Apppublic

PacificallyOc/bowhead-tensorboard

sourceHugging Facemitupdated 3mo agoView on Hugging Face

0likes

App README

Bowhead Whale Detector — training dashboard

Live TensorBoard for the supervised CNN bowhead-call detector (source repo).

Shows, per run (scratch vs warm-start arms):

—SCALARS — training loss + held-out validation ROC-AUC per epoch, and final test metrics at realistic (~1:8 call:transient) prevalence.
—HPARAMS — warm-start vs random-init ablation compared side by side.

The CNN mirrors the encoder of a 32-D convolutional autoencoder so the unsupervised features warm-start the classifier. Companion code for "Enhancing bowhead whale detection using convolutional autoencoders" (Thode 2026, JASA).

How the logs get here

TensorBoard event files are written to runs/<tag>/ during training and copied into this Space's logs/ directory. To refresh:

bash

rm -rf deploy/tensorboard_space/logs && cp -r runs deploy/tensorboard_space/logs
# then push this folder to the Space (see repo README "Portfolio / hosting")