check_split.py

"""Diagnostic: class distribution in the official 152 train / 152 val crops."""
import sys, os
sys.path.insert(0, os.path.dirname(__file__))

import numpy as np
from collections import Counter
import config as C
from dataset import train_patch_ids, val_patch_ids, get_meta

PASTIS_CLASSES = C.PASTIS_CLASSES   # single source of truth — see config.py


def crop_targets(patch_ids):
    """Yield the (i, j) label crops the finetune loop actually trains on."""
    s = C.PATCH_SIZE // C.PRETRAIN_CROP_GRID
    for pid in patch_ids:                     # duplicates counted twice
        tgt = np.load(C.ANNOT_DIR / f"TARGET_{pid}.npy")[0]
        for (i, j) in C.FT_CROP_IJ:
            yield tgt[i * s:(i + 1) * s, j * s:(j + 1) * s]


def main():
    meta = get_meta()
    tr, va = train_patch_ids(), val_patch_ids()
    print(f"Train: {len(tr)} patches x {len(C.FT_CROP_IJ)} crops = "
          f"{len(tr) * len(C.FT_CROP_IJ)} crops   (fold "
          f"{sorted({meta[p]['fold'] for p in tr})})")
    print(f"Val:   {len(va)} patches x {len(C.FT_CROP_IJ)} crops = "
          f"{len(va) * len(C.FT_CROP_IJ)} crops   (fold "
          f"{sorted({meta[p]['fold'] for p in va})})")
    print(f"Crop positions: {C.FT_CROP_IJ}\n")

    stats = {}
    for split_name, ids in [("TRAIN", tr), ("VAL", va)]:
        px = Counter()
        ncrop = Counter()
        total = 0
        for sem in crop_targets(ids):
            total += sem.size
            for c in np.unique(sem):
                if 0 < c < 19:
                    ncrop[int(c)] += 1
            for c in range(20):
                px[c] += int((sem == c).sum())
        stats[split_name] = (px, ncrop, total)

        print(f"{'='*66}")
        print(f" {split_name} — {len(ids) * len(C.FT_CROP_IJ)} crops, "
              f"{total:,} pixels")
        print(f"{'='*66}")
        print(f"{'Cls':>4} {'Name':<28} {'Pixels':>10} {'%':>7} {'Crops':>6} Status")
        print("-" * 66)
        for c in range(20):
            n = px.get(c, 0)
            pct = 100.0 * n / total if total else 0.0
            if c in (0, 19):
                status = "ignore"
            elif n == 0:
                status = "!! DEAD"
            elif n < 500:
                status = "! RARE"
            elif n < 2000:
                status = "~ sparse"
            else:
                status = "ok"
            print(f"{c:>4} {PASTIS_CLASSES[c]:<28} {n:>10,} {pct:>6.2f}% "
                  f"{ncrop.get(c, 0):>6} {status}")
        print()

    # train/val agreement on the scored classes
    trp, _, trt = stats["TRAIN"]
    vap, _, vat = stats["VAL"]
    print("=" * 66)
    print(" Train vs val share per scored class (ratio of pixel %)")
    print("=" * 66)
    print(f"{'Cls':>4} {'Name':<28} {'train %':>9} {'val %':>9} {'ratio':>8}")
    print("-" * 66)
    for c in range(1, 19):
        a = 100.0 * trp[c] / trt if trt else 0.0
        b = 100.0 * vap[c] / vat if vat else 0.0
        r = (a / b) if b > 0 else float("inf")
        flag = "  <-- skewed" if (r > 2 or r < 0.5) else ""
        print(f"{c:>4} {PASTIS_CLASSES[c]:<28} {a:>8.2f}% {b:>8.2f}% "
              f"{r:>8.2f}{flag}")


if __name__ == "__main__":
    main()