From 9f74d15ad9f0ea7f478ef273d60ca56d863a2b61 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz> Date: Tue, 6 Oct 2026 10:46:56 +0200 Subject: [PATCH] chore: point to recomputed tissue stats and filter tiles MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Both artifacts were regenerated after the tile->mask scale fix (#19). The corrected tissue coverage changes the training set substantially: the tissue filter used to drop 29.9% of annotated train tiles (1572719 -> 1102086) and 29.8% of annotated test tiles (483171 -> 339038), because tiles were looked up at half their true position in the mask. It now drops 0.14% and 0.28% respectively, which matches reality — annotations are drawn on tissue, so an annotated tile almost always has tissue. Downstream artifacts are stale and need regenerating: kfold split, virchow2 embeddings, the final classifier, its test metrics, and the tissue-tile embeddings used for prediction maps. Co-Authored-By: Claude Opus 5 --- configs/data/dataset.yaml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/configs/data/dataset.yaml b/configs/data/dataset.yaml index 766849f5..7299c1d3 100644 --- a/configs/data/dataset.yaml +++ b/configs/data/dataset.yaml @@ -13,14 +13,14 @@ dataset: train_split_filename: "split_mapping/train_split.csv" test_split_filename: "split_mapping/test_split.csv" tiling_run_id: "fdf7550a2004474f8c7a05dc0cf1fd86" - filter_tiles_run_id: "4e8f5d3c82124ea5a8f871a42d3ed9ba" + filter_tiles_run_id: "9a844082b48d47bfa5d3e7a37d7de268" stratified_kfold_run_id: "850c81506684450b9af92296acfd045a" stratified_group_kfold_run_id: "382b41d2fa894514908e8067949c4326" embedding_run_ids: virchow2: "262d680f52374614baaf19f1e4d05940" provgigapath: "5099ce57e886411dababfc1234a88b23" tissue_masks_run_id: "52bc0924f8624b259819c480c7cf213f" - tissue_stats_run_id: "16ae2d003d88471b924e5f332415232a" + tissue_stats_run_id: "f3401b702f044ac097a177bea30161e5" exclusions: bad_slides: