1 CoralLayer

rri_metrics.CoralLayer(in_dim, num_classes, *, preinit_bias=True)

CORAL output layer with shared weights and per-threshold biases.

1.1 This implements logits

logit_k = w^T x + b_k, k = 0..K-2

1.2 Methods

Name	Description
forward	Compute threshold logits.
init_bin_values	Initialize (or re-initialize) monotone bin representatives u_k.
init_bias_from_priors	Initialize CORAL biases from class priors.
expected_from_probs	Compute expected continuous value using learnable bin values.
expected_from_logits	Convert logits to marginals and compute expected continuous value.
bin_value_regularizer	L2 penalty to keep learnable bin values close to target values.

1.2.1 forward

rri_metrics.CoralLayer.forward(x)

Compute threshold logits.

1.2.1.1 Parameters

Name	Type	Description	Default
x	Tensor	Tensor["... in_dim"].	required

1.2.1.2 Returns

Name	Type	Description
	Tensor	Tensor["... K-1"].

1.2.2 init_bin_values

rri_metrics.CoralLayer.init_bin_values(values, *, overwrite=False)

Initialize (or re-initialize) monotone bin representatives u_k.

1.2.3 init_bias_from_priors

rri_metrics.CoralLayer.init_bias_from_priors(priors, *, overwrite=True)

Initialize CORAL biases from class priors.

1.2.3.1 Parameters

Name	Type	Description	Default
priors	Tensor	Tensor["K"] class priors that sum to 1.	required
overwrite	bool	If True, overwrite the current bias values.	True

1.2.4 expected_from_probs

rri_metrics.CoralLayer.expected_from_probs(probs)

Compute expected continuous value using learnable bin values.

1.2.5 expected_from_logits

rri_metrics.CoralLayer.expected_from_logits(logits)

Convert logits to marginals and compute expected continuous value.

1.2.6 bin_value_regularizer

rri_metrics.CoralLayer.bin_value_regularizer(target_values)

L2 penalty to keep learnable bin values close to target values.