JiaruiLi/DeepHealth
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Refactor bias handling in AUC evaluation scripts to improve robustness and prevent errors when bias is not defined

Browse Source
This commit is contained in:
Jiarui Li
2026-06-12 17:28:22 +08:00
parent 1240766aa9
commit 034d8065a7
2 changed files with 14 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
evaluate_auc.py
View File

@@ -632,15 +632,19 @@ def compute_logits_for_disease_chunk(
device.type == "cuda" and use_amp) else torch.float32 device.type == "cuda" and use_amp) else torch.float32
weight = model.risk_head.weight[disease_ids].detach().to( weight = model.risk_head.weight[disease_ids].detach().to(
device=device, dtype=compute_dtype) device=device, dtype=compute_dtype)
bias = model.risk_head.bias[disease_ids].detach().to( bias = None
device=device, dtype=compute_dtype) if model.risk_head.bias is not None:
bias = model.risk_head.bias[disease_ids].detach().to(
device=device, dtype=compute_dtype)
parts: List[np.ndarray] = [] parts: List[np.ndarray] = []
for start in tqdm(range(0, n, logit_batch_size), desc="Risk-head projection", leave=False, dynamic_ncols=True): for start in tqdm(range(0, n, logit_batch_size), desc="Risk-head projection", leave=False, dynamic_ncols=True):
end = min(start + logit_batch_size, n) end = min(start + logit_batch_size, n)
h = torch.from_numpy(hidden_all[start:end]).to( h = torch.from_numpy(hidden_all[start:end]).to(
device=device, dtype=compute_dtype, non_blocking=True) device=device, dtype=compute_dtype, non_blocking=True)
logits = torch.matmul(h, weight.t()) + bias logits = torch.matmul(h, weight.t())
if bias is not None:
logits = logits + bias
parts.append(logits.float().cpu().numpy().astype( parts.append(logits.float().cpu().numpy().astype(
np.float32, copy=False)) np.float32, copy=False))
del h, logits del h, logits

10
evaluate_auc_v2.py
View File

@@ -839,8 +839,10 @@ def project_distribution_chunk(
device.type == "cuda" and use_amp) else torch.float32 device.type == "cuda" and use_amp) else torch.float32
weight = model.risk_head.weight[disease_ids].detach().to( weight = model.risk_head.weight[disease_ids].detach().to(
device=device, dtype=compute_dtype) device=device, dtype=compute_dtype)
bias = model.risk_head.bias[disease_ids].detach().to( bias = None
device=device, dtype=compute_dtype) if model.risk_head.bias is not None:
bias = model.risk_head.bias[disease_ids].detach().to(
device=device, dtype=compute_dtype)
rho_weight = None rho_weight = None
rho_bias = None rho_bias = None
death_rho_weight = None death_rho_weight = None
@@ -868,7 +870,9 @@ def project_distribution_chunk(
end = min(start + logit_batch_size, n) end = min(start + logit_batch_size, n)
h = torch.from_numpy(hidden_all[start:end]).to( h = torch.from_numpy(hidden_all[start:end]).to(
device=device, dtype=compute_dtype, non_blocking=True) device=device, dtype=compute_dtype, non_blocking=True)
logits = torch.matmul(h, weight.t()) + bias logits = torch.matmul(h, weight.t())
if bias is not None:
logits = logits + bias
rho = None rho = None
if dist_mode == "weibull": if dist_mode == "weibull":
assert rho_weight is not None and rho_bias is not None assert rho_weight is not None and rho_bias is not None