From 2a9a59bf2c83371851e08c3f226ff439971741d7 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 22 Apr 2026 21:22:57 +0200
Subject: [PATCH 01/42] feat: add tiling run_id

---
 configs/data/dataset.yaml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/configs/data/dataset.yaml b/configs/data/dataset.yaml
index ef6cbe7..625d07d 100644
--- a/configs/data/dataset.yaml
+++ b/configs/data/dataset.yaml
@@ -12,6 +12,7 @@ dataset:
     train_test_split_run_id: "90d7c88418ae4b879c3febbb70e9e741"
     train_split_filename: "split_mapping/train_split.csv"
     test_split_filename: "split_mapping/test_split.csv"
+    tiling_run_id: "fdf7550a2004474f8c7a05dc0cf1fd86"
 
   exclusions:
     bad_slides:

From bb6dc1bcbf2fada46970c710e6eef45adab23da1 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 22 Apr 2026 21:25:19 +0200
Subject: [PATCH 02/42] feat: add configs for tile masks

---
 .../experiment/preprocessing/tile_masks_05mpp.yaml    |  8 ++++++++
 configs/preprocessing/tile_masks.yaml                 | 11 +++++++++++
 2 files changed, 19 insertions(+)
 create mode 100644 configs/experiment/preprocessing/tile_masks_05mpp.yaml
 create mode 100644 configs/preprocessing/tile_masks.yaml

diff --git a/configs/experiment/preprocessing/tile_masks_05mpp.yaml b/configs/experiment/preprocessing/tile_masks_05mpp.yaml
new file mode 100644
index 0000000..fee1fa7
--- /dev/null
+++ b/configs/experiment/preprocessing/tile_masks_05mpp.yaml
@@ -0,0 +1,8 @@
+# @package _global_
+
+defaults:
+  - /data: dataset
+  - _self_
+
+slides_artifact_path: train_split/slides.parquet
+tiles_artifact_path: train_split/tiles.parquet
diff --git a/configs/preprocessing/tile_masks.yaml b/configs/preprocessing/tile_masks.yaml
new file mode 100644
index 0000000..550c07e
--- /dev/null
+++ b/configs/preprocessing/tile_masks.yaml
@@ -0,0 +1,11 @@
+# @package _global_
+
+slides_artifact_path: ???
+tiles_artifact_path: ???
+
+max_concurrent: 8
+mlflow_artifact_path: tile_masks
+
+metadata:
+  run_name: Tile Masks ${dataset.name}
+  description: "Tile mask generation"

From 8b5e5ac5e00045916287f84e6e0d39d6a30e0244 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 22 Apr 2026 21:27:09 +0200
Subject: [PATCH 03/42] feat: implement tile mask pipeline

---
 preprocessing/tile_masks.py | 83 +++++++++++++++++++++++++++++++++++++
 1 file changed, 83 insertions(+)
 create mode 100644 preprocessing/tile_masks.py

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
new file mode 100644
index 0000000..4c8cf85
--- /dev/null
+++ b/preprocessing/tile_masks.py
@@ -0,0 +1,83 @@
+from pathlib import Path
+from tempfile import TemporaryDirectory
+
+import hydra
+import mlflow
+import pandas as pd
+import pyvips
+import ray
+from omegaconf import DictConfig
+from rationai.masks import process_items, tile_mask, write_big_tiff
+from rationai.mlkit import autolog, with_cli_args
+from rationai.mlkit.lightning.loggers import MLFlowLogger
+
+
+@ray.remote(num_cpus=1, memory=3 * 1024**3)
+def process_slide(
+    item: tuple[dict, pd.DataFrame],
+    output_dir: str,
+) -> None:
+    slide, slide_tiles = item
+    slide_path = Path(slide["path"])
+
+    mask = tile_mask(
+        slide_tiles,
+        tile_extent=(slide["tile_extent_x"], slide["tile_extent_y"]),
+        size=(slide["extent_x"], slide["extent_y"]),
+    )
+
+    width, height = mask.size
+    mask_bytes = mask.tobytes()
+    del mask
+
+    write_big_tiff(
+        image=pyvips.Image.new_from_memory(
+            data=mask_bytes, width=width, height=height, bands=1, format="uchar"
+        ),
+        path=Path(output_dir, slide_path.with_suffix(".tiff").name),
+        mpp_x=slide["mpp_x"],
+        mpp_y=slide["mpp_y"],
+    )
+
+
+@with_cli_args(["+preprocessing=tile_masks"])
+@hydra.main(config_path="../configs", config_name="preprocessing", version_base=None)
+@autolog
+def main(config: DictConfig, logger: MLFlowLogger) -> None:
+    tiling_run_id = config.dataset.mlflow_artifacts.tiling_run_id
+
+    slides = pd.read_parquet(
+        mlflow.artifacts.download_artifacts(
+            run_id=tiling_run_id,
+            artifact_path=config.slides_artifact_path,
+        )
+    )
+    tiles = pd.read_parquet(
+        mlflow.artifacts.download_artifacts(
+            run_id=tiling_run_id,
+            artifact_path=config.tiles_artifact_path,
+        )
+    )
+
+    items = [
+        (slide.to_dict(), tiles[tiles["slide_id"] == slide["id"]][["x", "y"]])
+        for _, slide in slides.iterrows()
+    ]
+
+    with TemporaryDirectory() as output_dir:
+        process_items(
+            items,
+            process_item=process_slide,
+            fn_kwargs={"output_dir": output_dir},
+            max_concurrent=config.max_concurrent,
+        )
+
+        logger.log_artifacts(
+            local_dir=output_dir, artifact_path=config.mlflow_artifact_path
+        )
+
+
+if __name__ == "__main__":
+    ray.init()
+    main()
+    ray.shutdown()

From 04283b2872fa5e9ee764a792896e1f72497d0292 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 22 Apr 2026 21:27:46 +0200
Subject: [PATCH 04/42] feat: add sumbission script for the tile masks

---
 scripts/submit_tile_masks.py | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)
 create mode 100644 scripts/submit_tile_masks.py

diff --git a/scripts/submit_tile_masks.py b/scripts/submit_tile_masks.py
new file mode 100644
index 0000000..1356a6a
--- /dev/null
+++ b/scripts/submit_tile_masks.py
@@ -0,0 +1,18 @@
+from kube_jobs import storage, submit_job
+
+
+submit_job(
+    job_name="tissue-classification-tile-masks",
+    username=...,
+    cpu=8,
+    memory="32Gi",
+    gpu=None,
+    public=False,
+    script=[
+        "git clone https://gitlab.ics.muni.cz/rationai/digital-pathology/pathology/tissue-classification.git workdir",
+        "cd workdir",
+        "uv sync",
+        "uv run python -m preprocessing.tile_masks +experiment=...",
+    ],
+    storage=[storage.secure.DATA],
+)

From 654710a51eafecbc5ba8b1b830eb2c8edbf4b811 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 22 Apr 2026 21:35:49 +0200
Subject: [PATCH 05/42] fix: group by slides before the loop

---
 preprocessing/tile_masks.py | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 4c8cf85..4ef0fde 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -59,8 +59,13 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
         )
     )
 
+    tiles_by_slide = {
+        slide_id: group[["x", "y"]]
+        for slide_id, group in tiles.groupby("slide_id")
+    }
+
     items = [
-        (slide.to_dict(), tiles[tiles["slide_id"] == slide["id"]][["x", "y"]])
+        (slide.to_dict(), tiles_by_slide.get(slide["id"], pd.DataFrame(columns=["x", "y"])))
         for _, slide in slides.iterrows()
     ]
 

From 93eecb9814ab8791f7be217631ef555350768d5d Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 22 Apr 2026 21:44:07 +0200
Subject: [PATCH 06/42] fix: update submission script to use GitHub repository
 URL

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 scripts/submit_tile_masks.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/scripts/submit_tile_masks.py b/scripts/submit_tile_masks.py
index 1356a6a..bb57a59 100644
--- a/scripts/submit_tile_masks.py
+++ b/scripts/submit_tile_masks.py
@@ -9,7 +9,7 @@
     gpu=None,
     public=False,
     script=[
-        "git clone https://gitlab.ics.muni.cz/rationai/digital-pathology/pathology/tissue-classification.git workdir",
+        "git clone https://github.com/RationAI/tissue-classification.git workdir",
         "cd workdir",
         "uv sync",
         "uv run python -m preprocessing.tile_masks +experiment=...",

From 682a5d3cd0876971e8d1e737d31796e8b4acc07e Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 22 Apr 2026 21:58:50 +0200
Subject: [PATCH 07/42] perf: load only required columns from tiles parquet

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 preprocessing/tile_masks.py | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 4ef0fde..cf3d0e3 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -56,11 +56,12 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
         mlflow.artifacts.download_artifacts(
             run_id=tiling_run_id,
             artifact_path=config.tiles_artifact_path,
-        )
+        ),
+        columns=["slide_id", "x", "y"],
     )
 
     tiles_by_slide = {
-        slide_id: group[["x", "y"]]
+        slide_id: group.drop(columns="slide_id")
         for slide_id, group in tiles.groupby("slide_id")
     }
 

From cb29d16aa99326a5cba5624f65f23315d3c3374a Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 22 Apr 2026 22:14:39 +0200
Subject: [PATCH 08/42] perf: replace PIL tile_mask with vectorized numpy
 implementation

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 preprocessing/tile_masks.py | 50 ++++++++++++++++++++++++++++++++-----
 1 file changed, 44 insertions(+), 6 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index cf3d0e3..02851d0 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -3,15 +3,55 @@
 
 import hydra
 import mlflow
+import numpy as np
 import pandas as pd
 import pyvips
 import ray
 from omegaconf import DictConfig
-from rationai.masks import process_items, tile_mask, write_big_tiff
+from rationai.masks import process_items, write_big_tiff
 from rationai.mlkit import autolog, with_cli_args
 from rationai.mlkit.lightning.loggers import MLFlowLogger
 
 
+def _draw_tile_outlines(
+    tiles: pd.DataFrame,
+    tile_extent: tuple[int, int],
+    size: tuple[int, int],
+    outline_width: int = 4,
+) -> np.ndarray:
+    tw, th = tile_extent
+    width, height = size
+    ow = outline_width
+    mask = np.zeros((height, width), dtype=np.uint8)
+
+    xs = tiles["x"].to_numpy()
+    ys = tiles["y"].to_numpy()
+
+    dc = np.arange(tw)
+    dr = np.arange(th)
+    dw = np.arange(ow)
+
+    # Horizontal borders (top and bottom)
+    h_rows_top = (ys[:, None, None] + dw[None, :, None]).ravel()
+    h_rows_bot = (ys[:, None, None] + (th - ow) + dw[None, :, None]).ravel()
+    h_cols = (xs[:, None, None] + dc[None, None, :]).ravel()
+    h_cols = np.tile(h_cols.reshape(len(xs), 1, tw), (1, ow, 1)).ravel()
+    mask[h_rows_top, h_cols] = 255
+    mask[h_rows_bot, h_cols] = 255
+
+    # Vertical borders (left and right)
+    v_rows = (ys[:, None, None] + dr[None, :, None]).ravel()
+    v_rows = np.tile(v_rows.reshape(len(xs), th, 1), (1, 1, ow)).ravel()
+    v_cols_left = (xs[:, None, None] + dw[None, None, :]).ravel()
+    v_cols_right = (xs[:, None, None] + (tw - ow) + dw[None, None, :]).ravel()
+    v_cols_left = np.tile(v_cols_left.reshape(len(xs), 1, ow), (1, th, 1)).ravel()
+    v_cols_right = np.tile(v_cols_right.reshape(len(xs), 1, ow), (1, th, 1)).ravel()
+    mask[v_rows, v_cols_left] = 255
+    mask[v_rows, v_cols_right] = 255
+
+    return mask
+
+
 @ray.remote(num_cpus=1, memory=3 * 1024**3)
 def process_slide(
     item: tuple[dict, pd.DataFrame],
@@ -20,19 +60,17 @@ def process_slide(
     slide, slide_tiles = item
     slide_path = Path(slide["path"])
 
-    mask = tile_mask(
+    mask = _draw_tile_outlines(
         slide_tiles,
         tile_extent=(slide["tile_extent_x"], slide["tile_extent_y"]),
         size=(slide["extent_x"], slide["extent_y"]),
     )
 
-    width, height = mask.size
-    mask_bytes = mask.tobytes()
-    del mask
+    height, width = mask.shape
 
     write_big_tiff(
         image=pyvips.Image.new_from_memory(
-            data=mask_bytes, width=width, height=height, bands=1, format="uchar"
+            data=mask.tobytes(), width=width, height=height, bands=1, format="uchar"
         ),
         path=Path(output_dir, slide_path.with_suffix(".tiff").name),
         mpp_x=slide["mpp_x"],

From 7d2e4759df6088afd909f8ba264d609fd204c1be Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 22 Apr 2026 22:19:05 +0200
Subject: [PATCH 09/42] fix: correct shape mismatch in vectorized tile outline
 drawing

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 preprocessing/tile_masks.py | 32 +++++++++++---------------------
 1 file changed, 11 insertions(+), 21 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 02851d0..edf3d7d 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -27,27 +27,17 @@ def _draw_tile_outlines(
     xs = tiles["x"].to_numpy()
     ys = tiles["y"].to_numpy()
 
-    dc = np.arange(tw)
-    dr = np.arange(th)
-    dw = np.arange(ow)
-
-    # Horizontal borders (top and bottom)
-    h_rows_top = (ys[:, None, None] + dw[None, :, None]).ravel()
-    h_rows_bot = (ys[:, None, None] + (th - ow) + dw[None, :, None]).ravel()
-    h_cols = (xs[:, None, None] + dc[None, None, :]).ravel()
-    h_cols = np.tile(h_cols.reshape(len(xs), 1, tw), (1, ow, 1)).ravel()
-    mask[h_rows_top, h_cols] = 255
-    mask[h_rows_bot, h_cols] = 255
-
-    # Vertical borders (left and right)
-    v_rows = (ys[:, None, None] + dr[None, :, None]).ravel()
-    v_rows = np.tile(v_rows.reshape(len(xs), th, 1), (1, 1, ow)).ravel()
-    v_cols_left = (xs[:, None, None] + dw[None, None, :]).ravel()
-    v_cols_right = (xs[:, None, None] + (tw - ow) + dw[None, None, :]).ravel()
-    v_cols_left = np.tile(v_cols_left.reshape(len(xs), 1, ow), (1, th, 1)).ravel()
-    v_cols_right = np.tile(v_cols_right.reshape(len(xs), 1, ow), (1, th, 1)).ravel()
-    mask[v_rows, v_cols_left] = 255
-    mask[v_rows, v_cols_right] = 255
+    # Precompute shared index spans — O(n × tile_size) each
+    h_cols = (xs[:, None] + np.arange(tw)[None, :]).ravel()  # (n × tw,)
+    v_rows = (ys[:, None] + np.arange(th)[None, :]).ravel()  # (n × th,)
+
+    for dy in range(ow):
+        mask[np.repeat(ys + dy, tw), h_cols] = 255              # top border
+        mask[np.repeat(ys + th - ow + dy, tw), h_cols] = 255    # bottom border
+
+    for dx in range(ow):
+        mask[v_rows, np.repeat(xs + dx, th)] = 255              # left border
+        mask[v_rows, np.repeat(xs + tw - ow + dx, th)] = 255    # right border
 
     return mask
 

From b94b3f850db7b71df2e8c126744c34d4b3012924 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 22 Apr 2026 22:23:06 +0200
Subject: [PATCH 10/42] fix: clip out-of-bounds tile pixels in vectorized
 outline drawing

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 preprocessing/tile_masks.py | 15 +++++++++++----
 1 file changed, 11 insertions(+), 4 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index edf3d7d..8a6fb1a 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -31,13 +31,20 @@ def _draw_tile_outlines(
     h_cols = (xs[:, None] + np.arange(tw)[None, :]).ravel()  # (n × tw,)
     v_rows = (ys[:, None] + np.arange(th)[None, :]).ravel()  # (n × th,)
 
+    h_cols_in = (h_cols >= 0) & (h_cols < width)
+    v_rows_in = (v_rows >= 0) & (v_rows < height)
+
     for dy in range(ow):
-        mask[np.repeat(ys + dy, tw), h_cols] = 255              # top border
-        mask[np.repeat(ys + th - ow + dy, tw), h_cols] = 255    # bottom border
+        for ry in (ys + dy, ys + th - ow + dy):
+            rows = np.repeat(ry, tw)
+            ok = h_cols_in & (rows >= 0) & (rows < height)
+            mask[rows[ok], h_cols[ok]] = 255
 
     for dx in range(ow):
-        mask[v_rows, np.repeat(xs + dx, th)] = 255              # left border
-        mask[v_rows, np.repeat(xs + tw - ow + dx, th)] = 255    # right border
+        for cx in (xs + dx, xs + tw - ow + dx):
+            cols = np.repeat(cx, th)
+            ok = v_rows_in & (cols >= 0) & (cols < width)
+            mask[v_rows[ok], cols[ok]] = 255
 
     return mask
 

From ef4e51ec263cb3dee3eed312a7400ab3d0c4c822 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 22 Apr 2026 22:30:02 +0200
Subject: [PATCH 11/42] perf: add configurable downsampling to tile mask
 generation

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 .../preprocessing/tile_masks_05mpp.yaml       |  1 +
 configs/preprocessing/tile_masks.yaml         |  5 +++-
 preprocessing/tile_masks.py                   | 27 ++++++++++++++-----
 3 files changed, 26 insertions(+), 7 deletions(-)

diff --git a/configs/experiment/preprocessing/tile_masks_05mpp.yaml b/configs/experiment/preprocessing/tile_masks_05mpp.yaml
index fee1fa7..b353855 100644
--- a/configs/experiment/preprocessing/tile_masks_05mpp.yaml
+++ b/configs/experiment/preprocessing/tile_masks_05mpp.yaml
@@ -6,3 +6,4 @@ defaults:
 
 slides_artifact_path: train_split/slides.parquet
 tiles_artifact_path: train_split/tiles.parquet
+downsample: 4
diff --git a/configs/preprocessing/tile_masks.yaml b/configs/preprocessing/tile_masks.yaml
index 550c07e..2a4d8d2 100644
--- a/configs/preprocessing/tile_masks.yaml
+++ b/configs/preprocessing/tile_masks.yaml
@@ -2,10 +2,13 @@
 
 slides_artifact_path: ???
 tiles_artifact_path: ???
+downsample: 1
 
 max_concurrent: 8
 mlflow_artifact_path: tile_masks
 
 metadata:
   run_name: Tile Masks ${dataset.name}
-  description: "Tile mask generation"
+  description: "Tile mask generation (downsample ${downsample}x)"
+  hyperparams:
+    downsample: ${downsample}
diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 8a6fb1a..5282a5c 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -53,14 +53,26 @@ def _draw_tile_outlines(
 def process_slide(
     item: tuple[dict, pd.DataFrame],
     output_dir: str,
+    downsample: int,
 ) -> None:
     slide, slide_tiles = item
     slide_path = Path(slide["path"])
 
+    scaled_tiles = slide_tiles.assign(
+        x=slide_tiles["x"] // downsample,
+        y=slide_tiles["y"] // downsample,
+    )
+
     mask = _draw_tile_outlines(
-        slide_tiles,
-        tile_extent=(slide["tile_extent_x"], slide["tile_extent_y"]),
-        size=(slide["extent_x"], slide["extent_y"]),
+        scaled_tiles,
+        tile_extent=(
+            slide["tile_extent_x"] // downsample,
+            slide["tile_extent_y"] // downsample,
+        ),
+        size=(
+            slide["extent_x"] // downsample,
+            slide["extent_y"] // downsample,
+        ),
     )
 
     height, width = mask.shape
@@ -70,8 +82,8 @@ def process_slide(
             data=mask.tobytes(), width=width, height=height, bands=1, format="uchar"
         ),
         path=Path(output_dir, slide_path.with_suffix(".tiff").name),
-        mpp_x=slide["mpp_x"],
-        mpp_y=slide["mpp_y"],
+        mpp_x=slide["mpp_x"] * downsample,
+        mpp_y=slide["mpp_y"] * downsample,
     )
 
 
@@ -109,7 +121,10 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
         process_items(
             items,
             process_item=process_slide,
-            fn_kwargs={"output_dir": output_dir},
+            fn_kwargs={
+                "output_dir": output_dir,
+                "downsample": config.downsample,
+            },
             max_concurrent=config.max_concurrent,
         )
 

From f96c636492c3f6c56af0f9f82be865812f427e0c Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 22 Apr 2026 22:33:54 +0200
Subject: [PATCH 12/42] style: drop comments with ambiguous unicode
 multiplication sign

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 preprocessing/tile_masks.py | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 5282a5c..eed8895 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -27,9 +27,8 @@ def _draw_tile_outlines(
     xs = tiles["x"].to_numpy()
     ys = tiles["y"].to_numpy()
 
-    # Precompute shared index spans — O(n × tile_size) each
-    h_cols = (xs[:, None] + np.arange(tw)[None, :]).ravel()  # (n × tw,)
-    v_rows = (ys[:, None] + np.arange(th)[None, :]).ravel()  # (n × th,)
+    h_cols = (xs[:, None] + np.arange(tw)[None, :]).ravel()
+    v_rows = (ys[:, None] + np.arange(th)[None, :]).ravel()
 
     h_cols_in = (h_cols >= 0) & (h_cols < width)
     v_rows_in = (v_rows >= 0) & (v_rows < height)
@@ -113,7 +112,10 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
     }
 
     items = [
-        (slide.to_dict(), tiles_by_slide.get(slide["id"], pd.DataFrame(columns=["x", "y"])))
+        (
+            slide.to_dict(),
+            tiles_by_slide.get(slide["id"], pd.DataFrame(columns=["x", "y"])),
+        )
         for _, slide in slides.iterrows()
     ]
 

From 8504b840c694eb84717a733a16ac152dda67f9e4 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 22 Apr 2026 22:42:57 +0200
Subject: [PATCH 13/42] chore: increase tile mask downsample to 8x to avoid OOM

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 configs/experiment/preprocessing/tile_masks_05mpp.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/configs/experiment/preprocessing/tile_masks_05mpp.yaml b/configs/experiment/preprocessing/tile_masks_05mpp.yaml
index b353855..af89e15 100644
--- a/configs/experiment/preprocessing/tile_masks_05mpp.yaml
+++ b/configs/experiment/preprocessing/tile_masks_05mpp.yaml
@@ -6,4 +6,4 @@ defaults:
 
 slides_artifact_path: train_split/slides.parquet
 tiles_artifact_path: train_split/tiles.parquet
-downsample: 4
+downsample: 8

From 6ad1b6221774169c05db4c2b237e0508b96bcf1c Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Sat, 25 Apr 2026 00:43:10 +0200
Subject: [PATCH 14/42] fix: tile boundaries

---
 preprocessing/tile_masks.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index eed8895..ad28e95 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -21,7 +21,7 @@ def _draw_tile_outlines(
 ) -> np.ndarray:
     tw, th = tile_extent
     width, height = size
-    ow = outline_width
+    ow = min(outline_width, tw // 2, th // 2)
     mask = np.zeros((height, width), dtype=np.uint8)
 
     xs = tiles["x"].to_numpy()

From d2d1e7cae3dee26631aa349b48a642c6d28645cf Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Sat, 25 Apr 2026 00:46:03 +0200
Subject: [PATCH 15/42] fix: set default downsample to 8 in base config, remove
 redundant override

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 configs/experiment/preprocessing/tile_masks_05mpp.yaml | 1 -
 configs/preprocessing/tile_masks.yaml                  | 2 +-
 2 files changed, 1 insertion(+), 2 deletions(-)

diff --git a/configs/experiment/preprocessing/tile_masks_05mpp.yaml b/configs/experiment/preprocessing/tile_masks_05mpp.yaml
index af89e15..fee1fa7 100644
--- a/configs/experiment/preprocessing/tile_masks_05mpp.yaml
+++ b/configs/experiment/preprocessing/tile_masks_05mpp.yaml
@@ -6,4 +6,3 @@ defaults:
 
 slides_artifact_path: train_split/slides.parquet
 tiles_artifact_path: train_split/tiles.parquet
-downsample: 8
diff --git a/configs/preprocessing/tile_masks.yaml b/configs/preprocessing/tile_masks.yaml
index 2a4d8d2..52f5f36 100644
--- a/configs/preprocessing/tile_masks.yaml
+++ b/configs/preprocessing/tile_masks.yaml
@@ -2,7 +2,7 @@
 
 slides_artifact_path: ???
 tiles_artifact_path: ???
-downsample: 1
+downsample: 8
 
 max_concurrent: 8
 mlflow_artifact_path: tile_masks

From fc51a29ca73fd9414fd755b75d1014176b5cdc11 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Sat, 25 Apr 2026 00:46:53 +0200
Subject: [PATCH 16/42] fix: use slide id as output filename to prevent
 cross-source collisions

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 preprocessing/tile_masks.py | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index ad28e95..8e689c0 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -55,7 +55,6 @@ def process_slide(
     downsample: int,
 ) -> None:
     slide, slide_tiles = item
-    slide_path = Path(slide["path"])
 
     scaled_tiles = slide_tiles.assign(
         x=slide_tiles["x"] // downsample,
@@ -80,7 +79,7 @@ def process_slide(
         image=pyvips.Image.new_from_memory(
             data=mask.tobytes(), width=width, height=height, bands=1, format="uchar"
         ),
-        path=Path(output_dir, slide_path.with_suffix(".tiff").name),
+        path=Path(output_dir, f"{slide['id']}.tiff"),
         mpp_x=slide["mpp_x"] * downsample,
         mpp_y=slide["mpp_y"] * downsample,
     )

From bf6ae43c165a0e1508b222b948f6de3d5fa3e345 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Sat, 25 Apr 2026 15:45:39 +0200
Subject: [PATCH 17/42] fix: mypy

---
 preprocessing/tile_masks.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 8e689c0..19875b1 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -50,7 +50,7 @@ def _draw_tile_outlines(
 
 @ray.remote(num_cpus=1, memory=3 * 1024**3)
 def process_slide(
-    item: tuple[dict, pd.DataFrame],
+    item: tuple[dict[str, object], pd.DataFrame],
     output_dir: str,
     downsample: int,
 ) -> None:

From b32c6b6b3d2b2a254c7eef970323215ff16552be Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Mon, 27 Apr 2026 17:07:32 +0200
Subject: [PATCH 18/42] fix(tile_masks): resolve mypy type errors

Use dict[str, Any] instead of dict[str, object] so arithmetic on slide
fields type-checks, and add an explicit annotation on items so
process_items can infer its type parameter.
---
 preprocessing/tile_masks.py | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 19875b1..25a419b 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -1,5 +1,6 @@
 from pathlib import Path
 from tempfile import TemporaryDirectory
+from typing import Any
 
 import hydra
 import mlflow
@@ -50,7 +51,7 @@ def _draw_tile_outlines(
 
 @ray.remote(num_cpus=1, memory=3 * 1024**3)
 def process_slide(
-    item: tuple[dict[str, object], pd.DataFrame],
+    item: tuple[dict[str, Any], pd.DataFrame],
     output_dir: str,
     downsample: int,
 ) -> None:
@@ -110,9 +111,9 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
         for slide_id, group in tiles.groupby("slide_id")
     }
 
-    items = [
+    items: list[tuple[dict[str, Any], pd.DataFrame]] = [
         (
-            slide.to_dict(),
+            {str(k): v for k, v in slide.to_dict().items()},
             tiles_by_slide.get(slide["id"], pd.DataFrame(columns=["x", "y"])),
         )
         for _, slide in slides.iterrows()

From de381f3401ccecb5da9be6441cbafa2ec5e8430f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Tue, 28 Apr 2026 10:29:00 +0200
Subject: [PATCH 19/42] refactor(tile_masks): use rationai.masks.tile_mask and
 name outputs by WSI stem

Replace the local numpy-based outline drawing with the library's tile_mask
helper, and name output TIFFs by the WSI stem so masks map directly to
their source slides.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---
 preprocessing/tile_masks.py | 44 ++++---------------------------------
 1 file changed, 4 insertions(+), 40 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 25a419b..780d054 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -4,51 +4,15 @@
 
 import hydra
 import mlflow
-import numpy as np
 import pandas as pd
 import pyvips
 import ray
 from omegaconf import DictConfig
-from rationai.masks import process_items, write_big_tiff
+from rationai.masks import process_items, tile_mask, write_big_tiff
 from rationai.mlkit import autolog, with_cli_args
 from rationai.mlkit.lightning.loggers import MLFlowLogger
 
 
-def _draw_tile_outlines(
-    tiles: pd.DataFrame,
-    tile_extent: tuple[int, int],
-    size: tuple[int, int],
-    outline_width: int = 4,
-) -> np.ndarray:
-    tw, th = tile_extent
-    width, height = size
-    ow = min(outline_width, tw // 2, th // 2)
-    mask = np.zeros((height, width), dtype=np.uint8)
-
-    xs = tiles["x"].to_numpy()
-    ys = tiles["y"].to_numpy()
-
-    h_cols = (xs[:, None] + np.arange(tw)[None, :]).ravel()
-    v_rows = (ys[:, None] + np.arange(th)[None, :]).ravel()
-
-    h_cols_in = (h_cols >= 0) & (h_cols < width)
-    v_rows_in = (v_rows >= 0) & (v_rows < height)
-
-    for dy in range(ow):
-        for ry in (ys + dy, ys + th - ow + dy):
-            rows = np.repeat(ry, tw)
-            ok = h_cols_in & (rows >= 0) & (rows < height)
-            mask[rows[ok], h_cols[ok]] = 255
-
-    for dx in range(ow):
-        for cx in (xs + dx, xs + tw - ow + dx):
-            cols = np.repeat(cx, th)
-            ok = v_rows_in & (cols >= 0) & (cols < width)
-            mask[v_rows[ok], cols[ok]] = 255
-
-    return mask
-
-
 @ray.remote(num_cpus=1, memory=3 * 1024**3)
 def process_slide(
     item: tuple[dict[str, Any], pd.DataFrame],
@@ -62,7 +26,7 @@ def process_slide(
         y=slide_tiles["y"] // downsample,
     )
 
-    mask = _draw_tile_outlines(
+    mask = tile_mask(
         scaled_tiles,
         tile_extent=(
             slide["tile_extent_x"] // downsample,
@@ -74,13 +38,13 @@ def process_slide(
         ),
     )
 
-    height, width = mask.shape
+    width, height = mask.size
 
     write_big_tiff(
         image=pyvips.Image.new_from_memory(
             data=mask.tobytes(), width=width, height=height, bands=1, format="uchar"
         ),
-        path=Path(output_dir, f"{slide['id']}.tiff"),
+        path=Path(output_dir, f"{Path(slide['wsi_path']).stem}.tiff"),
         mpp_x=slide["mpp_x"] * downsample,
         mpp_y=slide["mpp_y"] * downsample,
     )

From 1f6f1d9abf83a27cc32e9d406abd90043a3b7779 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Tue, 28 Apr 2026 10:33:27 +0200
Subject: [PATCH 20/42] fix(tile_masks): use 'path' column for WSI filename

The tiling run's slides parquet uses 'path' (from read_slides), not
'wsi_path' which only exists upstream in wsi_mapping output.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---
 preprocessing/tile_masks.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 780d054..3348ae7 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -44,7 +44,7 @@ def process_slide(
         image=pyvips.Image.new_from_memory(
             data=mask.tobytes(), width=width, height=height, bands=1, format="uchar"
         ),
-        path=Path(output_dir, f"{Path(slide['wsi_path']).stem}.tiff"),
+        path=Path(output_dir, f"{Path(slide['path']).stem}.tiff"),
         mpp_x=slide["mpp_x"] * downsample,
         mpp_y=slide["mpp_y"] * downsample,
     )

From 979d4ead16b1387d94d7d00e63a01e69ce9ed7d9 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Tue, 28 Apr 2026 10:36:57 +0200
Subject: [PATCH 21/42] fix(tile_masks): ceil canvas size to avoid clipping
 edge tiles

Floor on extent_x/y could shrink the canvas by up to downsample-1 px,
clipping the rightmost/bottommost tile outline. Use ceil for size while
keeping floor for tile_extent and tile coordinates so outlines still tile
cleanly at the scaled stride.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---
 preprocessing/tile_masks.py | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 3348ae7..e58c545 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -1,3 +1,4 @@
+from math import ceil
 from pathlib import Path
 from tempfile import TemporaryDirectory
 from typing import Any
@@ -33,8 +34,8 @@ def process_slide(
             slide["tile_extent_y"] // downsample,
         ),
         size=(
-            slide["extent_x"] // downsample,
-            slide["extent_y"] // downsample,
+            ceil(slide["extent_x"] / downsample),
+            ceil(slide["extent_y"] / downsample),
         ),
     )
 

From 97f3951b0ccc878958a79a0e2f505377ce425224 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 29 Apr 2026 15:17:21 +0200
Subject: [PATCH 22/42] fix: use stride as the box size

---
 preprocessing/tile_masks.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index e58c545..2f70c38 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -30,8 +30,8 @@ def process_slide(
     mask = tile_mask(
         scaled_tiles,
         tile_extent=(
-            slide["tile_extent_x"] // downsample,
-            slide["tile_extent_y"] // downsample,
+            slide["stride_x"] // downsample,
+            slide["stride_y"] // downsample,
         ),
         size=(
             ceil(slide["extent_x"] / downsample),

From a302ed74b8356d7d1a57d4d5a4317621b5c87873 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 29 Apr 2026 15:29:19 +0200
Subject: [PATCH 23/42] fix: tile masks allignment

---
 preprocessing/tile_masks.py | 25 +++++++------------------
 1 file changed, 7 insertions(+), 18 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 2f70c38..75cf291 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -1,4 +1,3 @@
-from math import ceil
 from pathlib import Path
 from tempfile import TemporaryDirectory
 from typing import Any
@@ -22,29 +21,19 @@ def process_slide(
 ) -> None:
     slide, slide_tiles = item
 
-    scaled_tiles = slide_tiles.assign(
-        x=slide_tiles["x"] // downsample,
-        y=slide_tiles["y"] // downsample,
-    )
-
     mask = tile_mask(
-        scaled_tiles,
-        tile_extent=(
-            slide["stride_x"] // downsample,
-            slide["stride_y"] // downsample,
-        ),
-        size=(
-            ceil(slide["extent_x"] / downsample),
-            ceil(slide["extent_y"] / downsample),
-        ),
+        slide_tiles,
+        tile_extent=(slide["tile_extent_x"], slide["tile_extent_y"]),
+        size=(slide["extent_x"], slide["extent_y"]),
     )
 
     width, height = mask.size
+    vips_image = pyvips.Image.new_from_memory(
+        data=mask.tobytes(), width=width, height=height, bands=1, format="uchar"
+    ).resize(1.0 / downsample)
 
     write_big_tiff(
-        image=pyvips.Image.new_from_memory(
-            data=mask.tobytes(), width=width, height=height, bands=1, format="uchar"
-        ),
+        image=vips_image,
         path=Path(output_dir, f"{Path(slide['path']).stem}.tiff"),
         mpp_x=slide["mpp_x"] * downsample,
         mpp_y=slide["mpp_y"] * downsample,

From a0f1011f853cf68ae39c089afc888d9f39d61d37 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 29 Apr 2026 15:38:56 +0200
Subject: [PATCH 24/42] feat: use ScalarMaskBuilder

---
 preprocessing/tile_masks.py | 39 +++++++++++++++++--------------------
 1 file changed, 18 insertions(+), 21 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 75cf291..4eba452 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -5,10 +5,11 @@
 import hydra
 import mlflow
 import pandas as pd
-import pyvips
 import ray
+import torch
 from omegaconf import DictConfig
-from rationai.masks import process_items, tile_mask, write_big_tiff
+from rationai.masks import process_items
+from rationai.masks.mask_builders import ScalarMaskBuilder
 from rationai.mlkit import autolog, with_cli_args
 from rationai.mlkit.lightning.loggers import MLFlowLogger
 
@@ -17,27 +18,26 @@
 def process_slide(
     item: tuple[dict[str, Any], pd.DataFrame],
     output_dir: str,
-    downsample: int,
 ) -> None:
     slide, slide_tiles = item
 
-    mask = tile_mask(
-        slide_tiles,
-        tile_extent=(slide["tile_extent_x"], slide["tile_extent_y"]),
-        size=(slide["extent_x"], slide["extent_y"]),
+    builder = ScalarMaskBuilder(
+        save_dir=output_dir,
+        filename=Path(slide["path"]).stem,
+        extent_x=slide["extent_x"],
+        extent_y=slide["extent_y"],
+        mpp_x=slide["mpp_x"],
+        mpp_y=slide["mpp_y"],
+        extent_tile=slide["tile_extent_x"],
+        stride=slide["stride_x"],
     )
 
-    width, height = mask.size
-    vips_image = pyvips.Image.new_from_memory(
-        data=mask.tobytes(), width=width, height=height, bands=1, format="uchar"
-    ).resize(1.0 / downsample)
+    xs = torch.tensor(slide_tiles["x"].values)
+    ys = torch.tensor(slide_tiles["y"].values)
+    data = torch.ones(len(slide_tiles), 1)
 
-    write_big_tiff(
-        image=vips_image,
-        path=Path(output_dir, f"{Path(slide['path']).stem}.tiff"),
-        mpp_x=slide["mpp_x"] * downsample,
-        mpp_y=slide["mpp_y"] * downsample,
-    )
+    builder.update(data, xs, ys)
+    builder.save()
 
 
 @with_cli_args(["+preprocessing=tile_masks"])
@@ -77,10 +77,7 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
         process_items(
             items,
             process_item=process_slide,
-            fn_kwargs={
-                "output_dir": output_dir,
-                "downsample": config.downsample,
-            },
+            fn_kwargs={"output_dir": output_dir},
             max_concurrent=config.max_concurrent,
         )
 

From 1884819090c63bc3d2a51a49427e0aee48e4d86f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 29 Apr 2026 15:50:00 +0200
Subject: [PATCH 25/42] feat: compute also tile coverage with heatmap

---
 preprocessing/tile_masks.py | 56 +++++++++++++++++++++++++------------
 1 file changed, 38 insertions(+), 18 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 4eba452..b00d0c4 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -4,11 +4,13 @@
 
 import hydra
 import mlflow
+import numpy as np
 import pandas as pd
+import pyvips
 import ray
 import torch
 from omegaconf import DictConfig
-from rationai.masks import process_items
+from rationai.masks import process_items, tile_mask, write_big_tiff
 from rationai.masks.mask_builders import ScalarMaskBuilder
 from rationai.mlkit import autolog, with_cli_args
 from rationai.mlkit.lightning.loggers import MLFlowLogger
@@ -18,33 +20,48 @@
 def process_slide(
     item: tuple[dict[str, Any], pd.DataFrame],
     output_dir: str,
+    tile_percentage_cols: list[str],
 ) -> None:
     slide, slide_tiles = item
+    filename = f"{Path(slide['path']).stem}.tiff"
 
-    builder = ScalarMaskBuilder(
-        save_dir=output_dir,
-        filename=Path(slide["path"]).stem,
-        extent_x=slide["extent_x"],
-        extent_y=slide["extent_y"],
+    for percentage_col in tile_percentage_cols:
+        builder = ScalarMaskBuilder(
+            save_dir=Path(output_dir, percentage_col),
+            filename=filename,
+            extent_x=slide["extent_x"],
+            extent_y=slide["extent_y"],
+            mpp_x=slide["mpp_x"],
+            mpp_y=slide["mpp_y"],
+            extent_tile=slide["tile_extent_x"],
+            stride=slide["stride_x"],
+        )
+        builder.update(
+            data=torch.tensor(slide_tiles[percentage_col].values).unsqueeze(1),
+            xs=torch.tensor(slide_tiles["x"].values),
+            ys=torch.tensor(slide_tiles["y"].values),
+        )
+        builder.save()
+
+    mask = tile_mask(
+        slide_tiles,
+        tile_extent=(slide["tile_extent_x"], slide["tile_extent_y"]),
+        size=(slide["extent_x"], slide["extent_y"]),
+    )
+    write_big_tiff(
+        pyvips.Image.new_from_array(np.array(mask)),
+        Path(output_dir, "outlines", filename),
         mpp_x=slide["mpp_x"],
         mpp_y=slide["mpp_y"],
-        extent_tile=slide["tile_extent_x"],
-        stride=slide["stride_x"],
     )
 
-    xs = torch.tensor(slide_tiles["x"].values)
-    ys = torch.tensor(slide_tiles["y"].values)
-    data = torch.ones(len(slide_tiles), 1)
-
-    builder.update(data, xs, ys)
-    builder.save()
-
 
 @with_cli_args(["+preprocessing=tile_masks"])
 @hydra.main(config_path="../configs", config_name="preprocessing", version_base=None)
 @autolog
 def main(config: DictConfig, logger: MLFlowLogger) -> None:
     tiling_run_id = config.dataset.mlflow_artifacts.tiling_run_id
+    tile_percentage_cols: list[str] = list(config.tile_percentage_cols)
 
     slides = pd.read_parquet(
         mlflow.artifacts.download_artifacts(
@@ -57,7 +74,7 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
             run_id=tiling_run_id,
             artifact_path=config.tiles_artifact_path,
         ),
-        columns=["slide_id", "x", "y"],
+        columns=["slide_id", "x", "y", *tile_percentage_cols],
     )
 
     tiles_by_slide = {
@@ -68,7 +85,7 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
     items: list[tuple[dict[str, Any], pd.DataFrame]] = [
         (
             {str(k): v for k, v in slide.to_dict().items()},
-            tiles_by_slide.get(slide["id"], pd.DataFrame(columns=["x", "y"])),
+            tiles_by_slide.get(slide["id"], pd.DataFrame(columns=["x", "y", *tile_percentage_cols])),
         )
         for _, slide in slides.iterrows()
     ]
@@ -77,7 +94,10 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
         process_items(
             items,
             process_item=process_slide,
-            fn_kwargs={"output_dir": output_dir},
+            fn_kwargs={
+                "output_dir": output_dir,
+                "tile_percentage_cols": tile_percentage_cols,
+            },
             max_concurrent=config.max_concurrent,
         )
 

From 760e13a314762dca4d73b58b3adaeeac57bd83d6 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 29 Apr 2026 15:55:54 +0200
Subject: [PATCH 26/42] feat(tile_masks): add per-class coverage heatmaps and
 configure submit script

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 configs/experiment/preprocessing/tile_masks_05mpp.yaml | 8 ++++++++
 configs/preprocessing/tile_masks.yaml                  | 6 ++----
 preprocessing/tile_masks.py                            | 4 +++-
 scripts/submit_tile_masks.py                           | 6 +++---
 4 files changed, 16 insertions(+), 8 deletions(-)

diff --git a/configs/experiment/preprocessing/tile_masks_05mpp.yaml b/configs/experiment/preprocessing/tile_masks_05mpp.yaml
index fee1fa7..8cfc5a9 100644
--- a/configs/experiment/preprocessing/tile_masks_05mpp.yaml
+++ b/configs/experiment/preprocessing/tile_masks_05mpp.yaml
@@ -6,3 +6,11 @@ defaults:
 
 slides_artifact_path: train_split/slides.parquet
 tiles_artifact_path: train_split/tiles.parquet
+tile_percentage_cols:
+  - tile_coverage_Nerve
+  - tile_coverage_Blood
+  - tile_coverage_Connective-Tissue
+  - tile_coverage_Fat
+  - tile_coverage_Epithelium
+  - tile_coverage_Muscle
+  - tile_coverage_Other
diff --git a/configs/preprocessing/tile_masks.yaml b/configs/preprocessing/tile_masks.yaml
index 52f5f36..10c1273 100644
--- a/configs/preprocessing/tile_masks.yaml
+++ b/configs/preprocessing/tile_masks.yaml
@@ -2,13 +2,11 @@
 
 slides_artifact_path: ???
 tiles_artifact_path: ???
-downsample: 8
+tile_percentage_cols: ???
 
 max_concurrent: 8
 mlflow_artifact_path: tile_masks
 
 metadata:
   run_name: Tile Masks ${dataset.name}
-  description: "Tile mask generation (downsample ${downsample}x)"
-  hyperparams:
-    downsample: ${downsample}
+  description: Tile mask generation
diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index b00d0c4..888d8c0 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -85,7 +85,9 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
     items: list[tuple[dict[str, Any], pd.DataFrame]] = [
         (
             {str(k): v for k, v in slide.to_dict().items()},
-            tiles_by_slide.get(slide["id"], pd.DataFrame(columns=["x", "y", *tile_percentage_cols])),
+            tiles_by_slide.get(
+                slide["id"], pd.DataFrame(columns=["x", "y", *tile_percentage_cols])
+            ),
         )
         for _, slide in slides.iterrows()
     ]
diff --git a/scripts/submit_tile_masks.py b/scripts/submit_tile_masks.py
index bb57a59..fc290ca 100644
--- a/scripts/submit_tile_masks.py
+++ b/scripts/submit_tile_masks.py
@@ -3,16 +3,16 @@
 
 submit_job(
     job_name="tissue-classification-tile-masks",
-    username=...,
+    username="vcifka",
     cpu=8,
     memory="32Gi",
     gpu=None,
     public=False,
     script=[
-        "git clone https://github.com/RationAI/tissue-classification.git workdir",
+        "git clone --branch feature/implement-tile-masks https://github.com/RationAI/tissue-classification.git workdir",
         "cd workdir",
         "uv sync",
-        "uv run python -m preprocessing.tile_masks +experiment=...",
+        "uv run python -m preprocessing.tile_masks +experiment=preprocessing/tile_masks_05mpp",
     ],
     storage=[storage.secure.DATA],
 )

From 9cff04d121a475d0dee24cbcc27bcd1134b70c3f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 29 Apr 2026 20:01:23 +0200
Subject: [PATCH 27/42] fix: remove heatmap generation

---
 configs/preprocessing/tile_masks.yaml |  1 -
 preprocessing/tile_masks.py           | 27 ++-------------------------
 2 files changed, 2 insertions(+), 26 deletions(-)

diff --git a/configs/preprocessing/tile_masks.yaml b/configs/preprocessing/tile_masks.yaml
index 10c1273..2480b95 100644
--- a/configs/preprocessing/tile_masks.yaml
+++ b/configs/preprocessing/tile_masks.yaml
@@ -2,7 +2,6 @@
 
 slides_artifact_path: ???
 tiles_artifact_path: ???
-tile_percentage_cols: ???
 
 max_concurrent: 8
 mlflow_artifact_path: tile_masks
diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 888d8c0..f39469d 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -8,10 +8,8 @@
 import pandas as pd
 import pyvips
 import ray
-import torch
 from omegaconf import DictConfig
 from rationai.masks import process_items, tile_mask, write_big_tiff
-from rationai.masks.mask_builders import ScalarMaskBuilder
 from rationai.mlkit import autolog, with_cli_args
 from rationai.mlkit.lightning.loggers import MLFlowLogger
 
@@ -20,29 +18,10 @@
 def process_slide(
     item: tuple[dict[str, Any], pd.DataFrame],
     output_dir: str,
-    tile_percentage_cols: list[str],
 ) -> None:
     slide, slide_tiles = item
     filename = f"{Path(slide['path']).stem}.tiff"
 
-    for percentage_col in tile_percentage_cols:
-        builder = ScalarMaskBuilder(
-            save_dir=Path(output_dir, percentage_col),
-            filename=filename,
-            extent_x=slide["extent_x"],
-            extent_y=slide["extent_y"],
-            mpp_x=slide["mpp_x"],
-            mpp_y=slide["mpp_y"],
-            extent_tile=slide["tile_extent_x"],
-            stride=slide["stride_x"],
-        )
-        builder.update(
-            data=torch.tensor(slide_tiles[percentage_col].values).unsqueeze(1),
-            xs=torch.tensor(slide_tiles["x"].values),
-            ys=torch.tensor(slide_tiles["y"].values),
-        )
-        builder.save()
-
     mask = tile_mask(
         slide_tiles,
         tile_extent=(slide["tile_extent_x"], slide["tile_extent_y"]),
@@ -61,7 +40,6 @@ def process_slide(
 @autolog
 def main(config: DictConfig, logger: MLFlowLogger) -> None:
     tiling_run_id = config.dataset.mlflow_artifacts.tiling_run_id
-    tile_percentage_cols: list[str] = list(config.tile_percentage_cols)
 
     slides = pd.read_parquet(
         mlflow.artifacts.download_artifacts(
@@ -74,7 +52,7 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
             run_id=tiling_run_id,
             artifact_path=config.tiles_artifact_path,
         ),
-        columns=["slide_id", "x", "y", *tile_percentage_cols],
+        columns=["slide_id", "x", "y"],
     )
 
     tiles_by_slide = {
@@ -86,7 +64,7 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
         (
             {str(k): v for k, v in slide.to_dict().items()},
             tiles_by_slide.get(
-                slide["id"], pd.DataFrame(columns=["x", "y", *tile_percentage_cols])
+                slide["id"], pd.DataFrame(columns=["x", "y"])
             ),
         )
         for _, slide in slides.iterrows()
@@ -98,7 +76,6 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
             process_item=process_slide,
             fn_kwargs={
                 "output_dir": output_dir,
-                "tile_percentage_cols": tile_percentage_cols,
             },
             max_concurrent=config.max_concurrent,
         )

From c35e29d6f2772eda20065b1f351a85031a500bcf Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 29 Apr 2026 20:08:17 +0200
Subject: [PATCH 28/42] fix: downsample the images

---
 preprocessing/tile_masks.py | 12 +++++++-----
 1 file changed, 7 insertions(+), 5 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index f39469d..b558729 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -1,3 +1,4 @@
+from math import gcd
 from pathlib import Path
 from tempfile import TemporaryDirectory
 from typing import Any
@@ -22,16 +23,17 @@ def process_slide(
     slide, slide_tiles = item
     filename = f"{Path(slide['path']).stem}.tiff"
 
+    d = gcd(int(slide["stride_x"]), int(slide["tile_extent_x"]))
     mask = tile_mask(
-        slide_tiles,
-        tile_extent=(slide["tile_extent_x"], slide["tile_extent_y"]),
-        size=(slide["extent_x"], slide["extent_y"]),
+        slide_tiles.assign(x=slide_tiles["x"] // d, y=slide_tiles["y"] // d),
+        tile_extent=(slide["tile_extent_x"] // d, slide["tile_extent_y"] // d),
+        size=(slide["extent_x"] // d, slide["extent_y"] // d),
     )
     write_big_tiff(
         pyvips.Image.new_from_array(np.array(mask)),
         Path(output_dir, "outlines", filename),
-        mpp_x=slide["mpp_x"],
-        mpp_y=slide["mpp_y"],
+        mpp_x=slide["mpp_x"] * d,
+        mpp_y=slide["mpp_y"] * d,
     )
 
 

From 3984431ae1b96a74e285b3ee9579e08fbb9eb86d Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 29 Apr 2026 20:11:41 +0200
Subject: [PATCH 29/42] fix: remove directory

---
 preprocessing/tile_masks.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index b558729..363c248 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -73,6 +73,7 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
     ]
 
     with TemporaryDirectory() as output_dir:
+        Path(output_dir, "outlines").mkdir()
         process_items(
             items,
             process_item=process_slide,

From 98f435b145c2d20ceb79661bf9742cb98775d298 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 29 Apr 2026 20:21:07 +0200
Subject: [PATCH 30/42] fix: remove downsample and free memory bounds

---
 preprocessing/tile_masks.py | 14 ++++++--------
 1 file changed, 6 insertions(+), 8 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 363c248..bdc3fb4 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -1,4 +1,3 @@
-from math import gcd
 from pathlib import Path
 from tempfile import TemporaryDirectory
 from typing import Any
@@ -15,7 +14,7 @@
 from rationai.mlkit.lightning.loggers import MLFlowLogger
 
 
-@ray.remote(num_cpus=1, memory=3 * 1024**3)
+@ray.remote
 def process_slide(
     item: tuple[dict[str, Any], pd.DataFrame],
     output_dir: str,
@@ -23,17 +22,16 @@ def process_slide(
     slide, slide_tiles = item
     filename = f"{Path(slide['path']).stem}.tiff"
 
-    d = gcd(int(slide["stride_x"]), int(slide["tile_extent_x"]))
     mask = tile_mask(
-        slide_tiles.assign(x=slide_tiles["x"] // d, y=slide_tiles["y"] // d),
-        tile_extent=(slide["tile_extent_x"] // d, slide["tile_extent_y"] // d),
-        size=(slide["extent_x"] // d, slide["extent_y"] // d),
+        slide_tiles,
+        tile_extent=(slide["tile_extent_x"], slide["tile_extent_y"]),
+        size=(slide["extent_x"], slide["extent_y"]),
     )
     write_big_tiff(
         pyvips.Image.new_from_array(np.array(mask)),
         Path(output_dir, "outlines", filename),
-        mpp_x=slide["mpp_x"] * d,
-        mpp_y=slide["mpp_y"] * d,
+        mpp_x=slide["mpp_x"],
+        mpp_y=slide["mpp_y"],
     )
 
 

From 64501fc328a66b9a013d15cbb1714ed414936966 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 29 Apr 2026 20:25:46 +0200
Subject: [PATCH 31/42] feat: use numpy

---
 preprocessing/tile_masks.py | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index bdc3fb4..c3ed501 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -4,12 +4,11 @@
 
 import hydra
 import mlflow
-import numpy as np
 import pandas as pd
 import pyvips
 import ray
 from omegaconf import DictConfig
-from rationai.masks import process_items, tile_mask, write_big_tiff
+from rationai.masks import process_items, write_big_tiff
 from rationai.mlkit import autolog, with_cli_args
 from rationai.mlkit.lightning.loggers import MLFlowLogger
 
@@ -22,13 +21,13 @@ def process_slide(
     slide, slide_tiles = item
     filename = f"{Path(slide['path']).stem}.tiff"
 
-    mask = tile_mask(
-        slide_tiles,
-        tile_extent=(slide["tile_extent_x"], slide["tile_extent_y"]),
-        size=(slide["extent_x"], slide["extent_y"]),
-    )
+    img = pyvips.Image.black(int(slide["extent_x"]), int(slide["extent_y"]))
+    te_x = int(slide["tile_extent_x"])
+    te_y = int(slide["tile_extent_y"])
+    for _, row in slide_tiles.iterrows():
+        img.draw_rect([255], int(row["x"]), int(row["y"]), te_x, te_y, fill=False)
     write_big_tiff(
-        pyvips.Image.new_from_array(np.array(mask)),
+        img,
         Path(output_dir, "outlines", filename),
         mpp_x=slide["mpp_x"],
         mpp_y=slide["mpp_y"],

From 67f0b2d201d7436330420f226d120316d317c2eb Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 29 Apr 2026 20:32:02 +0200
Subject: [PATCH 32/42] fix: lower concurrency

---
 configs/preprocessing/tile_masks.yaml |  2 +-
 preprocessing/tile_masks.py           | 15 ++++++++-------
 2 files changed, 9 insertions(+), 8 deletions(-)

diff --git a/configs/preprocessing/tile_masks.yaml b/configs/preprocessing/tile_masks.yaml
index 2480b95..c3849d2 100644
--- a/configs/preprocessing/tile_masks.yaml
+++ b/configs/preprocessing/tile_masks.yaml
@@ -3,7 +3,7 @@
 slides_artifact_path: ???
 tiles_artifact_path: ???
 
-max_concurrent: 8
+max_concurrent: 1
 mlflow_artifact_path: tile_masks
 
 metadata:
diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index c3ed501..bdc3fb4 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -4,11 +4,12 @@
 
 import hydra
 import mlflow
+import numpy as np
 import pandas as pd
 import pyvips
 import ray
 from omegaconf import DictConfig
-from rationai.masks import process_items, write_big_tiff
+from rationai.masks import process_items, tile_mask, write_big_tiff
 from rationai.mlkit import autolog, with_cli_args
 from rationai.mlkit.lightning.loggers import MLFlowLogger
 
@@ -21,13 +22,13 @@ def process_slide(
     slide, slide_tiles = item
     filename = f"{Path(slide['path']).stem}.tiff"
 
-    img = pyvips.Image.black(int(slide["extent_x"]), int(slide["extent_y"]))
-    te_x = int(slide["tile_extent_x"])
-    te_y = int(slide["tile_extent_y"])
-    for _, row in slide_tiles.iterrows():
-        img.draw_rect([255], int(row["x"]), int(row["y"]), te_x, te_y, fill=False)
+    mask = tile_mask(
+        slide_tiles,
+        tile_extent=(slide["tile_extent_x"], slide["tile_extent_y"]),
+        size=(slide["extent_x"], slide["extent_y"]),
+    )
     write_big_tiff(
-        img,
+        pyvips.Image.new_from_array(np.array(mask)),
         Path(output_dir, "outlines", filename),
         mpp_x=slide["mpp_x"],
         mpp_y=slide["mpp_y"],

From fa9c8eca70a46078ca996314ae47239698367631 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 29 Apr 2026 20:44:33 +0200
Subject: [PATCH 33/42] feat: bring back ray

---
 preprocessing/tile_masks.py  | 4 +---
 scripts/submit_tile_masks.py | 8 ++++----
 2 files changed, 5 insertions(+), 7 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index bdc3fb4..70440c6 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -63,9 +63,7 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
     items: list[tuple[dict[str, Any], pd.DataFrame]] = [
         (
             {str(k): v for k, v in slide.to_dict().items()},
-            tiles_by_slide.get(
-                slide["id"], pd.DataFrame(columns=["x", "y"])
-            ),
+            tiles_by_slide.get(slide["id"], pd.DataFrame(columns=["x", "y"])),
         )
         for _, slide in slides.iterrows()
     ]
diff --git a/scripts/submit_tile_masks.py b/scripts/submit_tile_masks.py
index fc290ca..7fc8a94 100644
--- a/scripts/submit_tile_masks.py
+++ b/scripts/submit_tile_masks.py
@@ -3,16 +3,16 @@
 
 submit_job(
     job_name="tissue-classification-tile-masks",
-    username="vcifka",
+    username=...,
     cpu=8,
-    memory="32Gi",
+    memory="64Gi",
     gpu=None,
     public=False,
     script=[
-        "git clone --branch feature/implement-tile-masks https://github.com/RationAI/tissue-classification.git workdir",
+        "git clone https://github.com/RationAI/tissue-classification.git workdir",
         "cd workdir",
         "uv sync",
-        "uv run python -m preprocessing.tile_masks +experiment=preprocessing/tile_masks_05mpp",
+        "uv run python -m preprocessing.tile_masks +experiment=...",
     ],
     storage=[storage.secure.DATA],
 )

From 4e243250edc13c6b29e081010be98df39a983797 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 29 Apr 2026 22:22:06 +0200
Subject: [PATCH 34/42] refactor: drop ray, add heatmaps

---
 configs/preprocessing/tile_masks.yaml |  2 +-
 preprocessing/tile_masks.py           | 45 ++++++++++++++++++---------
 2 files changed, 31 insertions(+), 16 deletions(-)

diff --git a/configs/preprocessing/tile_masks.yaml b/configs/preprocessing/tile_masks.yaml
index c3849d2..21dbf3f 100644
--- a/configs/preprocessing/tile_masks.yaml
+++ b/configs/preprocessing/tile_masks.yaml
@@ -2,8 +2,8 @@
 
 slides_artifact_path: ???
 tiles_artifact_path: ???
+tile_percentage_cols: ???
 
-max_concurrent: 1
 mlflow_artifact_path: tile_masks
 
 metadata:
diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 70440c6..4013cae 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -7,21 +7,41 @@
 import numpy as np
 import pandas as pd
 import pyvips
-import ray
+import torch
 from omegaconf import DictConfig
-from rationai.masks import process_items, tile_mask, write_big_tiff
+from rationai.masks import tile_mask, write_big_tiff
+from rationai.masks.mask_builders import ScalarMaskBuilder
 from rationai.mlkit import autolog, with_cli_args
 from rationai.mlkit.lightning.loggers import MLFlowLogger
+from tqdm import tqdm
 
 
-@ray.remote
 def process_slide(
     item: tuple[dict[str, Any], pd.DataFrame],
     output_dir: str,
+    tile_percentage_cols: list[str],
 ) -> None:
     slide, slide_tiles = item
     filename = f"{Path(slide['path']).stem}.tiff"
 
+    for percentage_col in tile_percentage_cols:
+        builder = ScalarMaskBuilder(
+            save_dir=Path(output_dir, percentage_col),
+            filename=filename,
+            extent_x=slide["extent_x"],
+            extent_y=slide["extent_y"],
+            mpp_x=slide["mpp_x"],
+            mpp_y=slide["mpp_y"],
+            extent_tile=slide["tile_extent_x"],
+            stride=slide["stride_x"],
+        )
+        builder.update(
+            data=torch.tensor(slide_tiles[percentage_col].values).unsqueeze(1),
+            xs=torch.tensor(slide_tiles["x"].values),
+            ys=torch.tensor(slide_tiles["y"].values),
+        )
+        builder.save()
+
     mask = tile_mask(
         slide_tiles,
         tile_extent=(slide["tile_extent_x"], slide["tile_extent_y"]),
@@ -40,6 +60,7 @@ def process_slide(
 @autolog
 def main(config: DictConfig, logger: MLFlowLogger) -> None:
     tiling_run_id = config.dataset.mlflow_artifacts.tiling_run_id
+    tile_percentage_cols: list[str] = list(config.tile_percentage_cols)
 
     slides = pd.read_parquet(
         mlflow.artifacts.download_artifacts(
@@ -52,7 +73,7 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
             run_id=tiling_run_id,
             artifact_path=config.tiles_artifact_path,
         ),
-        columns=["slide_id", "x", "y"],
+        columns=["slide_id", "x", "y", *tile_percentage_cols],
     )
 
     tiles_by_slide = {
@@ -63,21 +84,17 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
     items: list[tuple[dict[str, Any], pd.DataFrame]] = [
         (
             {str(k): v for k, v in slide.to_dict().items()},
-            tiles_by_slide.get(slide["id"], pd.DataFrame(columns=["x", "y"])),
+            tiles_by_slide.get(
+                slide["id"], pd.DataFrame(columns=["x", "y", *tile_percentage_cols])
+            ),
         )
         for _, slide in slides.iterrows()
     ]
 
     with TemporaryDirectory() as output_dir:
         Path(output_dir, "outlines").mkdir()
-        process_items(
-            items,
-            process_item=process_slide,
-            fn_kwargs={
-                "output_dir": output_dir,
-            },
-            max_concurrent=config.max_concurrent,
-        )
+        for item in tqdm(items):
+            process_slide(item, output_dir, tile_percentage_cols)
 
         logger.log_artifacts(
             local_dir=output_dir, artifact_path=config.mlflow_artifact_path
@@ -85,6 +102,4 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
 
 
 if __name__ == "__main__":
-    ray.init()
     main()
-    ray.shutdown()

From 432ad82e8f1767fdc5b7ea4ec473230c8e2f974e Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Wed, 29 Apr 2026 22:33:01 +0200
Subject: [PATCH 35/42] fix: bring back ray but dump the memory

---
 configs/preprocessing/tile_masks.yaml |  1 +
 preprocessing/tile_masks.py           | 18 ++++++++++++++----
 2 files changed, 15 insertions(+), 4 deletions(-)

diff --git a/configs/preprocessing/tile_masks.yaml b/configs/preprocessing/tile_masks.yaml
index 21dbf3f..a287440 100644
--- a/configs/preprocessing/tile_masks.yaml
+++ b/configs/preprocessing/tile_masks.yaml
@@ -4,6 +4,7 @@ slides_artifact_path: ???
 tiles_artifact_path: ???
 tile_percentage_cols: ???
 
+max_concurrent: 1
 mlflow_artifact_path: tile_masks
 
 metadata:
diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index 4013cae..cc26408 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -7,15 +7,16 @@
 import numpy as np
 import pandas as pd
 import pyvips
+import ray
 import torch
 from omegaconf import DictConfig
-from rationai.masks import tile_mask, write_big_tiff
+from rationai.masks import process_items, tile_mask, write_big_tiff
 from rationai.masks.mask_builders import ScalarMaskBuilder
 from rationai.mlkit import autolog, with_cli_args
 from rationai.mlkit.lightning.loggers import MLFlowLogger
-from tqdm import tqdm
 
 
+@ray.remote(max_calls=1)
 def process_slide(
     item: tuple[dict[str, Any], pd.DataFrame],
     output_dir: str,
@@ -93,8 +94,15 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
 
     with TemporaryDirectory() as output_dir:
         Path(output_dir, "outlines").mkdir()
-        for item in tqdm(items):
-            process_slide(item, output_dir, tile_percentage_cols)
+        process_items(
+            items,
+            process_item=process_slide,
+            fn_kwargs={
+                "output_dir": output_dir,
+                "tile_percentage_cols": tile_percentage_cols,
+            },
+            max_concurrent=config.max_concurrent,
+        )
 
         logger.log_artifacts(
             local_dir=output_dir, artifact_path=config.mlflow_artifact_path
@@ -102,4 +110,6 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
 
 
 if __name__ == "__main__":
+    ray.init()
     main()
+    ray.shutdown()

From 72f12a9f1fc6881357746c8c8a35686a188e45ab Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Thu, 30 Apr 2026 11:06:20 +0200
Subject: [PATCH 36/42] fix: generate the mask in place

---
 preprocessing/tile_masks.py | 57 +++++++++++++++++++------------------
 1 file changed, 29 insertions(+), 28 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index cc26408..a3df4e6 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -4,7 +4,6 @@
 
 import hydra
 import mlflow
-import numpy as np
 import pandas as pd
 import pyvips
 import ray
@@ -18,11 +17,16 @@
 
 @ray.remote(max_calls=1)
 def process_slide(
-    item: tuple[dict[str, Any], pd.DataFrame],
+    slide: dict[str, Any],
     output_dir: str,
     tile_percentage_cols: list[str],
+    tiles_path: str,
 ) -> None:
-    slide, slide_tiles = item
+    slide_tiles = pd.read_parquet(
+        tiles_path,
+        columns=["x", "y", *tile_percentage_cols],
+        filters=[("slide_id", "==", slide["id"])],
+    )
     filename = f"{Path(slide['path']).stem}.tiff"
 
     for percentage_col in tile_percentage_cols:
@@ -48,9 +52,19 @@ def process_slide(
         tile_extent=(slide["tile_extent_x"], slide["tile_extent_y"]),
         size=(slide["extent_x"], slide["extent_y"]),
     )
+    width, height = mask.size
+    mask_bytes = mask.tobytes()
+    del mask
+
     write_big_tiff(
-        pyvips.Image.new_from_array(np.array(mask)),
-        Path(output_dir, "outlines", filename),
+        image=pyvips.Image.new_from_memory(
+            data=mask_bytes,
+            width=width,
+            height=height,
+            bands=1,
+            format="uchar",
+        ),
+        path=Path(output_dir, "outlines", filename),
         mpp_x=slide["mpp_x"],
         mpp_y=slide["mpp_y"],
     )
@@ -63,32 +77,18 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
     tiling_run_id = config.dataset.mlflow_artifacts.tiling_run_id
     tile_percentage_cols: list[str] = list(config.tile_percentage_cols)
 
-    slides = pd.read_parquet(
-        mlflow.artifacts.download_artifacts(
-            run_id=tiling_run_id,
-            artifact_path=config.slides_artifact_path,
-        )
+    slides_path = mlflow.artifacts.download_artifacts(
+        run_id=tiling_run_id,
+        artifact_path=config.slides_artifact_path,
     )
-    tiles = pd.read_parquet(
-        mlflow.artifacts.download_artifacts(
-            run_id=tiling_run_id,
-            artifact_path=config.tiles_artifact_path,
-        ),
-        columns=["slide_id", "x", "y", *tile_percentage_cols],
+    tiles_path = mlflow.artifacts.download_artifacts(
+        run_id=tiling_run_id,
+        artifact_path=config.tiles_artifact_path,
     )
+    slides = pd.read_parquet(slides_path)
 
-    tiles_by_slide = {
-        slide_id: group.drop(columns="slide_id")
-        for slide_id, group in tiles.groupby("slide_id")
-    }
-
-    items: list[tuple[dict[str, Any], pd.DataFrame]] = [
-        (
-            {str(k): v for k, v in slide.to_dict().items()},
-            tiles_by_slide.get(
-                slide["id"], pd.DataFrame(columns=["x", "y", *tile_percentage_cols])
-            ),
-        )
+    items: list[dict[str, Any]] = [
+        {str(k): v for k, v in slide.to_dict().items()}
         for _, slide in slides.iterrows()
     ]
 
@@ -100,6 +100,7 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
             fn_kwargs={
                 "output_dir": output_dir,
                 "tile_percentage_cols": tile_percentage_cols,
+                "tiles_path": tiles_path,
             },
             max_concurrent=config.max_concurrent,
         )

From 096119af8662b03c49a69e212be9a50f9e1d22ba Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Thu, 30 Apr 2026 11:33:14 +0200
Subject: [PATCH 37/42] fix: raise concurrency

---
 configs/preprocessing/tile_masks.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/configs/preprocessing/tile_masks.yaml b/configs/preprocessing/tile_masks.yaml
index a287440..10c1273 100644
--- a/configs/preprocessing/tile_masks.yaml
+++ b/configs/preprocessing/tile_masks.yaml
@@ -4,7 +4,7 @@ slides_artifact_path: ???
 tiles_artifact_path: ???
 tile_percentage_cols: ???
 
-max_concurrent: 1
+max_concurrent: 8
 mlflow_artifact_path: tile_masks
 
 metadata:

From 79c343d85ecebb48b87b4daa882393e6db7587d8 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Thu, 30 Apr 2026 11:49:09 +0200
Subject: [PATCH 38/42] fix: lower concurrency to 4

---
 configs/preprocessing/tile_masks.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/configs/preprocessing/tile_masks.yaml b/configs/preprocessing/tile_masks.yaml
index 10c1273..3f344ca 100644
--- a/configs/preprocessing/tile_masks.yaml
+++ b/configs/preprocessing/tile_masks.yaml
@@ -4,7 +4,7 @@ slides_artifact_path: ???
 tiles_artifact_path: ???
 tile_percentage_cols: ???
 
-max_concurrent: 8
+max_concurrent: 4
 mlflow_artifact_path: tile_masks
 
 metadata:

From 1eecb8434a171c65f8c03fa5eed07cfbba863887 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Thu, 30 Apr 2026 21:03:49 +0200
Subject: [PATCH 39/42] feat: lower concurrency to 1

---
 configs/preprocessing/tissue_masks.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/configs/preprocessing/tissue_masks.yaml b/configs/preprocessing/tissue_masks.yaml
index 4de9a5c..58168f7 100644
--- a/configs/preprocessing/tissue_masks.yaml
+++ b/configs/preprocessing/tissue_masks.yaml
@@ -3,7 +3,7 @@
 mpp: ???
 disk_factor: ???
 
-max_concurrent: 8
+max_concurrent: 1
 mlflow_artifact_path: tissue_masks
 
 metadata:

From 8d02a50b1681792ad64280952f6da39cfb311e36 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Sun, 3 May 2026 13:00:45 +0200
Subject: [PATCH 40/42] refactor: drop ray

---
 configs/preprocessing/tile_masks.yaml   |  2 +-
 configs/preprocessing/tissue_masks.yaml |  2 +-
 preprocessing/tile_masks.py             | 29 +++++++++----------------
 3 files changed, 12 insertions(+), 21 deletions(-)

diff --git a/configs/preprocessing/tile_masks.yaml b/configs/preprocessing/tile_masks.yaml
index 3f344ca..a287440 100644
--- a/configs/preprocessing/tile_masks.yaml
+++ b/configs/preprocessing/tile_masks.yaml
@@ -4,7 +4,7 @@ slides_artifact_path: ???
 tiles_artifact_path: ???
 tile_percentage_cols: ???
 
-max_concurrent: 4
+max_concurrent: 1
 mlflow_artifact_path: tile_masks
 
 metadata:
diff --git a/configs/preprocessing/tissue_masks.yaml b/configs/preprocessing/tissue_masks.yaml
index 58168f7..4de9a5c 100644
--- a/configs/preprocessing/tissue_masks.yaml
+++ b/configs/preprocessing/tissue_masks.yaml
@@ -3,7 +3,7 @@
 mpp: ???
 disk_factor: ???
 
-max_concurrent: 1
+max_concurrent: 8
 mlflow_artifact_path: tissue_masks
 
 metadata:
diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index a3df4e6..edd0974 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -6,16 +6,15 @@
 import mlflow
 import pandas as pd
 import pyvips
-import ray
 import torch
 from omegaconf import DictConfig
-from rationai.masks import process_items, tile_mask, write_big_tiff
+from rationai.masks import tile_mask, write_big_tiff
 from rationai.masks.mask_builders import ScalarMaskBuilder
 from rationai.mlkit import autolog, with_cli_args
 from rationai.mlkit.lightning.loggers import MLFlowLogger
+from tqdm import tqdm
 
 
-@ray.remote(max_calls=1)
 def process_slide(
     slide: dict[str, Any],
     output_dir: str,
@@ -87,23 +86,17 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
     )
     slides = pd.read_parquet(slides_path)
 
-    items: list[dict[str, Any]] = [
-        {str(k): v for k, v in slide.to_dict().items()}
-        for _, slide in slides.iterrows()
-    ]
+    items = slides.to_dict(orient="records")
 
     with TemporaryDirectory() as output_dir:
         Path(output_dir, "outlines").mkdir()
-        process_items(
-            items,
-            process_item=process_slide,
-            fn_kwargs={
-                "output_dir": output_dir,
-                "tile_percentage_cols": tile_percentage_cols,
-                "tiles_path": tiles_path,
-            },
-            max_concurrent=config.max_concurrent,
-        )
+        for slide in tqdm(items):
+            process_slide(
+                slide,
+                output_dir=output_dir,
+                tile_percentage_cols=tile_percentage_cols,
+                tiles_path=tiles_path,
+            )
 
         logger.log_artifacts(
             local_dir=output_dir, artifact_path=config.mlflow_artifact_path
@@ -111,6 +104,4 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
 
 
 if __name__ == "__main__":
-    ray.init()
     main()
-    ray.shutdown()

From 524d116f163b53af3185c3e0c785a6a609704f80 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Sun, 3 May 2026 14:47:41 +0200
Subject: [PATCH 41/42] refactor: drop Ray from tile_masks, simplify slide
 iteration

Replace Ray remote dispatch and process_items with a plain tqdm loop.
Also simplify iterrows to to_dict(orient="records").

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 preprocessing/tile_masks.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index edd0974..fcf1f1b 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -1,6 +1,6 @@
 from pathlib import Path
 from tempfile import TemporaryDirectory
-from typing import Any
+from typing import Any, cast
 
 import hydra
 import mlflow
@@ -86,7 +86,7 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
     )
     slides = pd.read_parquet(slides_path)
 
-    items = slides.to_dict(orient="records")
+    items = cast(list[dict[str, Any]], slides.to_dict(orient="records"))
 
     with TemporaryDirectory() as output_dir:
         Path(output_dir, "outlines").mkdir()

From b9c6aab7673df4a21e51ec9e2f0efb1fa6f039bc Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Vojt=C4=9Bch=20C=C3=ADfka?= <550433@mail.muni.cz>
Date: Sun, 3 May 2026 14:49:06 +0200
Subject: [PATCH 42/42] fix: lint

---
 preprocessing/tile_masks.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/preprocessing/tile_masks.py b/preprocessing/tile_masks.py
index fcf1f1b..9738f3d 100644
--- a/preprocessing/tile_masks.py
+++ b/preprocessing/tile_masks.py
@@ -86,7 +86,7 @@ def main(config: DictConfig, logger: MLFlowLogger) -> None:
     )
     slides = pd.read_parquet(slides_path)
 
-    items = cast(list[dict[str, Any]], slides.to_dict(orient="records"))
+    items = cast("list[dict[str, Any]]", slides.to_dict(orient="records"))
 
     with TemporaryDirectory() as output_dir:
         Path(output_dir, "outlines").mkdir()