Update evaluation scripts

.gitignore

@@ -197,8 +197,9 @@ iterative_prompting_results/
 *MACOSX
 hela_ctc
 clf-test-data
-*.h5
 .claude/
+*.pdf
+*.h5
 .codex
 
 # Related to i2k workshop folders.

finetuning/v2/evaluation/baselines_common.py

@@ -0,0 +1,69 @@
+"""Shared data-loading utilities for evaluate_automatic_baselines and evaluate_interactive_baselines."""
+
+import os
+
+import numpy as np
+import imageio.v3 as imageio
+from skimage.measure import label as connected_components
+
+from elf.io import open_file
+from torch_em.transform.raw import normalize
+
+from common import get_data_paths, load_volume, _center_crop_roi
+
+CROP_SHAPE_2D = (512, 512)
+CROP_SHAPE_3D = (8, 512, 512)
+MAX_EVALUATION_SAMPLES = int(os.environ.get("MICRO_SAM_EVAL_MAX_SAMPLES", "200"))
+
+
+def _ensure_8bit_range(raw):
+    raw = raw.astype("float32", copy=False)
+    if raw.size == 0:
+        return raw
+    if raw.max() <= 1:
+        raw = raw * 255
+    elif raw.max() > 255 or raw.min() < 0:
+        raw = normalize(raw) * 255
+    return np.clip(raw, 0, 255).astype("float32")
+
+
+def _read_2d(path, key):
+    """Read a 2D array from an image file or from an H5/zarr file using key."""
+    if key is not None:
+        arr = open_file(path, mode="r")[key][:]
+    else:
+        arr = np.asarray(imageio.imread(path))
+    # Transpose channel-first (C, H, W) to channel-last (H, W, C).
+    if arr.ndim == 3 and arr.shape[0] <= 4 and arr.shape[1] > arr.shape[0] and arr.shape[2] > arr.shape[0]:
+        arr = arr.transpose(1, 2, 0)
+    return arr
+
+
+def _sorted_path_pairs(raw_paths, label_paths):
+    return sorted(zip(raw_paths, label_paths), key=lambda pair: (str(pair[0]), str(pair[1])))
+
+
+def _load_data(dataset_name, data_root, ndim):
+    """Yield (image_or_volume, labels, valid_roi) triples for the given dataset.
+
+    valid_roi is a boolean mask (True = annotated) for partially annotated datasets
+    (platynereis_nuclei), or None for all others.
+    """
+    if ndim == 3:
+        raw_paths, label_paths, raw_key, label_key = get_data_paths(dataset_name, data_root)
+        path_pairs = _sorted_path_pairs(raw_paths, label_paths)[:MAX_EVALUATION_SAMPLES]
+        for raw_path, label_path in path_pairs:
+            raw, labels, valid_roi = load_volume(raw_path, label_path, raw_key, label_key, dataset_name, CROP_SHAPE_3D)
+            raw = _ensure_8bit_range(raw)
+            yield raw, labels, valid_roi
+    else:
+        image_paths, gt_paths, raw_key, label_key = get_data_paths(dataset_name, data_root)
+        path_pairs = _sorted_path_pairs(image_paths, gt_paths)[:MAX_EVALUATION_SAMPLES]
+        for img_path, gt_path in path_pairs:
+            image = _read_2d(img_path, raw_key)
+            image = _ensure_8bit_range(image)
+            roi = _center_crop_roi(image.shape[:2], CROP_SHAPE_2D)
+            image = image[roi]
+            gt = _read_2d(gt_path, label_key)
+            gt = connected_components(gt[roi]).astype("uint32")
+            yield image, gt, None