tensorplay.func.batch_norm_without_running_stats

tensorplay.func.batch_norm_without_running_stats(module: Module) → None[source]

Drops the running statistics of one batch-normalization module.

Running statistics are updated in place on every forward pass. Under a transform that evaluates the module more than once – mapping over an ensemble, or differentiating through it – those updates would be applied repeatedly and in an order the caller never asked for, so the module is switched to normalizing with the batch’s own statistics instead.