nnpd.nre.training

Readable BCE training, including exact global-batch sharding for DDP.

Functions

predict_logits(→ numpy.ndarray)
classification(→ dict)
fit_binary(→ dict)	No sample padding or dropping: every training row contributes once per epoch.

Module Contents

nnpd.nre.training.predict_logits(model, x, theta, batch_size: int, device) → numpy.ndarray

nnpd.nre.training.classification(labels: numpy.ndarray, logits: numpy.ndarray) → dict

nnpd.nre.training.fit_binary(context, model) → dict

No sample padding or dropping: every training row contributes once per epoch.

A batch is split by rank. Local loss SUM * world/global_count compensates for DDP’s gradient averaging even for unequal or empty last shards. An empty shard runs a zero-weight dummy forward to participate in collectives. The configured batch size is GLOBAL, not per GPU.