Refactor LogNormalBasisHazardLoss to LogNormalBasisBinnedHazardCIFNLLLoss and update related configurations

This commit is contained in:
2026-01-13 21:11:38 +08:00
parent 1df02d85d7
commit f16596ed58
3 changed files with 320 additions and 154 deletions

264
losses.py
View File

@@ -260,24 +260,27 @@ class DiscreteTimeCIFNLLLoss(nn.Module):
return nll, reg
class LogNormalBasisHazardLoss(nn.Module):
"""Event-only competing risks loss using lognormal basis (Gaussian on log-time).
class LogNormalBasisBinnedHazardCIFNLLLoss(nn.Module):
r"""Route-3: continuous-time lognormal-basis hazards with discrete-time CIF likelihood.
This loss models cause-specific CIF as a mixture of lognormal basis CDFs:
This implements a cause-specific continuous-time hazard model:
F_j(t) = sum_r w_{j,r} * Phi((log t - mu_r) / sigma)
\lambda_j(t) = \sum_r \alpha_{j,r} b_r(t)
Training uses *bin probability mass* (Delta CIF / interval mass). There is
**no censoring**: every sample is an observed event with a valid cause.
where b_r(t) is the lognormal PDF basis implied by a Normal on log-time.
Logits interface:
logits: (B, 1 + J*R)
logits[:, 0] -> w0 (survival mass / never-event)
logits[:, 1:] -> flattened (j,r) in row-major order: j then r
index = 1 + j*R + r
Training objective is IDENTICAL in structure to DiscreteTimeCIFNLLLoss,
but per-bin categorical probabilities are derived from integrated hazards.
Expected logits interface (preferred):
logits: (B, J*R)
reshaped to (B, J, R)
For convenience/compatibility, also accepts:
logits: (B, 1+J*R) and ignores the first column.
Forward interface (must match):
forward(logits, target_events, dt, reduction)
forward(logits, target_events, dt, reduction) -> (nll, reg)
"""
def __init__(
@@ -286,17 +289,20 @@ class LogNormalBasisHazardLoss(nn.Module):
centers: Sequence[float],
*,
eps: float = 1e-8,
bandwidth_init: float = 0.5,
alpha_floor: float = 0.0,
bandwidth_init: float = 0.7,
bandwidth_min: float = 1e-3,
bandwidth_max: float = 10.0,
lambda_sigma_reg: float = 0.0,
sigma_reg_target: Optional[float] = None,
return_dict: bool = False,
lambda_reg: float = 0.0,
):
super().__init__()
if len(bin_edges) < 2:
raise ValueError("bin_edges must have length >= 2")
raise ValueError("bin_edges must have length >= 2 (n_bins >= 1)")
if float(bin_edges[0]) != 0.0:
raise ValueError("bin_edges[0] must equal 0")
# allow last edge to be +inf
for i in range(1, len(bin_edges)):
prev = float(bin_edges[i - 1])
@@ -310,20 +316,19 @@ class LogNormalBasisHazardLoss(nn.Module):
else:
if not (cur > prev):
raise ValueError("bin_edges must be strictly increasing")
if float(bin_edges[0]) < 0.0:
raise ValueError("bin_edges[0] must be >= 0")
if len(centers) < 1:
raise ValueError("centers must have length >= 1")
self.eps = float(eps)
self.alpha_floor = float(alpha_floor)
self.bandwidth_min = float(bandwidth_min)
self.bandwidth_max = float(bandwidth_max)
self.bandwidth_init = float(bandwidth_init)
self.lambda_sigma_reg = float(lambda_sigma_reg)
self.sigma_reg_target = None if sigma_reg_target is None else float(
sigma_reg_target)
self.bandwidth_init = float(bandwidth_init)
self.return_dict = bool(return_dict)
self.lambda_reg = float(lambda_reg)
self.register_buffer(
"bin_edges",
@@ -343,69 +348,32 @@ class LogNormalBasisHazardLoss(nn.Module):
@staticmethod
def _normal_cdf(z: torch.Tensor) -> torch.Tensor:
# Stable normal CDF via erf.
z = torch.clamp(z, -12.0, 12.0)
return 0.5 * (1.0 + torch.erf(z / math.sqrt(2.0)))
@staticmethod
def _normal_sf(z: torch.Tensor) -> torch.Tensor:
# Stable normal survival function via erfc.
z = torch.clamp(z, -12.0, 12.0)
return 0.5 * torch.erfc(z / math.sqrt(2.0))
def forward(
def _compute_delta_basis_all_bins(
self,
logits: torch.Tensor,
target_events: torch.Tensor,
dt: torch.Tensor,
reduction: str = "mean",
) -> Union[Tuple[torch.Tensor, torch.Tensor], Dict[str, Any]]:
if logits.ndim != 2:
raise ValueError(
f"logits must be 2D with shape (B, 1+J*R); got {tuple(logits.shape)}")
if target_events.ndim != 1 or dt.ndim != 1:
raise ValueError("target_events and dt must be 1D tensors")
if logits.shape[0] != target_events.shape[0] or logits.shape[0] != dt.shape[0]:
raise ValueError(
"Batch size mismatch among logits, target_events, dt")
if reduction not in {"mean", "sum", "none"}:
raise ValueError("reduction must be one of {'mean','sum','none'}")
device = logits.device
dtype = logits.dtype
*,
device: torch.device,
dtype: torch.dtype,
) -> torch.Tensor:
"""Compute ΔB[k,r] for bins k=1..n_bins (shape: (n_bins, R))."""
bin_edges = self.bin_edges.to(device=device, dtype=dtype)
centers = self.centers.to(device=device, dtype=dtype)
bsz = logits.shape[0]
r = int(centers.numel())
jr = int(logits.shape[1] - 1)
if jr <= 0:
raise ValueError(
"logits.shape[1] must be >= 2 (w0 + at least one (j,r) weight)")
if jr % r != 0:
raise ValueError(
f"(logits.shape[1]-1) must be divisible by R={r}; got {jr}")
j = jr // r
# 1) Stable global weights (includes w0).
w_all = torch.softmax(logits, dim=-1) # (B, 1+J*R)
w0 = w_all[:, 0]
w = w_all[:, 1:].view(bsz, j, r)
# 2) Determine event bin index.
k = int(bin_edges.numel() - 1)
if k < 1:
n_bins = int(bin_edges.numel() - 1)
if n_bins < 1:
raise ValueError("bin_edges must define at least one bin")
# v2: dt is always continuous time (float), map to bin via searchsorted.
dt_f = dt.to(device=device, dtype=dtype)
bin_idx = torch.searchsorted(bin_edges, dt_f, right=True) - 1
bin_idx = torch.clamp(bin_idx, 0, k - 1).to(torch.long)
left = bin_edges[:-1] # (n_bins,)
right = bin_edges[1:] # (n_bins,)
left = bin_edges[bin_idx]
right = bin_edges[bin_idx + 1]
# 3) Stable log(t) clamp.
if float(self.bin_edges[1]) > 0.0:
t_min = float(self.bin_edges[1]) * 1e-6
else:
@@ -413,19 +381,19 @@ class LogNormalBasisHazardLoss(nn.Module):
t_min_t = torch.tensor(t_min, device=device, dtype=dtype)
left_is_zero = left <= 0
# For log() we still need a positive clamp, but we will treat CDF(left)=0 exactly
# when left<=0 (instead of approximating via t_min).
left_clamped = torch.clamp(left, min=t_min_t)
log_left = torch.log(left_clamped)
is_inf = torch.isinf(right)
# right might be +inf for last bin; avoid log(+inf) by substituting a safe finite value.
right_safe = torch.where(is_inf, left_clamped,
torch.clamp(right, min=t_min_t))
right_safe = torch.where(
is_inf, left_clamped, torch.clamp(right, min=t_min_t))
log_right = torch.log(right_safe)
sigma = torch.clamp(self.log_sigma.to(
device=device, dtype=dtype).exp(), self.bandwidth_min, self.bandwidth_max)
sigma = torch.clamp(
self.log_sigma.to(device=device, dtype=dtype).exp(),
self.bandwidth_min,
self.bandwidth_max,
)
z_left = (log_left.unsqueeze(-1) - centers.unsqueeze(0)) / sigma
z_right = (log_right.unsqueeze(-1) - centers.unsqueeze(0)) / sigma
@@ -433,58 +401,152 @@ class LogNormalBasisHazardLoss(nn.Module):
z_right = torch.clamp(z_right, -12.0, 12.0)
cdf_left = self._normal_cdf(z_left)
# Treat the first-bin left boundary exactly as 0 in CDF.
if left_is_zero.any():
cdf_left = torch.where(
left_is_zero.unsqueeze(-1), torch.zeros_like(cdf_left), cdf_left)
cdf_right = self._normal_cdf(z_right)
delta_finite = cdf_right - cdf_left
# Last bin: ΔB = 1 - CDF(left) = SF(left), computed via erfc for stability.
delta_last = self._normal_sf(z_left)
# If left<=0, SF(left)=1 exactly.
if left_is_zero.any():
delta_last = torch.where(
left_is_zero.unsqueeze(-1), torch.ones_like(delta_last), delta_last)
delta_basis = torch.where(
is_inf.unsqueeze(-1), delta_last, delta_finite)
delta_basis = torch.clamp(delta_basis, min=0.0)
return delta_basis
def forward(
self,
logits: torch.Tensor,
target_events: torch.Tensor,
dt: torch.Tensor,
reduction: str = "mean",
) -> Tuple[torch.Tensor, torch.Tensor]:
if logits.ndim not in {2, 3}:
raise ValueError(
f"logits must be 2D (B, J*R) (or (B, 1+J*R)) or 3D (B, J, R); got {tuple(logits.shape)}"
)
if target_events.ndim != 1 or dt.ndim != 1:
raise ValueError(
f"target_events and dt must be 1D tensors; got target_events.ndim={target_events.ndim}, dt.ndim={dt.ndim}"
)
if logits.shape[0] != target_events.shape[0] or logits.shape[0] != dt.shape[0]:
raise ValueError(
"Batch size mismatch: logits.shape[0] must equal target_events.shape[0] and dt.shape[0]"
)
if reduction not in {"mean", "sum", "none"}:
raise ValueError("reduction must be one of {'mean','sum','none'}")
if not torch.all(dt > 0):
raise ValueError("dt must be strictly positive")
device = logits.device
dtype = logits.dtype
centers = self.centers.to(device=device, dtype=dtype)
r = int(centers.numel())
if r < 1:
raise ValueError("centers must have length >= 1")
if logits.ndim == 3:
if logits.shape[2] != r:
raise ValueError(
f"logits.shape[2] must equal R={r}; got {int(logits.shape[2])}"
)
j = int(logits.shape[1])
if j < 1:
raise ValueError("Inferred number of causes J must be >= 1")
alpha = F.softplus(logits) + self.alpha_floor # (B, J, R)
else:
d = int(logits.shape[1])
offset = 0
if d % r == 0:
jr = d
elif (d - 1) % r == 0:
offset = 1
jr = d - 1
else:
raise ValueError(
f"logits.shape[1] must be divisible by R={r} (or 1+J*R); got {d}"
)
j = jr // r
if j < 1:
raise ValueError("Inferred number of causes J must be >= 1")
logits_used = logits[:, offset:]
alpha = F.softplus(logits_used).view(-1, j, r) + \
self.alpha_floor # (B, J, R)
delta_basis = self._compute_delta_basis_all_bins(
device=device,
dtype=dtype,
) # (n_bins, R)
n_bins = int(delta_basis.shape[0])
# H_{j,k} = sum_r alpha_{j,r} * ΔB_{k,r}
h_jk = torch.einsum("mjr,kr->mjk", alpha,
delta_basis) # (B, J, n_bins)
h_k = h_jk.sum(dim=1) # (B, n_bins)
# Map continuous dt to discrete event bin index k* in {1..n_bins}.
bin_edges = self.bin_edges.to(device=dt.device, dtype=dt.dtype)
time_bin = torch.bucketize(dt, bin_edges)
time_bin = torch.clamp(time_bin, min=1, max=n_bins).to(torch.long)
# 4) Gather per-sample cause weights and compute event mass.
cause = target_events.to(device=device, dtype=torch.long)
if (cause < 0).any() or (cause >= j).any():
raise ValueError(f"target_events must be in [0, J-1] where J={j}")
b_idx = torch.arange(bsz, device=device)
w_cause = w[b_idx, cause, :] # (B, R)
# Previous survival term: sum_{u<k*} H_u
bins = torch.arange(
1, n_bins + 1, device=device).unsqueeze(0) # (1, n_bins)
mask_prev = bins < time_bin.unsqueeze(1) # (B, n_bins)
loss_prev = (h_k * mask_prev.to(h_k.dtype)).sum(dim=1) # (B,)
m = (w_cause * delta_basis).sum(dim=-1)
m = torch.clamp(m, min=self.eps)
nll_vec = -torch.log(m)
# Event term at k*: -log p_{k*}(cause)
b_idx = torch.arange(target_events.shape[0], device=device)
k0 = time_bin - 1 # (B,) index into 0..n_bins-1
h_event_total = h_k[b_idx, k0]
h_event_total = torch.clamp(h_event_total, min=self.eps)
h_event_cause = h_jk[b_idx, cause, k0]
h_event_cause = torch.clamp(h_event_cause, min=self.eps)
# log(1 - exp(-H)) stably
log1mexp = torch.log(-torch.expm1(-h_event_total))
loss_event = -log1mexp - \
torch.log(h_event_cause) + torch.log(h_event_total)
loss = loss_prev + loss_event
if reduction == "mean":
nll: torch.Tensor = nll_vec.mean()
nll = loss.mean()
elif reduction == "sum":
nll = nll_vec.sum()
nll = loss.sum()
else:
nll = nll_vec
nll = loss
reg = torch.zeros((), device=device, dtype=dtype)
if self.lambda_reg > 0.0:
# Regularize the within-bin cause competition via NLL on log ratios.
# ratio_j = H_{j,k*} / H_{k*}
h_event_all = h_jk[b_idx, :, k0] # (B, J)
denom = torch.clamp(h_event_total, min=self.eps).unsqueeze(1)
ratio = torch.clamp(h_event_all / denom, min=self.eps)
log_ratio = torch.log(ratio)
reg = reg + self.lambda_reg * F.nll_loss(
log_ratio, cause, reduction="mean")
sigma_penalty = torch.zeros((), device=device, dtype=dtype)
if self.lambda_sigma_reg > 0.0:
target = self.bandwidth_init if self.sigma_reg_target is None else self.sigma_reg_target
sigma_penalty = (self.log_sigma.to(
device=device, dtype=dtype) - math.log(float(target))) ** 2
reg = sigma_penalty * float(self.lambda_sigma_reg)
reg = reg + sigma_penalty * self.lambda_sigma_reg
if not self.return_dict:
return nll, reg
return {
"nll": nll,
"reg": reg,
"nll_vec": nll_vec,
"sigma": sigma.detach(),
"avg_w0": w0.mean().detach(),
"min_delta_basis": delta_basis.min().detach(),
"mean_m": m.mean().detach(),
"sigma_penalty": sigma_penalty.detach(),
"bin_idx": bin_idx.detach(),
}
return nll, reg