Use precomputed exposure embeddings

This commit is contained in:
2026-07-09 16:49:49 +08:00
parent 552e09aa01
commit e439c3c98c
7 changed files with 311 additions and 281 deletions

153
models.py
View File

@@ -7,7 +7,6 @@ import torch.nn.functional as F
from backbones import (
AgeSinusoidalEncoding,
GPTBlock,
TimesNetExposureEncoder,
)
from targets import PAD_IDX
@@ -30,16 +29,7 @@ class DeepHealth(nn.Module):
target_mode: str = "next_token", # "next_token" or "all_future"
dist_mode: str = "exponential", # "exponential", "weibull" or "mixed"
dropout: float = 0.0,
use_exposure_encoder: bool = False,
exposure_daily_input_dim: int = 4,
exposure_monthly_input_dim: int = 2,
exposure_d_model: int | None = 64,
exposure_n_layers: int = 2,
exposure_top_k: int = 2,
exposure_n_backbone_blocks: int = 1,
exposure_backbone_kernel_size: int = 5,
exposure_backbone_expansion: float = 2.0,
exposure_use_gate: bool = True,
use_exposure_embeddings: bool = False,
):
super().__init__()
if target_mode not in ["next_token", "all_future"]:
@@ -53,26 +43,9 @@ class DeepHealth(nn.Module):
2, n_embd) # Assuming binary gender
self.target_mode = target_mode
self.dist_mode = dist_mode
self.use_exposure_encoder = use_exposure_encoder
self.use_exposure_embeddings = bool(use_exposure_embeddings)
self.n_embd = n_embd
self.vocab_size = vocab_size
self.exposure_encoder = (
TimesNetExposureEncoder(
n_embd=n_embd,
daily_input_dim=exposure_daily_input_dim,
monthly_input_dim=exposure_monthly_input_dim,
d_model=exposure_d_model,
n_layers=exposure_n_layers,
top_k=exposure_top_k,
n_backbone_blocks=exposure_n_backbone_blocks,
backbone_kernel_size=exposure_backbone_kernel_size,
backbone_expansion=exposure_backbone_expansion,
dropout=dropout,
use_gate=exposure_use_gate,
)
if use_exposure_encoder
else None
)
nn.init.normal_(self.token_embedding.weight, mean=0.0, std=0.02)
nn.init.zeros_(self.token_embedding.weight[0])
nn.init.normal_(self.gender_embedding.weight, mean=0.0, std=0.02)
@@ -121,95 +94,6 @@ class DeepHealth(nn.Module):
dtype=dtype,
).masked_fill(~valid, -1e4)[:, None, :, :]
def _encode_event_exposure(
self,
exposure_daily: torch.Tensor | None,
exposure_monthly: torch.Tensor | None,
exposure_daily_mask: torch.Tensor | None,
exposure_monthly_mask: torch.Tensor | None,
event_shape: tuple[int, int],
) -> torch.Tensor | None:
if self.exposure_encoder is None:
if exposure_daily is not None or exposure_monthly is not None:
raise ValueError(
"Exposure tensors were provided but use_exposure_encoder=False"
)
return None
if exposure_daily is None or exposure_monthly is None:
raise ValueError(
"exposure_daily and exposure_monthly are required when "
"use_exposure_encoder=True"
)
batch_size, event_len = event_shape
if exposure_daily.shape[:2] != event_shape:
raise ValueError(
"exposure_daily must have shape (B, L, T_daily, C_daily), got "
f"{tuple(exposure_daily.shape)} for event shape {event_shape}"
)
if exposure_monthly.shape[:2] != event_shape:
raise ValueError(
"exposure_monthly must have shape (B, L, T_monthly, C_monthly), got "
f"{tuple(exposure_monthly.shape)} for event shape {event_shape}"
)
if exposure_daily.dim() != 4 or exposure_monthly.dim() != 4:
raise ValueError(
"exposure_daily and exposure_monthly must both be 4D tensors"
)
def flatten_mask(mask: torch.Tensor | None, ref: torch.Tensor) -> torch.Tensor | None:
if mask is None:
return None
if mask.shape[:2] != event_shape:
raise ValueError(
"exposure mask must start with shape (B, L), got "
f"{tuple(mask.shape)}"
)
if mask.dim() not in {3, 4}:
raise ValueError(
"exposure mask must have shape (B, L, T) or (B, L, T, C), "
f"got {tuple(mask.shape)}"
)
if mask.shape[2] != ref.shape[2]:
raise ValueError(
"exposure mask time dimension does not match exposure tensor"
)
if mask.dim() == 4 and mask.shape[3] != ref.shape[3]:
raise ValueError(
"exposure mask channel dimension does not match exposure tensor"
)
return mask.reshape(batch_size * event_len, *mask.shape[2:])
daily = exposure_daily.reshape(
batch_size * event_len,
exposure_daily.size(2),
exposure_daily.size(3),
)
monthly = exposure_monthly.reshape(
batch_size * event_len,
exposure_monthly.size(2),
exposure_monthly.size(3),
)
daily_mask = flatten_mask(exposure_daily_mask, exposure_daily)
monthly_mask = flatten_mask(exposure_monthly_mask, exposure_monthly)
exposure_device = exposure_daily.device
exposure_dtype = self.token_embedding.weight.dtype
daily = daily.to(device=exposure_device, dtype=exposure_dtype)
monthly = monthly.to(device=exposure_device, dtype=exposure_dtype)
if daily_mask is not None:
daily_mask = daily_mask.to(device=exposure_device)
if monthly_mask is not None:
monthly_mask = monthly_mask.to(device=exposure_device)
exposure_emb = self.exposure_encoder(
daily=daily,
monthly=monthly,
daily_mask=daily_mask,
monthly_mask=monthly_mask,
)
return exposure_emb.reshape(batch_size, event_len, self.n_embd)
def _forward_shared(
self,
event_seq: torch.LongTensor,
@@ -218,10 +102,7 @@ class DeepHealth(nn.Module):
mode: str,
padding_mask: torch.Tensor | None = None,
t_query: torch.FloatTensor | None = None,
exposure_daily: torch.Tensor | None = None,
exposure_monthly: torch.Tensor | None = None,
exposure_daily_mask: torch.Tensor | None = None,
exposure_monthly_mask: torch.Tensor | None = None,
exposure_embedding: torch.Tensor | None = None,
return_output: bool = False,
**unused_kwargs,
) -> torch.Tensor | DeepHealthOutput:
@@ -240,16 +121,24 @@ class DeepHealth(nn.Module):
event_len = event_seq.size(1)
h_disease = self.token_embedding(event_seq)
h_exposure = self._encode_event_exposure(
exposure_daily=exposure_daily,
exposure_monthly=exposure_monthly,
exposure_daily_mask=exposure_daily_mask,
exposure_monthly_mask=exposure_monthly_mask,
event_shape=(event_seq.size(0), event_len),
)
if h_exposure is not None:
h_exposure = h_exposure.to(device=event_seq.device, dtype=h_disease.dtype)
h_disease = h_disease + h_exposure
if self.use_exposure_embeddings:
if exposure_embedding is None:
raise ValueError(
"exposure_embedding is required when "
"use_exposure_embeddings=True"
)
if exposure_embedding.shape != h_disease.shape:
raise ValueError(
"exposure_embedding must have shape "
f"{tuple(h_disease.shape)}, got {tuple(exposure_embedding.shape)}"
)
h_disease = h_disease + exposure_embedding.to(
device=h_disease.device, dtype=h_disease.dtype
)
elif exposure_embedding is not None:
raise ValueError(
"exposure_embedding provided but use_exposure_embeddings=False"
)
t_disease = time_seq
h_disease = h_disease * padding_mask.unsqueeze(-1).to(h_disease.dtype)