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:
co-authored by
Claude Opus 4.6
wang.yuqi
parent
2a3c32ce67
commit
fa6ae31177
@@ -273,27 +273,25 @@ Affine Score Calibration, also known as [Platt Scaling](https://en.wikipedia.org
|
|||||||
|
|
||||||
The calibration follows the transformation:
|
The calibration follows the transformation:
|
||||||
|
|
||||||
`activation(logit_scale * (logit - logit_bias))`
|
`activation((logit - logit_mean) / logit_sigma)`
|
||||||
|
|
||||||
| Parameter | Default | Description |
|
| Parameter | Default | Description |
|
||||||
| --------- | ------- | ----------- |
|
| --------- | ------- | ----------- |
|
||||||
| `logit_bias` | `None` | Bias subtracted from logits before activation |
|
| `logit_mean` | `None` | Mean subtracted from logits (centers scores) |
|
||||||
| `logit_scale` | `None` | Scale factor applied to logits after bias subtraction |
|
| `logit_sigma` | `None` | Standard deviation used to scale logits after mean subtraction |
|
||||||
|
|
||||||
Note: `logit_bias` is **subtracted** from the logits (not added), consistent with the `sigmoid_normalize` convention where `sigmoid(x - bias)` centers the sigmoid around the bias value.
|
|
||||||
|
|
||||||
The computation order is as follows:
|
The computation order is as follows:
|
||||||
|
|
||||||
```python
|
```python
|
||||||
logits -= logit_bias # subtract bias (center scores)
|
logits -= logit_mean # subtract mean (center scores)
|
||||||
logits *= logit_scale # scale logits
|
logits /= logit_sigma # divide by sigma (scale)
|
||||||
logits = activation(logits) # e.g. sigmoid
|
logits = activation(logits) # e.g. sigmoid
|
||||||
```
|
```
|
||||||
|
|
||||||
Example configuration:
|
Example configuration:
|
||||||
|
|
||||||
```bash
|
```bash
|
||||||
--pooler-config '{"use_activation": true, "logit_bias": 4.5, "logit_scale": 1.0}'
|
--pooler-config '{"use_activation": true, "logit_mean": 4.5, "logit_sigma": 1.0}'
|
||||||
```
|
```
|
||||||
|
|
||||||
## Removed Features
|
## Removed Features
|
||||||
@@ -301,3 +299,7 @@ Example configuration:
|
|||||||
### Remove softmax from PoolingParams
|
### Remove softmax from PoolingParams
|
||||||
|
|
||||||
We have already removed `softmax` and `activation` from PoolingParams. Instead, use `use_activation`, since we allow `classify` and `token_classify` to use any activation function.
|
We have already removed `softmax` and `activation` from PoolingParams. Instead, use `use_activation`, since we allow `classify` and `token_classify` to use any activation function.
|
||||||
|
|
||||||
|
### Remove `logit_bias` and `logit_scale`
|
||||||
|
|
||||||
|
`logit_bias` and `logit_scale` are deprecated aliases for `logit_mean` and `logit_sigma` respectively. When using `logit_scale`, it is automatically converted to `logit_sigma = 1/logit_scale`. These deprecated parameters will be removed in v0.21.
|
||||||
|
|||||||
+52
-5
@@ -77,17 +77,31 @@ class PoolerConfig:
|
|||||||
Defaults to None (i.e. set to max_model_len).
|
Defaults to None (i.e. set to max_model_len).
|
||||||
"""
|
"""
|
||||||
|
|
||||||
## for classification models
|
## for classification models — affine score calibration
|
||||||
|
logit_mean: float | None = None
|
||||||
|
"""
|
||||||
|
If provided, subtract this value from classification logits before
|
||||||
|
activation. Used for affine score calibration (Platt scaling):
|
||||||
|
activation((logit - logit_mean) / logit_sigma). Defaults to None.
|
||||||
|
"""
|
||||||
|
|
||||||
|
logit_sigma: float | None = None
|
||||||
|
"""
|
||||||
|
If provided, divide the classification logits by this value after
|
||||||
|
mean subtraction. Used for affine score calibration (Platt scaling):
|
||||||
|
activation((logit - logit_mean) / logit_sigma). Defaults to None.
|
||||||
|
"""
|
||||||
|
|
||||||
|
# Deprecated aliases — will be removed in v0.21
|
||||||
logit_bias: float | None = None
|
logit_bias: float | None = None
|
||||||
"""
|
"""
|
||||||
If provided, apply classification logit biases. Defaults to None.
|
Deprecated: Use logit_mean instead. Will be removed in v0.21.
|
||||||
"""
|
"""
|
||||||
|
|
||||||
logit_scale: float | None = None
|
logit_scale: float | None = None
|
||||||
"""
|
"""
|
||||||
If provided, scale the classification logits by this factor before
|
Deprecated: Use logit_sigma instead (note: logit_sigma = 1/logit_scale).
|
||||||
activation. Combined with logit_bias, enables affine score calibration:
|
Will be removed in v0.21.
|
||||||
activation(logit_scale * (score - logit_bias)). Defaults to None.
|
|
||||||
"""
|
"""
|
||||||
|
|
||||||
## for reward models
|
## for reward models
|
||||||
@@ -105,6 +119,39 @@ class PoolerConfig:
|
|||||||
"""
|
"""
|
||||||
|
|
||||||
def __post_init__(self) -> None:
|
def __post_init__(self) -> None:
|
||||||
|
# Handle deprecated logit_bias → logit_mean
|
||||||
|
if self.logit_bias is not None:
|
||||||
|
if self.logit_mean is not None:
|
||||||
|
raise ValueError(
|
||||||
|
"Cannot set both `logit_bias` and `logit_mean`. "
|
||||||
|
"`logit_bias` is deprecated, use `logit_mean` instead."
|
||||||
|
)
|
||||||
|
logger.warning(
|
||||||
|
"`logit_bias` is deprecated and will be removed in v0.21. "
|
||||||
|
"Use `logit_mean` instead."
|
||||||
|
)
|
||||||
|
self.logit_mean = self.logit_bias
|
||||||
|
self.logit_bias = None
|
||||||
|
|
||||||
|
# Handle deprecated logit_scale → logit_sigma
|
||||||
|
if self.logit_scale is not None:
|
||||||
|
if self.logit_sigma is not None:
|
||||||
|
raise ValueError(
|
||||||
|
"Cannot set both `logit_scale` and `logit_sigma`. "
|
||||||
|
"`logit_scale` is deprecated, use `logit_sigma` instead."
|
||||||
|
)
|
||||||
|
logger.warning(
|
||||||
|
"`logit_scale` is deprecated and will be removed in v0.21. "
|
||||||
|
"Use `logit_sigma` instead (logit_sigma = 1/logit_scale)."
|
||||||
|
)
|
||||||
|
if self.logit_scale == 0:
|
||||||
|
raise ValueError("logit_scale cannot be 0 (division by zero)")
|
||||||
|
self.logit_sigma = 1.0 / self.logit_scale
|
||||||
|
self.logit_scale = None
|
||||||
|
|
||||||
|
if self.logit_sigma is not None and self.logit_sigma == 0:
|
||||||
|
raise ValueError("logit_sigma cannot be 0 (division by zero)")
|
||||||
|
|
||||||
if pooling_type := self.pooling_type:
|
if pooling_type := self.pooling_type:
|
||||||
if self.seq_pooling_type is not None:
|
if self.seq_pooling_type is not None:
|
||||||
raise ValueError(
|
raise ValueError(
|
||||||
|
|||||||
@@ -103,16 +103,16 @@ class ClassifierPoolerHead(SequencePoolerHead):
|
|||||||
def __init__(
|
def __init__(
|
||||||
self,
|
self,
|
||||||
classifier: ClassifierFn | None = None,
|
classifier: ClassifierFn | None = None,
|
||||||
logit_bias: float | None = None,
|
logit_mean: float | None = None,
|
||||||
logit_scale: float | None = None,
|
logit_sigma: float | None = None,
|
||||||
head_dtype: torch.dtype | str | None = None,
|
head_dtype: torch.dtype | str | None = None,
|
||||||
activation: ActivationFn | None = None,
|
activation: ActivationFn | None = None,
|
||||||
) -> None:
|
) -> None:
|
||||||
super().__init__()
|
super().__init__()
|
||||||
|
|
||||||
self.classifier = classifier
|
self.classifier = classifier
|
||||||
self.logit_bias = logit_bias
|
self.logit_mean = logit_mean
|
||||||
self.logit_scale = logit_scale
|
self.logit_sigma = logit_sigma
|
||||||
self.head_dtype = head_dtype
|
self.head_dtype = head_dtype
|
||||||
self.activation = activation
|
self.activation = activation
|
||||||
|
|
||||||
@@ -140,10 +140,11 @@ class ClassifierPoolerHead(SequencePoolerHead):
|
|||||||
logits = pooled_data
|
logits = pooled_data
|
||||||
|
|
||||||
# logits shape: [batchsize, num_labels]
|
# logits shape: [batchsize, num_labels]
|
||||||
if self.logit_bias is not None:
|
# Affine score calibration: activation((logit - mean) / sigma)
|
||||||
logits -= self.logit_bias
|
if self.logit_mean is not None:
|
||||||
if self.logit_scale is not None:
|
logits = logits - self.logit_mean
|
||||||
logits *= self.logit_scale
|
if self.logit_sigma is not None:
|
||||||
|
logits = logits / self.logit_sigma
|
||||||
|
|
||||||
if self.activation is not None:
|
if self.activation is not None:
|
||||||
flags = [p.use_activation for p in pooling_params]
|
flags = [p.use_activation for p in pooling_params]
|
||||||
|
|||||||
@@ -118,8 +118,8 @@ def pooler_for_classify(
|
|||||||
head = ClassifierPoolerHead(
|
head = ClassifierPoolerHead(
|
||||||
head_dtype=model_config.head_dtype,
|
head_dtype=model_config.head_dtype,
|
||||||
classifier=classifier,
|
classifier=classifier,
|
||||||
logit_bias=model_config.pooler_config.logit_bias,
|
logit_mean=model_config.pooler_config.logit_mean,
|
||||||
logit_scale=model_config.pooler_config.logit_scale,
|
logit_sigma=model_config.pooler_config.logit_sigma,
|
||||||
activation=resolve_classifier_act_fn(
|
activation=resolve_classifier_act_fn(
|
||||||
model_config, static_num_labels=True, act_fn=act_fn
|
model_config, static_num_labels=True, act_fn=act_fn
|
||||||
),
|
),
|
||||||
|
|||||||
@@ -92,16 +92,16 @@ class TokenClassifierPoolerHead(TokenPoolerHead):
|
|||||||
def __init__(
|
def __init__(
|
||||||
self,
|
self,
|
||||||
classifier: ClassifierFn | None = None,
|
classifier: ClassifierFn | None = None,
|
||||||
logit_bias: float | None = None,
|
logit_mean: float | None = None,
|
||||||
logit_scale: float | None = None,
|
logit_sigma: float | None = None,
|
||||||
head_dtype: torch.dtype | str | None = None,
|
head_dtype: torch.dtype | str | None = None,
|
||||||
activation: ActivationFn | None = None,
|
activation: ActivationFn | None = None,
|
||||||
) -> None:
|
) -> None:
|
||||||
super().__init__()
|
super().__init__()
|
||||||
|
|
||||||
self.classifier = classifier
|
self.classifier = classifier
|
||||||
self.logit_bias = logit_bias
|
self.logit_mean = logit_mean
|
||||||
self.logit_scale = logit_scale
|
self.logit_sigma = logit_sigma
|
||||||
self.head_dtype = head_dtype
|
self.head_dtype = head_dtype
|
||||||
self.activation = activation
|
self.activation = activation
|
||||||
|
|
||||||
@@ -127,10 +127,11 @@ class TokenClassifierPoolerHead(TokenPoolerHead):
|
|||||||
logits = pooled_data
|
logits = pooled_data
|
||||||
# logits shape: [n_token, num_labels]
|
# logits shape: [n_token, num_labels]
|
||||||
|
|
||||||
if self.logit_bias is not None:
|
# Affine score calibration: activation((logit - mean) / sigma)
|
||||||
logits -= self.logit_bias
|
if self.logit_mean is not None:
|
||||||
if self.logit_scale is not None:
|
logits = logits - self.logit_mean
|
||||||
logits *= self.logit_scale
|
if self.logit_sigma is not None:
|
||||||
|
logits = logits / self.logit_sigma
|
||||||
|
|
||||||
if self.activation is not None and pooling_param.use_activation:
|
if self.activation is not None and pooling_param.use_activation:
|
||||||
logits = self.activation(logits)
|
logits = self.activation(logits)
|
||||||
|
|||||||
@@ -127,8 +127,8 @@ def pooler_for_token_classify(
|
|||||||
head = TokenClassifierPoolerHead(
|
head = TokenClassifierPoolerHead(
|
||||||
head_dtype=model_config.head_dtype,
|
head_dtype=model_config.head_dtype,
|
||||||
classifier=classifier,
|
classifier=classifier,
|
||||||
logit_bias=model_config.pooler_config.logit_bias,
|
logit_mean=model_config.pooler_config.logit_mean,
|
||||||
logit_scale=model_config.pooler_config.logit_scale,
|
logit_sigma=model_config.pooler_config.logit_sigma,
|
||||||
activation=resolve_classifier_act_fn(
|
activation=resolve_classifier_act_fn(
|
||||||
model_config, static_num_labels=False, act_fn=act_fn
|
model_config, static_num_labels=False, act_fn=act_fn
|
||||||
),
|
),
|
||||||
|
|||||||
@@ -232,8 +232,8 @@ class JinaVLForSequenceClassificationConfig(VerifyAndUpdateConfig):
|
|||||||
config = model_config.hf_config
|
config = model_config.hf_config
|
||||||
config.num_labels = 1
|
config.num_labels = 1
|
||||||
pooler_config = model_config.pooler_config
|
pooler_config = model_config.pooler_config
|
||||||
if pooler_config.logit_bias is None:
|
if pooler_config.logit_mean is None:
|
||||||
pooler_config.logit_bias = 2.65
|
pooler_config.logit_mean = 2.65
|
||||||
|
|
||||||
|
|
||||||
class LlamaBidirectionalConfig(VerifyAndUpdateConfig):
|
class LlamaBidirectionalConfig(VerifyAndUpdateConfig):
|
||||||
|
|||||||
Reference in New Issue
Block a user