Logging#

Currently we support monitoring the training progress in Tensorboard while basic metrics are reported to the console as well. All metrics that include comparison to ground truth emitters are based on the parameters (implicitly) provided in the configuration .yaml file. Those include match dimensionality (i.e., in 2D or 3D) and max. allowed distances. The threshold on the detection filters the detections prior to matching.

Tensorboard#

Metrics#

Abbreviation

Name

Description

prec

Precision

Number of true positives over all detections

rec

Recall

Number of true positives over all (ground truth) localizations

jac

Jaccard

Intersection over Union

f1

F1 Score

2 x Precision x Recall over (Precision + Recall)

effcy_lat

Efficiency lateral

SMLM Challenge defined efficiency lateral

effcy_ax

Efficiency axial

SMLM Challenge defined efficiency axial

effcy_vol

Efficiency volumetric

SMLM Challenge defined efficiency volumetric

rmse_lat

RMSE lateral

Root Mean Square Error lateral

rmse_ax

RMSE axial

Root Mean Square Error axial (z)

rmse_vol

RMSE volumetric

Root Mean Square Error volumetric

mad_lat

MAD lateral

Mean Absolute Deviation lateral

mad_ax

MAD axial

Mean Absolute Deviation axial

mad_vol

MAD volumetric

Mean Absolute Deviation volumetric

Other (non or less-semantic) metrics

Abbreviation

Description

learning_rate

Learning Rate

test_ep

Test loss per epoch

train_batch

Train loss per batch

train_ep

Train loss per epoch

loss_cmp_0

Gaussian localization loss

loss_cmp_1

Background loss

Distributions#

Abbreviation

Description

dist/prob

Probability distribution of active pixels (active typically means above p >= 0.6)

dist/x_offset

Distribution of relative offset to pixel border in x direction

dist/y_offset

Distribution of relative offset to pixel border in y direction

residuals/phot_gt_pred

Predicted vs ground truth photon counts

residuals/z_gt_pred

Predicted vs ground truth z values

In- Output Logging#

Input shows a sampled input frame with ground truth emitters. Output shows the respective output channels of the model.

Emitter-Out and Emitter-Match#

Abbreviation

Description

em_match

All matched emitter pairs

em_match_3d

All matched emitter pairs in 3D

em_out

All ground truth emitters and all predicted emitters

em_out_3d

All ground truth emitters and all predicted emitters in 3D