import torch
from torch import nn

class AdvancedCrimeDetectionModel(nn.Module):
    def __init__(self, input_dim=1536, hidden=512):
        super().__init__()

        self.input_proj = nn.Sequential(
            nn.Linear(input_dim, hidden),
            nn.LayerNorm(hidden),
            nn.ReLU(),
            nn.Dropout(0.3)
        )

        encoder_layer = nn.TransformerEncoderLayer(
            d_model=hidden,
            nhead=8,
            dim_feedforward=hidden * 2,
            dropout=0.2,
            batch_first=True
        )
        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=3)

        self.gru = nn.GRU(
            hidden,
            hidden // 2,
            num_layers=3,
            bidirectional=True,
            batch_first=True,
            dropout=0.3
        )

        self.attention = nn.MultiheadAttention(
            embed_dim=hidden,
            num_heads=8,
            dropout=0.2,
            batch_first=True
        )

        self.fc1 = nn.Linear(hidden, 384)
        self.bn1 = nn.BatchNorm1d(384)
        self.fc2 = nn.Linear(384, 192)
        self.bn2 = nn.BatchNorm1d(192)
        self.fc3 = nn.Linear(192, 96)
        self.bn3 = nn.BatchNorm1d(96)
        self.fc4 = nn.Linear(96, 1)

        self.dropout = nn.Dropout(0.4)
        self.relu = nn.ReLU()

    def forward(self, x):
        x = self.input_proj(x)
        x = self.transformer(x)
        gru_out, _ = self.gru(x)
        attn_out, _ = self.attention(gru_out, gru_out, gru_out)
        pooled = torch.mean(attn_out, dim=1)

        x1 = self.dropout(self.relu(self.bn1(self.fc1(pooled))))
        x2 = self.dropout(self.relu(self.bn2(self.fc2(x1))))
        x3 = self.dropout(self.relu(self.bn3(self.fc3(x2))))
        out = self.fc4(x3)

        return out.squeeze(1)