import torch
import torch.nn as nn
import torch.nn.functional as F

"""
含残差的双层三明治结构模型
第一层：时间 -> 空间 -> 时间 -> Conv -> Residual差分 -> 输入第二层
第二层：时间 -> 空间 -> 时间 -> Conv -> 最终输出
无效但接近
"""

class DynamicGraphConstructor(nn.Module):
    def __init__(self, node_num, embed_dim):
        super().__init__()
        self.nodevec1 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
        self.nodevec2 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)

    def forward(self):
        # (N, D) @ (D, N) -> (N, N)
        adj = torch.matmul(self.nodevec1, self.nodevec2.T)
        adj = F.relu(adj)
        adj = F.softmax(adj, dim=-1)
        return adj


class GraphConvBlock(nn.Module):
    def __init__(self, input_dim, output_dim):
        super().__init__()
        self.theta = nn.Linear(input_dim, output_dim)
        self.residual = (input_dim == output_dim)
        if not self.residual:
            self.res_proj = nn.Linear(input_dim, output_dim)

    def forward(self, x, adj):
        # x: (B, N, C); adj: (N, N)
        res = x
        x = torch.matmul(adj, x)         # 空间卷积
        x = self.theta(x)
        # 残差
        x = x + (res if self.residual else self.res_proj(res))
        return F.relu(x)


class MANBA_Block(nn.Module):
    def __init__(self, input_dim, hidden_dim):
        super().__init__()
        self.attn = nn.MultiheadAttention(embed_dim=input_dim, num_heads=4, batch_first=True)
        self.ffn = nn.Sequential(
            nn.Linear(input_dim, hidden_dim),
            nn.ReLU(),
            nn.Linear(hidden_dim, input_dim)
        )
        self.norm1 = nn.LayerNorm(input_dim)
        self.norm2 = nn.LayerNorm(input_dim)

    def forward(self, x):
        # x: (B, N, C) 视 N 维为时间序列长度
        res = x
        x_attn, _ = self.attn(x, x, x)
        x = self.norm1(res + x_attn)
        res2 = x
        x_ffn = self.ffn(x)
        x = self.norm2(res2 + x_ffn)
        return x


class SandwichBlock(nn.Module):
    """
    时间-空间-时间 三明治结构
    输入/输出： (B, N, hidden_dim)
    """
    def __init__(self, num_nodes, embed_dim, hidden_dim):
        super().__init__()
        self.manba1 = MANBA_Block(hidden_dim, hidden_dim * 2)
        self.graph_constructor = DynamicGraphConstructor(num_nodes, embed_dim)
        self.gc = GraphConvBlock(hidden_dim, hidden_dim)
        self.manba2 = MANBA_Block(hidden_dim, hidden_dim * 2)

    def forward(self, h):
        # h: (B, N, hidden_dim)
        h1 = self.manba1(h)
        adj = self.graph_constructor()         # (N, N)
        h2 = self.gc(h1, adj)
        h3 = self.manba2(h2)
        return h3


class EXP(nn.Module):
    def __init__(self, args):
        super().__init__()
        self.horizon = args['horizon']
        self.output_dim = args['output_dim']
        self.seq_len = args.get('in_len', 12)
        self.hidden_dim = args.get('hidden_dim', 64)
        self.num_nodes = args['num_nodes']
        self.embed_dim = args.get('embed_dim', 16)

        # 输入映射: (batch*N, seq_len) -> hidden_dim
        self.input_proj = nn.Linear(self.seq_len, self.hidden_dim)
        # 两层三明治块
        self.sandwich1 = SandwichBlock(self.num_nodes, self.embed_dim, self.hidden_dim)
        self.sandwich2 = SandwichBlock(self.num_nodes, self.embed_dim, self.hidden_dim)
        # 卷积层用于残差处理
        self.res_conv1 = nn.Conv1d(in_channels=self.hidden_dim, out_channels=self.hidden_dim, kernel_size=1)
        self.res_conv2 = nn.Conv1d(in_channels=self.hidden_dim, out_channels=self.hidden_dim, kernel_size=1)
        # 输出映射
        self.out_proj = nn.Linear(self.hidden_dim, self.horizon * self.output_dim)

    def forward(self, x):
        # x: (B, T, N, D_total)
        x_main = x[..., 0]              # (B, T, N)
        B, T, N = x_main.shape
        assert T == self.seq_len

        # 输入投影 (B, T, N) -> (B*N, T) -> (B, N, hidden_dim)
        x_flat = x_main.permute(0, 2, 1).reshape(B * N, T)
        h0 = self.input_proj(x_flat).view(B, N, self.hidden_dim)

        # 第一层三明治
        h1_sand = self.sandwich1(h0)   # (B, N, hidden_dim)
        # 卷积残差 (节点维度视为长度)
        h1_perm = h1_sand.permute(0, 2, 1)  # (B, C, N)
        h1_conv = self.res_conv1(h1_perm)
        h1 = h1_conv.permute(0, 2, 1)       # (B, N, hidden_dim)

        # 计算差分残差作为第二层输入
        h2_input = h1 - h0

        # 第二层三明治
        h2_sand = self.sandwich2(h2_input)
        # 再次卷积处理
        h2_perm = h2_sand.permute(0, 2, 1)
        h2 = self.res_conv2(h2_perm).permute(0, 2, 1)

        # 输出映射
        out = self.out_proj(h2)          # (B, N, H*D_out)
        out = out.view(B, N, self.horizon, self.output_dim)
        out = out.permute(0, 2, 1, 3)    # (B, horizon, N, output_dim)
        return out