Results of a training epoch. More...

#include <sb.h>

Data Fields
float	loss
	SVL loss. More...

float	validation_loss
	SVL validation loss. More...

float	accuracy
	SVL accuracy metric. More...

float	validation_accuracy
	SVL validation accuracy metric. More...

Detailed Description

Results of a training epoch.

Used only by Deep Surface and Deep Cortex projects.

Definition at line 12622 of file sb.h.

Field Documentation

◆ accuracy

float sb_t_svl_res_epoch::accuracy

SVL accuracy metric.

SVL accuracy of the epoch. The value is computed by a weighted average of all the batch accuracies in the epoch.
According to the project type, the accuracy calculation is different:

Deep Cortex :
$accuracy = \frac{TP}{TP + FP + FN}$
where counters are considered at image level.
In multi-model case may happen that a FALSE NEGATIVE sample is associated with a FALSE POSITIVE sample. In this case the error weighs double in the accuracy calculation.
Deep Surface :
$accuracy = \frac{2*TP}{2*TP + FP + FN}$
where counters are considered at pixel level.

Attention: The value may differ from sb_t_stat::accuracy computed on SVL images at detection time. Mainly the differences are caused by the perturbations, if enabled.

Definition at line 12657 of file sb.h.

◆ loss

float sb_t_svl_res_epoch::loss

SVL loss.

Average batch training loss of the epoch.
It is a measure of the error between predictions and ground truth values on training images at the end of an epoch.
The value is available only in global results.

Note: SVL loss is not computed for each model.

Definition at line 12631 of file sb.h.

◆ validation_accuracy

float sb_t_svl_res_epoch::validation_accuracy

SVL validation accuracy metric.

SVL validation accuracy of the epoch. The value is computed by a weighted average of all validation batch accuracies in the epoch.
If sb_t_svl_dl_par::validation_percentage = 0, it is equal to zero.
See sb_t_svl_res_epoch::accuracy for metric calculation.

Attention: The value may differ from sb_t_stat::accuracy computed on validation images at detection time. Mainly the differences are caused by the perturbations, if enabled.

Definition at line 12666 of file sb.h.

◆ validation_loss

float sb_t_svl_res_epoch::validation_loss

SVL validation loss.

Average batch validation loss of the epoch.
It is a measure of the error between predictions and ground truth values on validation images at the end of an epoch. If sb_t_svl_dl_par::validation_percentage = 0, it is equal to zero.
The value is available only in global results.

Note: SVL validation loss is not computed for each model.

Definition at line 12640 of file sb.h.

The documentation for this struct was generated from the following file: