import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange, repeat
from timm.models.layers import DropPath, to_2tuple, trunc_normal_
import timm
from timm.models import MobileNetV3

class ConvBNReLU(nn.Sequential):
    def __init__(self, in_channels, out_channels, kernel_size=3, dilation=1, stride=1, norm_layer=nn.BatchNorm2d,
                 bias=False):
        super(ConvBNReLU, self).__init__(
            nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, bias=bias,
                      dilation=dilation, stride=stride, padding=((stride - 1) + dilation * (kernel_size - 1)) // 2),
            norm_layer(out_channels),
            nn.ReLU6()
        )



class ConvBN(nn.Sequential):
    def __init__(self, in_channels, out_channels, kernel_size=3, dilation=1, stride=1, norm_layer=nn.BatchNorm2d,
                 bias=False):
        super(ConvBN, self).__init__(
            nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, bias=bias,
                      dilation=dilation, stride=stride, padding=((stride - 1) + dilation * (kernel_size - 1)) // 2),
            norm_layer(out_channels)
        )



class Conv(nn.Sequential):
    def __init__(self, in_channels, out_channels, kernel_size=3, dilation=1, stride=1, bias=False):
        super(Conv, self).__init__(
            nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, bias=bias,
                      dilation=dilation, stride=stride, padding=((stride - 1) + dilation * (kernel_size - 1)) // 2)
        )


class SeparableConvBNReLU(nn.Sequential):
    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, dilation=1,
                 norm_layer=nn.BatchNorm2d):
        super(SeparableConvBNReLU, self).__init__(
            nn.Conv2d(in_channels, in_channels, kernel_size, stride=stride, dilation=dilation,
                      padding=((stride - 1) + dilation * (kernel_size - 1)) // 2,
                      groups=in_channels, bias=False),
            norm_layer(out_channels),
            # 逐点卷积
            nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False),
            nn.ReLU6()
        )


class SeparableConvBN(nn.Sequential):
    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, dilation=1,
                 norm_layer=nn.BatchNorm2d):
        super(SeparableConvBN, self).__init__(
            nn.Conv2d(in_channels, in_channels, kernel_size, stride=stride, dilation=dilation,
                      padding=((stride - 1) + dilation * (kernel_size - 1)) // 2,
                      groups=in_channels, bias=False),
            norm_layer(out_channels),
            nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
        )


class SeparableConv(nn.Sequential):
    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, dilation=1):
        super(SeparableConv, self).__init__(
            nn.Conv2d(in_channels, in_channels, kernel_size, stride=stride, dilation=dilation,
                      padding=((stride - 1) + dilation * (kernel_size - 1)) // 2,
                      groups=in_channels, bias=False),
            nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
        )



class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.ReLU6, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        # 这行代码的意思是，如果out_features已经被提供了一个值，那么out_features就等于这个提供的值。如果没有为out_features提供值（即out_features为None），
        # 那么out_features将被设置为in_features的值。

        hidden_features = hidden_features or in_features
        self.fc1 = nn.Conv2d(in_features, hidden_features, 1, 1, 0, bias=True)

        self.act = act_layer()
        self.fc2 = nn.Conv2d(hidden_features, out_features, 1, 1, 0, bias=True)
        self.drop = nn.Dropout(drop, inplace=True)


    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 GlobalLocalAttention(nn.Module):
    def __init__(self,
                 dim=256,
                 num_heads=16,
                 qkv_bias=False,
                 window_size=8,
                 relative_pos_embedding=True
                 ):

        super().__init__()
        self.num_heads = num_heads  # 初始化注意力头的数量
        head_dim = dim // self.num_heads  # 计算每个注意力头的特征维度
        self.scale = head_dim ** -0.5  # 计算缩放因子，用于注意力计算中的点积。
        self.ws = window_size  # 初始化局部窗口的大小。

        self.qkv = Conv(dim, 3 * dim, kernel_size=1, bias=qkv_bias)  # 初始化一个卷积层，用于生成Query、Key和Value。
        self.local1 = ConvBN(dim, dim, kernel_size=3)  # 初始化第一个卷积层和批量归一化层，用于处理局部特征。
        self.local2 = ConvBN(dim, dim, kernel_size=1)  # 初始化第二个卷积层和批量归一化层，用于处理局部特征。
        self.proj = SeparableConvBN(dim, dim, kernel_size=window_size)  # 初始化一个可分离卷积层，用于投影输出。

        self.attn_x = nn.AvgPool2d(kernel_size=(window_size, 1), stride=1,
                                   padding=(window_size // 2 - 1, 0))  # 初始化水平方向的平均池化层，用于整合全局信息。
        self.attn_y = nn.AvgPool2d(kernel_size=(1, window_size), stride=1,
                                   padding=(0, window_size // 2 - 1))  # 初始化垂直方向的平均池化层，用于整合全局信息。
        self.relative_pos_embedding = relative_pos_embedding
        # 初始化是否使用相对位置嵌入的标志。
        if self.relative_pos_embedding:  # 如果使用了相对位置嵌入，会定义一个相对位置偏置表
            # define a parameter table of relative position bias
            self.relative_position_bias_table = nn.Parameter(
                torch.zeros((2 * window_size - 1) * (2 * window_size - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH

            coords_h = torch.arange(self.ws)  # 创建一个包含窗口大小ws内所有水平坐标的张量。
            coords_w = torch.arange(self.ws)  # 创建一个包含窗口大小ws内所有垂直坐标的张量。
            coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww #使用meshgrid函数创建一个包含所有水平和垂直坐标的张量。
            coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww将三维坐标张量展平为一维张量。

            relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]
            # 2, Wh*Ww, Wh*Ww 计算所有坐标对之间的相对位置，即每个坐标相对于其他所有坐标的差

            relative_coords = relative_coords.permute(1, 2, 0).contiguous()

            relative_coords[:, :, 0] += self.ws - 1  # shift to start from 0将相对坐标的第一个维度增加ws - 1，以确保坐标从0开始。
            relative_coords[:, :, 1] += self.ws - 1  # 将相对坐标的第二个维度增加ws - 1，以确保坐标从0开始。
            relative_coords[:, :, 0] *= 2 * self.ws - 1  # 调整相对坐标的第一个维度，使其范围变为[-2*ws+1, 2*ws-1]。

            relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww将相对坐标的两个维度合并为一个单一的索引，用于访问相对位置偏置表。
            self.register_buffer("relative_position_index", relative_position_index)  # 将相对位置索引注册为一个缓冲区，以便在模型训练过程中重复使用。

            trunc_normal_(self.relative_position_bias_table, std=.02)
            # 使用trunc_normal_函数初始化相对位置偏置表，这是一种常用的初始化技术，用于生成服从截断正态分布的参数。

    def pad(self, x, ps):  # 定义一个函数，接受一个特征图x和一个填充大小ps作为参数。
        _, _, H, W = x.size()  # 获取特征图x的形状，并提取高度H和宽度W。
        if W % ps != 0:  # 如果特征图的宽度W不能被填充大小ps整除，则需要进行填充。
            x = F.pad(x, (0, ps - W % ps),
                      mode='reflect')  # 使用F.pad函数在特征图的右侧添加填充，填充大小为ps - W % ps，填充模式为'reflect'，这意味着新的像素值将反映原始像素值。
        if H % ps != 0:  # 如果特征图的高度H不能被填充大小ps整除，则需要进行额外的填充。
            x = F.pad(x, (0, 0, 0, ps - H % ps),
                      mode='reflect')  # 使用F.pad函数在特征图的下方添加填充，填充大小为ps - H % ps，填充模式为'reflect'。
        return x  # 返回填充后的特征图。

    def pad_out(self, x):  # 定义一个函数，接受一个特征图x作为参数。
        x = F.pad(x, pad=(0, 1, 0, 1), mode='reflect')  # 使用F.pad函数在特征图的右侧和下方添加填充，填充大小为1，填充模式为'reflect'。
        return x

    def forward(self, x):
        B, C, H, W = x.shape
        local = self.local2(x) + self.local1(x)  # 计算局部特征，通过两个卷积层。
        x = self.pad(x, self.ws)  # 填充输入特征图以适应窗口大小。
        B, C, Hp, Wp = x.shape  # 获取填充后的特征图的形状。
        qkv = self.qkv(x)  # 生成Query、Key和Value。
        q, k, v = rearrange(qkv, 'b (qkv h d) (hh ws1) (ww ws2) -> qkv (b hh ww) h (ws1 ws2) d', h=self.num_heads,
                            d=C // self.num_heads, hh=Hp // self.ws, ww=Wp // self.ws, qkv=3, ws1=self.ws,
                            ws2=self.ws)  # 重新排列Query、Key和Value以适应注意力机制的计算。
        dots = (q @ k.transpose(-2, -1)) * self.scale  # 计算点积，并应用缩放因子。
        # 如果使用了相对位置嵌入，将相对位置偏置加到点积上。
        if self.relative_pos_embedding:  # 如果启用了相对位置嵌入
            relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
                self.ws * self.ws, self.ws * self.ws, -1)  # Wh*Ww,Wh*Ww,nH获取相对位置偏置表，并根据相对位置索引进行调整。
            relative_position_bias = relative_position_bias.permute(2, 0,
                                                                    1).contiguous()  # nH, Wh*Ww, Wh*Ww重新排列相对位置偏置，以便与点积的形状匹配
            dots += relative_position_bias.unsqueeze(0)  # 将相对位置偏置加到点积上。

        attn = dots.softmax(dim=-1)  # 应用softmax函数计算注意力权重。
        attn = attn @ v  # 注意力权重应用于Value。

        attn = rearrange(attn, '(b hh ww) h (ws1 ws2) d -> b (h d) (hh ws1) (ww ws2)', h=self.num_heads,
                         d=C // self.num_heads, hh=Hp // self.ws, ww=Wp // self.ws, ws1=self.ws, ws2=self.ws)
        attn = attn[:, :, :H, :W]  # 裁剪注意力权重，使其与原始输入特征图的形状匹配。

        out = self.attn_x(F.pad(attn, pad=(0, 0, 0, 1), mode='reflect')) + \
              self.attn_y(F.pad(attn, pad=(0, 1, 0, 0), mode='reflect'))
        out = out + local  # 将局部特征与全局特征相加。
        out = self.pad_out(out)  # 添加额外的填充，以适应输出特征图的尺寸。
        out = self.proj(out)
        # print(out.size())
        out = out[:, :, :H, :W]
        return out


class Block(nn.Module):
    def __init__(self, dim=256, num_heads=16, mlp_ratio=4., qkv_bias=False, drop=0., attn_drop=0.,
                 drop_path=0., act_layer=nn.ReLU6, norm_layer=nn.BatchNorm2d, window_size=8):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = GlobalLocalAttention(dim, num_heads=num_heads, qkv_bias=qkv_bias, window_size=window_size)
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, out_features=dim, act_layer=act_layer,
                       drop=drop)
        self.norm2 = norm_layer(dim)
    def forward(self, x):
        x = x + self.drop_path(self.attn(self.norm1(x)))
        x = x + self.drop_path(self.mlp(self.norm2(x)))
        return x

class WF(nn.Module):
    def __init__(self, in_channels=128, decode_channels=128, eps=1e-8):
        super(WF, self).__init__()
        self.pre_conv = Conv(in_channels, decode_channels, kernel_size=1)

        self.weights = nn.Parameter(torch.ones(2, dtype=torch.float32), requires_grad=True)
        self.eps = eps
        self.post_conv = ConvBNReLU(decode_channels, decode_channels, kernel_size=3)

    def forward(self, x, res):
        x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)
        weights = nn.ReLU()(
            self.weights)
        fuse_weights = weights / (torch.sum(weights, dim=0) + self.eps)
        x = fuse_weights[0] * self.pre_conv(res) + fuse_weights[1] * x
        x = self.post_conv(x)
        return x


class FeatureRefinementHead(nn.Module):
    def __init__(self, in_channels=64, decode_channels=64):
        super().__init__()
        self.pre_conv = Conv(in_channels, decode_channels, kernel_size=1)

        self.weights = nn.Parameter(torch.ones(2, dtype=torch.float32), requires_grad=True)
        self.eps = 1e-8

        self.post_conv = ConvBNReLU(decode_channels, decode_channels, kernel_size=3)
        self.pa = nn.Sequential(
            nn.Conv2d(decode_channels, decode_channels, kernel_size=3, padding=1, groups=decode_channels),
            nn.Sigmoid())
        self.ca = nn.Sequential(nn.AdaptiveAvgPool2d(1),
                                Conv(decode_channels, decode_channels // 16, kernel_size=1),
                                nn.ReLU6(),
                                Conv(decode_channels // 16, decode_channels, kernel_size=1),
                                nn.Sigmoid())

        self.shortcut = ConvBN(decode_channels, decode_channels, kernel_size=1)
        self.proj = SeparableConvBN(decode_channels, decode_channels, kernel_size=3)
        self.act = nn.ReLU6()

    def forward(self, x, res):
        x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)
        weights = nn.ReLU()(self.weights)
        fuse_weights = weights / (torch.sum(weights, dim=0) + self.eps)
        x = fuse_weights[0] * self.pre_conv(res) + fuse_weights[1] * x
        x = self.post_conv(x)
        shortcut = self.shortcut(x)
        pa = self.pa(x) * x
        ca = self.ca(x) * x
        x = pa + ca
        x = self.proj(x) + shortcut
        x = self.act(x)
        return x

class AuxHead(nn.Module):
    def __init__(self, in_channels=64, num_classes=8):
        super().__init__()
        self.conv = ConvBNReLU(in_channels, in_channels)
        self.drop = nn.Dropout(0.1)
        self.conv_out = Conv(in_channels, num_classes, kernel_size=1)

    def forward(self, x, h, w):
        feat = self.conv(x)
        feat = self.drop(feat)
        feat = self.conv_out(feat)
        feat = F.interpolate(feat, size=(h, w), mode='bilinear', align_corners=False)
        return feat

class Decoder(nn.Module):
    def __init__(self,
                 encoder_channels=(64, 128, 256, 512),
                 decode_channels=64,
                 dropout=0.1,
                 window_size=8,
                 num_classes=6):
        super(Decoder, self).__init__()
        self.pre_conv = ConvBN(encoder_channels[-1], decode_channels, kernel_size=1)
        self.b4 = Block(dim=decode_channels, num_heads=8, window_size=window_size)
        self.b3 = Block(dim=decode_channels, num_heads=8, window_size=window_size)
        self.p3 = WF(encoder_channels[-2], decode_channels)  # 三个WS模块
        self.b2 = Block(dim=decode_channels, num_heads=8, window_size=window_size)
        self.p2 = WF(encoder_channels[-3], decode_channels)

        if self.training:
            self.up4 = nn.UpsamplingBilinear2d(scale_factor=4)
            self.up3 = nn.UpsamplingBilinear2d(scale_factor=2)
            self.aux_head = AuxHead(decode_channels, num_classes)
        self.p1 = FeatureRefinementHead(encoder_channels[-4], decode_channels)
        self.segmentation_head = nn.Sequential(ConvBNReLU(decode_channels, decode_channels),
                                               nn.Dropout2d(p=dropout, inplace=True),
                                               Conv(decode_channels, num_classes, kernel_size=1))
        self.init_weight()

    def forward(self, res1, res2, res3, res4, h, w):
        if self.training == True:
            x = self.b4(self.pre_conv(res4))
            h4 = self.up4(x)
            x = self.p3(x, res3)
            x = self.b3(x)
            h3 = self.up3(x)
            x = self.p2(x, res2)
            x = self.b2(x)
            h2 = x
            x = self.p1(x, res1)
            x = self.segmentation_head(x)
            x = F.interpolate(x, size=(h, w), mode='bilinear', align_corners=False)
            ah = h4 + h3 + h2
            ah = self.aux_head(ah, h, w)
            return x, ah
        else:
            x = self.b4(self.pre_conv(res4))
            x = self.p3(x, res3)
            x = self.b3(x)

            x = self.p2(x, res2)
            x = self.b2(x)

            x = self.p1(x, res1)

            x = self.segmentation_head(x)
            x = F.interpolate(x, size=(h, w), mode='bilinear', align_corners=False)

            return x

    def init_weight(self):
        for m in self.children():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, a=1)
                if m.bias is not None:
                    nn.init.constant_(m.bias, 0)


class FuseDisNet(nn.Module):
    def __init__(self,
                 decode_channels=64,
                 dropout=0.1,
                 backbone_name='swsl_resnet18',
                 pretrained=True,
                 window_size=8,
                 num_classes=6
                 ):
        super().__init__()

        self.backbone = timm.create_model(backbone_name, features_only=True, output_stride=32,
                                          out_indices=(1, 2, 3, 4), pretrained=pretrained)

        encoder_channels = self.backbone.feature_info.channels()

        self.decoder = Decoder(encoder_channels, decode_channels, dropout, window_size, num_classes)
        # 定义了一个解码器Decoder，传入了主干网络提取的特征通道数、解码器的参数等。

    def forward(self, x):
        h, w = x.size()[-2:]

        res1, res2, res3, res4 = self.backbone(x)
        if self.training:
            x, ah = self.decoder(res1, res2, res3, res4, h, w)
            return x, ah
        else:
            x = self.decoder(res1, res2, res3, res4, h, w)
            return x