pyhgf.model.fused.step_report#
- pyhgf.model.fused.step_report(pipeline)[source]#
Per-part step magnitudes after a training step, for rate calibration.
Returns one entry per learning part of the last
FusedPipeline.step()call: the norm of its applied weight change and of the error it received, labelled by the part’s path inside the model. Reading the two side by side shows at a glance which parts are being over- or under-driven — the practical symptom of a mis-calibrated learning rate (see the architecture guide on choosing the optimiser and rate).- Parameters:
pipeline (FusedPipeline) – The executor to read. Its part tree must have taken at least one step.
- Returns:
One entry per learning part, in pipeline order, with keys
"part"(its path inside the model),"layer_sizes"(the wrapped network’s layout),"update_norm"and"error_norm".- Return type: