feat: rename logit_bias/logit_scale to logit_mean/logit_sigma for affine score calibration (#39530)

Signed-off-by: Jesus Federico <jefp@amazon.com>
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: wang.yuqi <yuqi.wang@daocloud.io>
This commit is contained in:
Jesus Federico
2026-04-13 04:43:44 +00:00
committed by GitHub
co-authored by Claude Opus 4.6 wang.yuqi
parent 2a3c32ce67
commit fa6ae31177
7 changed files with 86 additions and 35 deletions
@@ -103,16 +103,16 @@ class ClassifierPoolerHead(SequencePoolerHead):
def __init__(
self,
classifier: ClassifierFn | None = None,
logit_bias: float | None = None,
logit_scale: float | None = None,
logit_mean: float | None = None,
logit_sigma: float | None = None,
head_dtype: torch.dtype | str | None = None,
activation: ActivationFn | None = None,
) -> None:
super().__init__()
self.classifier = classifier
self.logit_bias = logit_bias
self.logit_scale = logit_scale
self.logit_mean = logit_mean
self.logit_sigma = logit_sigma
self.head_dtype = head_dtype
self.activation = activation
@@ -140,10 +140,11 @@ class ClassifierPoolerHead(SequencePoolerHead):
logits = pooled_data
# logits shape: [batchsize, num_labels]
if self.logit_bias is not None:
logits -= self.logit_bias
if self.logit_scale is not None:
logits *= self.logit_scale
# Affine score calibration: activation((logit - mean) / sigma)
if self.logit_mean is not None:
logits = logits - self.logit_mean
if self.logit_sigma is not None:
logits = logits / self.logit_sigma
if self.activation is not None:
flags = [p.use_activation for p in pooling_params]
@@ -118,8 +118,8 @@ def pooler_for_classify(
head = ClassifierPoolerHead(
head_dtype=model_config.head_dtype,
classifier=classifier,
logit_bias=model_config.pooler_config.logit_bias,
logit_scale=model_config.pooler_config.logit_scale,
logit_mean=model_config.pooler_config.logit_mean,
logit_sigma=model_config.pooler_config.logit_sigma,
activation=resolve_classifier_act_fn(
model_config, static_num_labels=True, act_fn=act_fn
),
@@ -92,16 +92,16 @@ class TokenClassifierPoolerHead(TokenPoolerHead):
def __init__(
self,
classifier: ClassifierFn | None = None,
logit_bias: float | None = None,
logit_scale: float | None = None,
logit_mean: float | None = None,
logit_sigma: float | None = None,
head_dtype: torch.dtype | str | None = None,
activation: ActivationFn | None = None,
) -> None:
super().__init__()
self.classifier = classifier
self.logit_bias = logit_bias
self.logit_scale = logit_scale
self.logit_mean = logit_mean
self.logit_sigma = logit_sigma
self.head_dtype = head_dtype
self.activation = activation
@@ -127,10 +127,11 @@ class TokenClassifierPoolerHead(TokenPoolerHead):
logits = pooled_data
# logits shape: [n_token, num_labels]
if self.logit_bias is not None:
logits -= self.logit_bias
if self.logit_scale is not None:
logits *= self.logit_scale
# Affine score calibration: activation((logit - mean) / sigma)
if self.logit_mean is not None:
logits = logits - self.logit_mean
if self.logit_sigma is not None:
logits = logits / self.logit_sigma
if self.activation is not None and pooling_param.use_activation:
logits = self.activation(logits)
@@ -127,8 +127,8 @@ def pooler_for_token_classify(
head = TokenClassifierPoolerHead(
head_dtype=model_config.head_dtype,
classifier=classifier,
logit_bias=model_config.pooler_config.logit_bias,
logit_scale=model_config.pooler_config.logit_scale,
logit_mean=model_config.pooler_config.logit_mean,
logit_sigma=model_config.pooler_config.logit_sigma,
activation=resolve_classifier_act_fn(
model_config, static_num_labels=False, act_fn=act_fn
),