| """Define evaluation splits. Each split file: splits/<name>/fold<k>.csv with columns tile_id, split. |
| |
| orion_official : the user's train/val/test_dataframe.csv (slide-disjoint) |
| random_tile : tile-level random 80/10/10 (slides shared across splits -> optimistic upper bound) |
| patient_5fold : 5-fold CV grouped by patient (CRC33_01/_02 kept together); val = 3 patients from train folds |
| batch_lobo : leave-one-IF-acquisition-batch-out (C59kX/mX/nX/qX); val = 2 patients from training batches |
| """ |
| import os |
|
|
| import numpy as np |
|
|
| from common import SPLITS, load_all_tiles |
|
|
| df = load_all_tiles() |
|
|
|
|
| def save(name, k, split): |
| os.makedirs(os.path.join(SPLITS, name), exist_ok=True) |
| out = df[["tile_id", "in_slide_name", "patient", "if_batch"]].copy() |
| out["split"] = split |
| out.to_csv(os.path.join(SPLITS, name, f"fold{k}.csv"), index=False) |
| print(name, k, out.split.value_counts().to_dict(), |
| {s: out[out.split == s].patient.nunique() for s in ["train", "val", "test"]}) |
|
|
|
|
| |
| save("orion_official", 0, df.orion_split.values) |
|
|
| |
| rng = np.random.default_rng(0) |
| u = rng.random(len(df)) |
| save("random_tile", 0, np.where(u < 0.8, "train", np.where(u < 0.9, "val", "test"))) |
|
|
| |
| pats = np.array(sorted(df.patient.unique())) |
| rng = np.random.default_rng(0) |
| perm = rng.permutation(pats) |
| folds = np.array_split(perm, 5) |
| for k in range(5): |
| test_p = set(folds[k]) |
| rest = [p for p in perm if p not in test_p] |
| val_p = set(rest[:3]) |
| split = np.where(df.patient.isin(test_p), "test", np.where(df.patient.isin(val_p), "val", "train")) |
| save("patient_5fold", k, split) |
|
|
| |
| for k, b in enumerate(sorted(df.if_batch.unique())): |
| rest = sorted(df[df.if_batch != b].patient.unique()) |
| val_p = set(np.random.default_rng(k).choice(rest, 2, replace=False)) |
| split = np.where(df.if_batch == b, "test", np.where(df.patient.isin(val_p), "val", "train")) |
| save("batch_lobo", k, split) |
|
|
