1 utils.metric_reductions

utils.metric_reductions

Shared finalization of additive metric totals and observation counts.

Metric owners choose which observations contribute to each state. After local or distributed accumulation, this operation computes their mean without turning an unobserved population into a measured zero.

1.1 Functions

Name	Description
mean_from_sum_count	Divide accumulated totals by counts, yielding NaN for nonpositive counts.

1.1.1 mean_from_sum_count

utils.metric_reductions.mean_from_sum_count(total, count)

Divide accumulated totals by counts, yielding NaN for nonpositive counts.

1.1.1.1 Parameters

Name	Type	Description	Default
total `Tensor["...", float32\|float64]`		Accumulated metric sums.	required
count `Tensor["...", float32\|float64]`		Observation counts on the same device, broadcast-compatible with `total`. Counts are normally nonnegative integers stored in floating-point metric state.	required

1.1.1.2 Returns

Name	Type	Description
	Tensor	`Tensor["...", float32\\|float64]`: Broadcast result, retaining PyTorch
	Tensor	dtype promotion and autograd. Positive denominators retain the existing
	Tensor	clamp to at least one; nonpositive counts produce NaN. Nonfinite totals
	Tensor	propagate for positive counts. Neither input is mutated.

# utils.metric_reductions { #aria_nbv.utils.metric_reductions } `utils.metric_reductions` Shared finalization of additive metric totals and observation counts. Metric owners choose which observations contribute to each state. After local or distributed accumulation, this operation computes their mean without turning an unobserved population into a measured zero. ## Functions | Name | Description | | --- | --- | | [mean_from_sum_count](#aria_nbv.utils.metric_reductions.mean_from_sum_count) | Divide accumulated totals by counts, yielding NaN for nonpositive counts. | ### mean_from_sum_count { #aria_nbv.utils.metric_reductions.mean_from_sum_count } ```python utils.metric_reductions.mean_from_sum_count(total, count) ``` Divide accumulated totals by counts, yielding NaN for nonpositive counts. #### Parameters {.doc-section .doc-section-parameters} | Name | Type | Description | Default | |------------------------------------------|--------|-------------------------------------------------------------------------------------------------------------------------------------------------------------|------------| | total ``Tensor["...", float32|float64]`` | | Accumulated metric sums. | _required_ | | count ``Tensor["...", float32|float64]`` | | Observation counts on the same device, broadcast-compatible with ``total``. Counts are normally nonnegative integers stored in floating-point metric state. | _required_ | #### Returns {.doc-section .doc-section-returns} | Name | Type | Description | |--------|--------------------------|--------------------------------------------------------------------------| | | [Tensor](`torch.Tensor`) | ``Tensor["...", float32\|float64]``: Broadcast result, retaining PyTorch | | | [Tensor](`torch.Tensor`) | dtype promotion and autograd. Positive denominators retain the existing | | | [Tensor](`torch.Tensor`) | clamp to at least one; nonpositive counts produce NaN. Nonfinite totals | | | [Tensor](`torch.Tensor`) | propagate for positive counts. Neither input is mutated. |