import torch
import torch.nn as nn

class EgoHRVPipeline(nn.Module):
    """
    EgoHRV: 从眼动视频到HRV的两阶段pipeline
    
    Stage 1: 跨域预训练 - 在接触式BVP数据上训练频域编码器
    Stage 2: 推理 - 将冻结的编码器应用于眼动视频重建的BVP
    
    核心创新：频域（Fourier幅度谱）跨域迁移
    - 丢弃相位，只保留幅度 → 消除传感器位置差异
    """
    
    def __init__(self):
        super().__init__()
        # Stage 1: 3D backbone + 低高频分解
        self.backbone = ThreeDBackbone()
        self.low_high_decomp = LowHighDecomposition()
        
        # Stage 2: 频域编码器（在接触式BVP上预训练后冻结）
        self.frequency_encoder = UNetEncoder(
            input_channels=1,
            feat_dim=64
        )
        self.ibi_predictor = IBIPredictor(feat_dim=64)
    
    def forward(self, gaze_video):
        """
        Args:
            gaze_video: 眼动摄像头视频 (B, T, C, H, W)
        
        Returns:
            ibi: 预测的inter-beat-interval
            log_var: 不确定性估计
        """
        # 1. 3D backbone提取时空特征
        features = self.backbone(gaze_video)
        
        # 2. 低高频分解（分离运动噪声和脉搏信号）
        bvp_signal = self.low_high_decomp(features)
        
        # 3. Fourier幅度谱（丢弃相位）
        bvp_fft = torch.fft.rfft(bvp_signal).abs()
        
        # 4. 预训练编码器推理（冻结）
        encoded = self.frequency_encoder(bvp_fft)
        
        # 5. IBI预测 + 不确定性
        ibi, log_var = self.ibi_predictor(encoded)
        
        return ibi, log_var


class LowHighDecomposition(nn.Module):
    """
    低高频分解模块：可学习的自适应滤波器
    
    分离：
    - 低频：脉搏信号（0.5-4Hz对应30-240bpm）
    - 高频：运动噪声
    
    通过门控残差连接融合
    """
    
    def __init__(self):
        super().__init__()
        self.low_filter = nn.Conv1d(1, 1, 15, 1, 7)
        self.high_filter = nn.Conv1d(1, 1, 5, 1, 2)
        self.gate = nn.Sigmoid()
    
    def forward(self, x):
        low = self.low_filter(x)
        high = self.high_filter(x)
        gate = self.gate(low.mean(dim=-1, keepdim=True))
        return low + gate * high


class PhysFusion(nn.Module):
    """
    PhysFusion: 不确定性感知的HRV+视频融合
    
    用预测的log_var作为不确定性权重
    下游任务（驾驶技能评估）提升17.8%
    """
    
    def __init__(self, feat_dim=64, num_heads=4):
        super().__init__()
        self.attention = nn.MultiheadAttention(
            feat_dim, num_heads, batch_first=True
        )
        self.uncertainty_weight = nn.Linear(1, feat_dim)
    
    def forward(self, video_features, hrv_features, log_var):
        # 不确定性加权：高不确定性窗口降权
        weight = torch.exp(-self.uncertainty_weight(log_var))
        weighted_hrv = hrv_features * weight
        
        # 跨模态注意力融合
        fused, _ = self.attention(
            video_features, weighted_hrv, weighted_hrv
        )
        
        return fused

"""
跨域预训练核心思想：

1. 在大规模接触式BVP数据集（WildPPG, DeepPPG）上训练
   - 输入：BVP信号的Fourier幅度谱
   - 目标：ECG推导的IBI
   - 损失：负对数似然（输出均值+log方差）

2. 冻结编码器，应用于眼动视频重建的BVP
   - 只使用幅度谱（丢弃相位）
   - 消除传感器位置差异（接触式vs摄像头）
   - 频域特征对运动噪声更鲁棒

关键发现：频域比时域更适合跨域迁移
"""

pretraining_config = {
    "source_domain": "Contact BVP (WildPPG, DeepPPG)",
    "target_domain": "Egocentric gaze video → reconstructed BVP",
    "transfer_method": "Fourier magnitude spectrum",
    "phase": "Discarded (eliminates sensor-site mismatch)",
    "encoder": "U-Net (frozen after pretraining)",
    "loss": "NLL: -0.5 * (1 + log_var - (target - pred)^2 / exp(log_var))",
    "window": "4.3 seconds",
    "key_insight": "频域幅度谱在接触式和摄像头BVP间具有更好的域不变性"
}

class EgoHRVDMSIntegration:
    """
    EgoHRV + 视觉DMS融合架构
    
    利用现有DMS摄像头实现：
    - 视觉层：疲劳/分心/视线检测（已有）
    - 生理层：HRV压力/认知负荷监测（新增，零BOM）
    - 融合层：多维度驾驶员状态评估
    """
    
    def assess_driver_state(self, video_frame, gaze_frame):
        # 视觉DMS（已有）
        visual_state = self.visual_dms(video_frame)
        # {fatigue_level, distraction_score, gaze_zone, head_pose}
        
        # EgoHRV（新增）
        hrv_metrics = self.egohrv(gaze_frame)
        # {heart_rate, sdnn, rmssd, stress_index, cognitive_load}
        
        # 融合评估
        if visual_state.fatigue_level > 0.6 and hrv_metrics.sdnn < baseline:
            return "SEVERE_FATIGUE", "LEVEL_3_WARNING"
        elif hrv_metrics.stress_index > 0.7:
            return "HIGH_STRESS", "COMFORT_MODE_ADJUST"
        elif visual_state.distraction_score > 0.5:
            return "DISTRACTED", "LEVEL_1_WARNING"
        else:
            return "NORMAL", "MAINTAIN"