_dim * 4, self.patch_resolution // (2 ** 2),
                                         num_heads[2], window_size, patch_size, depth[2], depth_tr[2], mlp_ratio,
                                         qkv_bias, qk_scale, drop, attn_drop, drop_path, norm_layer, use_checkpoint,
                                         drop_key=drop_key)
        self.downsampler_2 = DownSampler(self.dim * 4, self.dim * 8)

        self.encoderlayer_3 = CATB_Layer(self.token_dim * 8, self.patch_resolution // (2 ** 3),
                                         num_heads[3], window_size, patch_size, depth[3], depth_tr[3], mlp_ratio,
                                         qkv_bias, qk_scale, drop, attn_drop, drop_path, norm_layer, use_checkpoint,
                                         drop_key=drop_key)
        self.downsampler_3 = DownSampler(self.dim * 8, self.dim * 16)

        # bottleneck
        self.bottleneck = GEB(self.token_dim * 16, self.patch_resolution // (2 ** 4),
                              num_heads[4], patch_size, depth[4], mlp_ratio, qkv_bias, qk_scale, drop, attn_drop,
                              drop_path, norm_layer, use_checkpoint, mode=mode, drop_key=drop_key)

        # decoder
        self.upsampler_0 = UpSampler(self.dim * 16, self.dim * 8)
        self.decoderlayer_0 = CATB_Layer(self.token_dim * 16, self.patch_resolution // (2 ** 3),
                                         num_heads[5], window_size, patch_size, depth[5], depth_tr[4], mlp_ratio,
                                         qkv_bias, qk_scale, drop, attn_drop, drop_path, norm_layer, use_checkpoint,
                                         drop_key=drop_key)

        self.upsampler_1 = UpSampler(self.dim * 16, self.dim * 4)
        self.decoderlayer_1 = CATB_Layer(self.token_dim * 8, self.patch_resolution // (2 ** 2),
                                         num_heads[6], window_size, patch_size, depth[6], depth_tr[5], mlp_ratio,
                                         qkv_bias, qk_scale, drop, attn_drop, drop_path, norm_layer, use_checkpoint,
                                         drop_key=drop_key)

        self.upsampler_2 = UpSampler(self.dim * 8, self.dim * 2)
        self.decoderlayer_2 = CATB_Layer(self.token_dim * 4, self.patch_resolution // (2 ** 1),
                                         num_heads[7], window_size, patch_size, depth[7], depth_tr[6], mlp_ratio,
                                         qkv_bias, qk_scale, drop, attn_drop, drop_path, norm_layer, use_checkpoint,
                                         drop_key=drop_key)

        self.upsampler_3 = UpSampler(self.dim * 4, self.dim)
        self.decoderlayer_3 = CATB_Layer(self.token_dim * 2, self.patch_resolution,
                                         num_heads[8], window_size, patch_size, depth[8], depth_tr[7], mlp_ratio,
                                         qkv_bias, qk_scale, drop, attn_drop, drop_path, norm_layer, use_checkpoint,
                                         drop_key=drop_key)

        self.output_proj = nn.Conv2d(self.dim * 2, output_dim, 3, 1, 1)
        self.pretrain_encoder = pretrain_encoder

    def forward(self, x):
        # if self.mode == 'hide':
        #     latent = self.pretrain_encoder(x)
        # elif self.mode == 'reveal' and auxiliary is not None:
        #     latent = auxiliary
        # else:
        #     print("input error")
        latent = self.pretrain_encoder(x)
        x = self.embedding(x)  # (8,256,256)

        # encode
        conv0 = self.encoderlayer_0(x)
        pool0 = self.downsampler_0(conv0)

        conv1 = self.encoderlayer_1(pool0)
        pool1 = self.downsampler_1(conv1)

        conv2 = self.encoderlayer_2(pool1)
        pool2 = self.downsampler_2(conv2)

        conv3 = self.encoderlayer_3(pool2)
        pool3 = self.downsampler_3(conv3)

        # bottle neck
        bottle = self.bottleneck(pool3, latent)  # (128,16,16)

        # decode
        up0 = self.upsampler_0(bottle)
        deconv0 = torch.cat([up0, conv3], 1)
        deconv0 = self.decoderlayer_0(deconv0)

        up1 = self.upsampler_1(deconv0)
        deconv1 = torch.cat([up1, conv2], 1)
        deconv1 = self.decoderlayer_1(deconv1)

        up2 = self.upsampler_2(deconv1)
        deconv2 = torch.cat([up2, conv1], 1)
        deconv2 = self.decoderlayer_2(deconv2)

        up3 = self.upsampler_3(deconv2)
        deconv3 = torch.cat([up3, conv0], 1)
        deconv3 = self.decoderlayer_3(deconv3)

        output = self.output_proj(deconv3)
        if self.output_act:
            output = self.output_act(output)
        return output