src/model/deep.py

# Generated by Claude Code -- 2026-02-08
"""Model 3: Physics-Informed Temporal Fusion Transformer (PI-TFT).

Architecture overview (think of it like reading serial lab values):

1. VARIABLE SELECTION: Not all 22 CDM features matter equally. The model
   learns attention weights over features -- e.g., miss_distance and
   covariance shrinkage rate might matter more than raw orbital elements.
   This is like a doctor learning which labs to focus on.

2. STATIC CONTEXT: Object properties (altitude, size, eccentricity) don't
   change between CDM updates. They're encoded once and injected as context
   into the temporal processing. Like knowing the patient's age and history.

3. CONTINUOUS TIME EMBEDDING: CDMs arrive at irregular intervals (not evenly
   spaced). Instead of positional encoding (position 1, 2, 3...), we embed
   the actual time_to_tca value. The model knows "this CDM was 3.2 days
   before closest approach" vs "this one was 0.5 days before."

4. TEMPORAL SELF-ATTENTION: The Transformer reads the full CDM sequence and
   learns which updates were most informative. A sudden miss distance drop
   at day -2 gets more attention than a stable reading at day -5.

5. PREDICTION HEADS: The final hidden state (from the most recent CDM)
   feeds into two prediction heads:
   - Risk classifier: sigmoid probability of high-risk collision
   - Miss distance regressor: predicted log(miss distance in km)

6. PHYSICS LOSS: The training loss includes a penalty when the model predicts
   a miss distance BELOW the Minimum Orbital Intersection Distance (MOID).
   MOID is the closest the two orbits can geometrically get. Predicting
   closer than MOID is physically impossible (without a maneuver), so we
   penalize it. This is like penalizing a model for predicting negative
   blood pressure -- constraining outputs to the physically possible range.
"""

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


class GatedResidualNetwork(nn.Module):
    """
    Gated skip connection with ELU activation and layer norm.

    Think of this as a "smart residual block" -- it learns how much of the
    transformed input to mix with the original. The gate (sigmoid) controls
    this: gate=0 means pass through unchanged, gate=1 means fully transformed.
    """

    def __init__(self, d_model: int, d_hidden: int = None, dropout: float = 0.1):
        super().__init__()
        d_hidden = d_hidden or d_model
        self.fc1 = nn.Linear(d_model, d_hidden)
        self.fc2 = nn.Linear(d_hidden, d_model)
        self.gate_fc = nn.Linear(d_hidden, d_model)
        self.norm = nn.LayerNorm(d_model)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        residual = x
        h = F.elu(self.fc1(x))
        h = self.dropout(h)
        transform = self.fc2(h)
        gate = torch.sigmoid(self.gate_fc(h))
        return self.norm(residual + gate * transform)


class VariableSelectionNetwork(nn.Module):
    """
    Learns which input features matter most via softmax attention.

    For N input features, produces N attention weights that sum to 1.
    Each feature is independently projected to d_model, then weighted
    and summed. The weights are interpretable -- they tell you which
    CDM columns the model found most predictive.
    """

    def __init__(self, n_features: int, d_model: int, dropout: float = 0.1):
        super().__init__()
        self.n_features = n_features
        self.d_model = d_model

        # Each feature gets its own linear projection: scalar -> d_model vector
        self.feature_projections = nn.ModuleList([
            nn.Linear(1, d_model) for _ in range(n_features)
        ])

        # Gating network: takes flattened projections -> feature weights
        self.gate_network = nn.Sequential(
            nn.Linear(n_features * d_model, n_features),
            nn.Softmax(dim=-1),
        )

        self.grn = GatedResidualNetwork(d_model, dropout=dropout)

    def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        """
        Args:
            x: (..., n_features) — can be (B, F) for static or (B, S, F) for temporal

        Returns:
            output: (..., d_model) — weighted combination of projected features
            weights: (..., n_features) — attention weights (sum to 1)
        """
        # Project each feature independently
        # x[..., i:i+1] is the i-th feature, shape (..., 1)
        projected = [proj(x[..., i:i+1]) for i, proj in enumerate(self.feature_projections)]
        # projected[i] shape: (..., d_model)

        # Stack for gating: (..., n_features, d_model)
        stacked = torch.stack(projected, dim=-2)

        # Flatten for gate computation: (..., n_features * d_model)
        flat = stacked.reshape(*stacked.shape[:-2], -1)
        weights = self.gate_network(flat)  # (..., n_features)

        # Weighted sum: (..., d_model)
        output = (stacked * weights.unsqueeze(-1)).sum(dim=-2)
        output = self.grn(output)

        return output, weights


class PhysicsInformedTFT(nn.Module):
    """
    Physics-Informed Temporal Fusion Transformer for conjunction assessment.

    Input flow:
      temporal_features (B, S, F_t) → Variable Selection → time embedding → self-attention → attention pool → heads
      static_features   (B, F_s)    → Variable Selection → context injection ↗

    Output:
      risk_logit: (B, 1) — raw logit for risk classification (apply sigmoid for probability)
      miss_log:   (B, 1) — predicted log1p(miss_distance_km)
      pc_log10:   (B, 1) — predicted log10(Pc) collision probability (when has_pc_head=True)
      feature_weights: (B, S, F_t) — which temporal features mattered
    """

    def __init__(
        self,
        n_temporal_features: int,
        n_static_features: int,
        d_model: int = 128,
        n_heads: int = 4,
        n_layers: int = 2,
        dropout: float = 0.15,
        max_seq_len: int = 30,
    ):
        super().__init__()
        self.d_model = d_model
        self.max_seq_len = max_seq_len

        # --- Variable Selection Networks ---
        self.temporal_vsn = VariableSelectionNetwork(n_temporal_features, d_model, dropout)
        self.static_vsn = VariableSelectionNetwork(n_static_features, d_model, dropout)

        # --- Static context encoding ---
        self.static_encoder = nn.Sequential(
            nn.Linear(d_model, d_model),
            nn.GELU(),
            nn.Dropout(dropout),
        )
        # Static -> enrichment vector that's added to each temporal step
        self.static_to_enrichment = nn.Linear(d_model, d_model)

        # --- Continuous time embedding ---
        # Instead of fixed positional encoding, we embed the actual time_to_tca
        self.time_embedding = nn.Sequential(
            nn.Linear(1, d_model // 2),
            nn.GELU(),
            nn.Linear(d_model // 2, d_model),
        )

        # --- Transformer encoder layers ---
        encoder_layer = nn.TransformerEncoderLayer(
            d_model=d_model,
            nhead=n_heads,
            dim_feedforward=d_model * 2,
            dropout=dropout,
            activation="gelu",
            batch_first=True,
            norm_first=True,
        )
        self.transformer_encoder = nn.TransformerEncoder(
            encoder_layer, num_layers=n_layers
        )

        # --- Pre/post attention processing ---
        self.pre_attn_grn = GatedResidualNetwork(d_model, dropout=dropout)
        self.post_attn_grn = GatedResidualNetwork(d_model, dropout=dropout)

        # --- Attention-weighted pooling ---
        # Learns which time steps matter most instead of just taking the last one.
        # Softmax attention over all real positions, with padding masked out.
        self.pool_attention = nn.Sequential(
            nn.Linear(d_model, d_model // 2),
            nn.Tanh(),
            nn.Linear(d_model // 2, 1),
        )

        # --- Prediction heads ---
        self.risk_head = nn.Sequential(
            nn.LayerNorm(d_model),
            nn.Linear(d_model, 64),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(64, 1),
        )

        self.miss_head = nn.Sequential(
            nn.LayerNorm(d_model),
            nn.Linear(d_model, 64),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(64, 1),
        )

        # --- Collision probability head ---
        # Predicts log10(Pc) directly instead of binary risk classification.
        # Pc ranges from ~1e-20 to ~1e-1, so log10 scale maps to [-20, -1].
        # The Kelvins `risk` column is already log10(Pc).
        self.pc_head = nn.Sequential(
            nn.LayerNorm(d_model),
            nn.Linear(d_model, 64),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(64, 1),
        )

    def encode_sequence(
        self,
        temporal_features: torch.Tensor,  # (B, S, F_t)
        static_features: torch.Tensor,    # (B, F_s)
        time_to_tca: torch.Tensor,        # (B, S, 1)
        mask: torch.Tensor,               # (B, S) — True for real, False for padding
    ):
        """Encode CDM sequence into per-timestep hidden states.

        Returns:
            hidden: (B, S, D) per-timestep representations after Transformer
            temporal_weights: (B, S, F_t) variable selection attention weights
        """
        # 1. Variable selection -- learn which features matter
        temporal_selected, temporal_weights = self.temporal_vsn(temporal_features)
        # temporal_selected: (B, S, D), temporal_weights: (B, S, F_t)

        static_selected, static_weights = self.static_vsn(static_features)
        # static_selected: (B, D)

        # 2. Static context -- compute enrichment vector
        static_ctx = self.static_encoder(static_selected)  # (B, D)
        enrichment = self.static_to_enrichment(static_ctx)  # (B, D)

        # 3. Continuous time embedding
        t_embed = self.time_embedding(time_to_tca)  # (B, S, D)

        # 4. Combine: temporal + time + static context
        x = temporal_selected + t_embed + enrichment.unsqueeze(1)

        # 5. Pre-attention GRN
        x = self.pre_attn_grn(x)

        # 6. Transformer self-attention
        # Convert mask: True=real -> need to invert for PyTorch's src_key_padding_mask
        # PyTorch expects True=ignore, so we flip
        padding_mask = ~mask  # (B, S), True = pad position to ignore
        x = self.transformer_encoder(x, src_key_padding_mask=padding_mask)

        # 7. Post-attention GRN
        x = self.post_attn_grn(x)

        return x, temporal_weights

    def forward(
        self,
        temporal_features: torch.Tensor,  # (B, S, F_t)
        static_features: torch.Tensor,    # (B, F_s)
        time_to_tca: torch.Tensor,        # (B, S, 1)
        mask: torch.Tensor,               # (B, S) — True for real, False for padding
    ):
        B, S, _ = temporal_features.shape

        # Steps 1-7: encode sequence into per-timestep hidden states
        x, temporal_weights = self.encode_sequence(
            temporal_features, static_features, time_to_tca, mask
        )

        # 8. Attention-weighted pooling over all real positions
        # Instead of just the last CDM, learn which time steps matter most
        attn_scores = self.pool_attention(x).squeeze(-1)  # (B, S)
        # Mask padding positions with -inf so they get zero attention
        attn_scores = attn_scores.masked_fill(~mask, float("-inf"))
        attn_weights = F.softmax(attn_scores, dim=-1)  # (B, S)
        # Handle all-padding edge case (shouldn't happen but be safe)
        attn_weights = attn_weights.nan_to_num(0.0)
        x_pooled = (x * attn_weights.unsqueeze(-1)).sum(dim=1)  # (B, D)

        # 9. Prediction heads
        risk_logit = self.risk_head(x_pooled)    # (B, 1)
        miss_log = self.miss_head(x_pooled)      # (B, 1)
        pc_log10 = self.pc_head(x_pooled)        # (B, 1) — log10(Pc)

        return risk_logit, miss_log, pc_log10, temporal_weights

    def count_parameters(self) -> int:
        return sum(p.numel() for p in self.parameters() if p.requires_grad)


class SigmoidFocalLoss(nn.Module):
    """
    Focal Loss for binary classification (Lin et al., 2017).

    Down-weights well-classified examples so the model focuses on hard cases.
    FL(p_t) = -alpha_t * (1 - p_t)^gamma * log(p_t)

    With gamma=0, this reduces to standard weighted BCE.
    With gamma=2, easy examples (p_t > 0.9) get ~100x less weight.
    """

    def __init__(self, alpha: float = 0.75, gamma: float = 2.0, reduction: str = "mean"):
        super().__init__()
        self.alpha = alpha
        self.gamma = gamma
        self.reduction = reduction

    def forward(self, logits: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
        p = torch.sigmoid(logits)
        # p_t = probability of the true class
        p_t = targets * p + (1 - targets) * (1 - p)
        # alpha_t = alpha for positive class, (1-alpha) for negative
        alpha_t = targets * self.alpha + (1 - targets) * (1 - self.alpha)
        # focal modulator: (1 - p_t)^gamma
        focal_weight = (1 - p_t) ** self.gamma
        # BCE per-element (numerically stable via log-sum-exp)
        bce = F.binary_cross_entropy_with_logits(logits, targets, reduction="none")
        loss = alpha_t * focal_weight * bce
        if self.reduction == "none":
            return loss
        return loss.mean()


class PhysicsInformedLoss(nn.Module):
    """
    Combined task loss + physics regularization.

    Total loss = risk_weight * FocalLoss(risk) + miss_weight * MSE(miss_distance)
                 + pc_weight * MSE(log10_Pc) + physics_weight * ReLU(MOID - predicted_miss)

    The physics term: MOID (Minimum Orbital Intersection Distance) is the
    geometric minimum distance between two orbits. The actual miss distance
    at closest approach CANNOT be less than MOID (without a maneuver).
    If the model predicts miss < MOID, we penalize it.

    The Pc term: direct regression on log10(collision probability). The Kelvins
    `risk` column is log10(Pc), giving us 162K labeled examples. This lets
    the model output calibrated collision probabilities, not just binary risk.

    For the Kelvins dataset, we approximate MOID from the orbital elements
    in the CDM features. When MOID isn't available, the physics term is 0.
    """

    def __init__(
        self,
        risk_weight: float = 1.0,
        miss_weight: float = 0.1,
        pc_weight: float = 0.3,
        physics_weight: float = 0.2,
        pos_weight: float = 50.0,
        use_focal: bool = False,
        focal_alpha: float = 0.75,
        focal_gamma: float = 2.0,
    ):
        super().__init__()
        self.risk_weight = risk_weight
        self.miss_weight = miss_weight
        self.pc_weight = pc_weight
        self.physics_weight = physics_weight
        if use_focal:
            self.risk_loss = SigmoidFocalLoss(alpha=focal_alpha, gamma=focal_gamma)
        else:
            self.risk_loss = nn.BCEWithLogitsLoss(
                pos_weight=torch.tensor(pos_weight)
            )
        self.miss_loss = nn.MSELoss()

    def forward(
        self,
        risk_logit: torch.Tensor,      # (B, 1)
        miss_pred_log: torch.Tensor,    # (B, 1)
        risk_target: torch.Tensor,      # (B,)
        miss_target_log: torch.Tensor,  # (B,)
        pc_pred_log10: torch.Tensor = None,  # (B, 1) predicted log10(Pc)
        pc_target_log10: torch.Tensor = None,  # (B,) target log10(Pc)
        moid_log: torch.Tensor = None,  # (B,) optional, log1p(MOID_km)
        domain_weight: torch.Tensor = None,  # (B,) per-sample weight
    ) -> tuple[torch.Tensor, dict]:

        # Risk classification loss (BCE with class weighting)
        if domain_weight is not None and not isinstance(self.risk_loss, SigmoidFocalLoss):
            # Per-sample weighted BCE: compute element-wise then weight
            bce_per_sample = F.binary_cross_entropy_with_logits(
                risk_logit.squeeze(-1), risk_target,
                pos_weight=self.risk_loss.pos_weight.to(risk_logit.device),
                reduction="none",
            )
            L_risk = (bce_per_sample * domain_weight).mean()
        else:
            L_risk = self.risk_loss(risk_logit.squeeze(-1), risk_target)

        # Miss distance regression loss — also domain-weighted
        miss_residual = (miss_pred_log.squeeze(-1) - miss_target_log) ** 2
        if domain_weight is not None:
            L_miss = (miss_residual * domain_weight).mean()
        else:
            L_miss = miss_residual.mean()

        # Collision probability regression loss
        L_pc = torch.tensor(0.0, device=risk_logit.device)
        if pc_pred_log10 is not None and pc_target_log10 is not None:
            pc_residual = (pc_pred_log10.squeeze(-1) - pc_target_log10) ** 2
            if domain_weight is not None:
                L_pc = (pc_residual * domain_weight).mean()
            else:
                L_pc = pc_residual.mean()

        # Physics constraint: predicted miss >= MOID
        L_physics = torch.tensor(0.0, device=risk_logit.device)
        if moid_log is not None:
            # Violation = how much below MOID the prediction is
            violation = F.relu(moid_log - miss_pred_log.squeeze(-1))
            L_physics = violation.mean()

        total = (self.risk_weight * L_risk
                 + self.miss_weight * L_miss
                 + self.pc_weight * L_pc
                 + self.physics_weight * L_physics)

        metrics = {
            "loss": total.item(),
            "risk_loss": L_risk.item(),
            "miss_loss": L_miss.item(),
            "pc_loss": L_pc.item(),
            "physics_loss": L_physics.item(),
        }

        return total, metrics