| from __future__ import annotations |
|
|
| import torch |
| from torch import nn |
|
|
| from .blocks import MLP |
|
|
|
|
| class ClinicalEncoder(nn.Module): |
| """Clinical MLP that explicitly receives missingness indicators.""" |
|
|
| def __init__(self, clinical_dim: int, feature_dim: int = 256, use_missingness: bool = True) -> None: |
| super().__init__() |
| self.clinical_dim = clinical_dim |
| self.feature_dim = feature_dim |
| self.use_missingness = use_missingness |
| self.net = MLP(clinical_dim * 2, max(feature_dim // 2, 16), feature_dim) |
|
|
| def forward(self, values: torch.Tensor | None, missing: torch.Tensor | None) -> torch.Tensor: |
| if self.clinical_dim == 0: |
| batch = 1 if values is None else values.shape[0] |
| device = None if values is None else values.device |
| return torch.zeros(batch, self.feature_dim, device=device) |
| if values is None or missing is None: |
| raise ValueError("Clinical values and missingness mask are required when clinical_dim > 0") |
| if not self.use_missingness: |
| missing = torch.zeros_like(missing) |
| return self.net(torch.cat([values.float(), missing.float()], dim=-1)) |
|
|
