Spaces:
Sleeping
Sleeping
Test-set results on an 80/20 split, auto-navigation and full-screen layout
#4
by F4RUKYLDRM - opened
This PR adds the following:
- Full-screen Space (
header: mini). The Space page was stretching the app's frame to full height, so the pinned action bar from PR #3 never stuck. Withheader: mini, the action bar stays visible. - 80/20 train/test split. The split is stratified, and the 20% test set is never used for training or for picking the early-stopping checkpoint. Early stopping uses 1/8 of the training part.
- Results show test-set scores: Test accuracy, Test macro F1 and Test macro AUC, plus the test-set confusion matrix and predictions. Models saved before this change keep their Validation labels.
- Automatic navigation: Build โ Train โ Export โ Predict โ Results. Export gets a "Continue to Predict" button.
- scTrioseq2 is the default example dataset.
- Readability on the Results page: white cards, a readable table font, and numbers rounded to 3 decimals on screen. Downloads keep full precision.
- Diagnostics in
diagnostics/reproducing the paper's results. With the upstream notebook's dense-then-mask first layer, scTrioseq2 reaches test macro F1 0.937 (paper: 0.95). That layer is not in the app yet; adding it needs the lab's decision. COAD'slabel.csvdoes not match its samples and needs the correct file.
All 25 tests pass. This PR is based on the current main and merges without conflicts.
bozdaglab changed pull request status to open
bozdaglab changed pull request status to merged