Upload folder using huggingface_hub

820fab0 verified 4 months ago

22 kB

"""
Evaluation script for RockGlacierNet — baseline 128x128 full-image model.
Each test GeoTIFF is normalized and padded/cropped to 128x128 exactly like
training, then predicted. Metrics are computed on the VALID (non-padded) region
of each image so zero-padding does not inflate the scores.
Outputs (outputs/predictions/baseline_model_output/)
-----------------------------------------------------
    per_image_metrics.csv      — ONE ROW PER TEST IMAGE, sorted worst IoU first
                                  (this is the file to inspect for problem images)
    evaluation_metrics.csv     — global threshold sweep 0.30–0.70
    confusion_matrix.csv       — aggregate CM at best IoU threshold
    evaluation_summary.txt     — human-readable summary
    pred_*.png                 — RGB / GT / prediction comparison figures
"""

import os
import glob
import json
import numpy as np
np.Inf = np.inf
import pandas as pd
import rasterio
import tensorflow as tf
from sklearn.metrics import confusion_matrix

from data_generator import GLOBAL_MEAN, GLOBAL_STD
from models.swin_unet import build_model
from models.attention_unet import build_attention_unet
from models.unet import build_unet
from models.segformer import build_segformer

# ─────────────────────────────────────────────────────────────────────────────
# Config
# ─────────────────────────────────────────────────────────────────────────────
import glob as _glob
DATA_DIR    = os.environ.get("DATA_DIR", "all_data/patches")
TEST_FEAT   = os.environ.get("TEST_FEAT", f"{DATA_DIR}/test/features")
TEST_MASK   = os.environ.get("TEST_MASK", f"{DATA_DIR}/test/masks")

# Resolve the run name -> checkpoint/result dirs (mirrors train.py).
RUN_NAME = os.environ.get("RUN_NAME")
if RUN_NAME is None:
    _sel = os.environ.get("BANDS", "all")
    RUN_NAME = "all" if _sel == "all" else _sel + "band"
CKPT_DIR  = os.environ.get("CKPT_DIR",
    "all_data/checkpoint" if RUN_NAME == "all" else f"all_data/checkpoint_{RUN_NAME}")
CKPT_PATH = os.environ.get("CKPT_PATH", f"{CKPT_DIR}/best_model.weights.h5")
OUT_DIR   = os.environ.get("OUT_DIR",
    "all_data/result" if RUN_NAME == "all" else f"all_data/result_{RUN_NAME}")

# Bands MUST match training. Prefer run_config.json written by train.py;
# else BAND_LIST env; else all bands. This makes a mismatch impossible.
_band_names = json.load(open(f"{DATA_DIR}/metadata/norm_stats.json"))["band_names"]
_rc = os.path.join(CKPT_DIR, "run_config.json")
if os.environ.get("BAND_LIST"):
    _bands = [b.strip() for b in os.environ["BAND_LIST"].split(",") if b.strip()]
elif os.path.exists(_rc):
    _bands = json.load(open(_rc)).get("bands")
else:
    _bands = None
if _bands and list(_bands) != list(_band_names):
    BAND_INDICES = [_band_names.index(b) for b in _bands]
else:
    BAND_INDICES = None

# Patch size auto-detected from a sample test patch.
_ph, _pw = np.load(sorted(_glob.glob(f"{TEST_FEAT}/*.npy"))[0]).shape[:2]
INPUT_SHAPE = (_ph, _pw, len(BAND_INDICES) if BAND_INDICES else len(_band_names))
TARGET      = INPUT_SHAPE[:2]
print(f"[eval] run={RUN_NAME}  bands={INPUT_SHAPE[2]}  patch={_ph}x{_pw}  ckpt={CKPT_PATH}")
BATCH_SIZE  = 8
THRESHOLDS  = [0.30, 0.35, 0.40, 0.45, 0.50, 0.55, 0.60, 0.65, 0.70]
USE_TTA     = True   # 8-way test-time augmentation (4 rotations x 2 flips)

os.makedirs(OUT_DIR, exist_ok=True)


# ─────────────────────────────────────────────────────────────────────────────
# Normalization: load the exact stats the checkpoint was trained with
# (checkpoint_dataset/norm_stats.json, written by train.py). Falls back to the
# built-in GLOBAL_MEAN/STD when absent so OLD checkpoints stay consistent.
# ─────────────────────────────────────────────────────────────────────────────
def _load_eval_norm():
    cand = os.path.join(os.path.dirname(CKPT_PATH), "norm_stats.json")
    if os.path.exists(cand):
        st = json.load(open(cand))
        m = np.array(st["mean"], np.float32).reshape(1, 1, -1)
        s = np.array(st["std"], np.float32).reshape(1, 1, -1)
        print(f"[eval] normalization: loaded {cand} (matches training)")
        return m, s
    print("[eval] normalization: built-in GLOBAL_MEAN/STD (no saved stats in checkpoint dir)")
    return GLOBAL_MEAN, GLOBAL_STD


NORM_MEAN, NORM_STD = _load_eval_norm()


# ─────────────────────────────────────────────────────────────────────────────
# Loss / metric helpers (needed for model.compile)
# ─────────────────────────────────────────────────────────────────────────────
def dice_loss(y_true, y_pred):
    smooth = 1e-6
    yt = tf.cast(y_true, tf.float32)
    i = tf.reduce_sum(yt * y_pred, axis=[1, 2, 3])
    u = tf.reduce_sum(yt, axis=[1, 2, 3]) + tf.reduce_sum(y_pred, axis=[1, 2, 3])
    return 1.0 - tf.reduce_mean((2.0 * i + smooth) / (u + smooth))


def combined_loss(y_true, y_pred):
    bce = tf.reduce_mean(tf.keras.losses.binary_crossentropy(y_true, y_pred)) * 2.0
    return bce + dice_loss(y_true, y_pred)


def dice_coef(y_true, y_pred):
    yp = tf.cast(y_pred > 0.5, tf.float32); yt = tf.cast(y_true, tf.float32)
    return (2 * tf.reduce_sum(yt * yp) + 1e-6) / (tf.reduce_sum(yt) + tf.reduce_sum(yp) + 1e-6)


# ─────────────────────────────────────────────────────────────────────────────
# Preprocess one image exactly like the (non-augment) data generator.
# Returns the padded/cropped 128x128 feature + mask AND the valid region size.
# ─────────────────────────────────────────────────────────────────────────────
def load_and_prep(f_path, m_path):
    with rasterio.open(f_path) as s:
        feat = np.moveaxis(s.read(), 0, -1)              # (H, W, 12)
    with rasterio.open(m_path) as s:
        mask = np.expand_dims(s.read(1), axis=-1)        # (H, W, 1)

    # Force mask to match feature shape (legacy .tif misalignment fix)
    H_f, W_f = feat.shape[:2]
    mask = mask[:H_f, :W_f, :]
    pad_mh, pad_mw = max(0, H_f - mask.shape[0]), max(0, W_f - mask.shape[1])
    if pad_mh > 0 or pad_mw > 0:
        mask = np.pad(mask, ((0, pad_mh), (0, pad_mw), (0, 0)), mode='constant')

    H, W = feat.shape[:2]
    feat = np.where((feat == -9999) | ~np.isfinite(feat), 0, feat).astype(np.float32)
    feat = (feat - NORM_MEAN) / (NORM_STD + 1e-7)

    th, tw = TARGET
    pad_h, pad_w = max(0, th - H), max(0, tw - W)
    if pad_h or pad_w:
        feat = np.pad(feat, ((0, pad_h), (0, pad_w), (0, 0)), mode='constant')
        mask = np.pad(mask, ((0, pad_h), (0, pad_w), (0, 0)), mode='constant')

    h, w = feat.shape[:2]
    if h > th or w > tw:                                  # center-crop large images
        hs = (h - th) // 2 if h > th else 0
        ws = (w - tw) // 2 if w > tw else 0
        feat = feat[hs:hs + th, ws:ws + tw, :]
        mask = mask[hs:hs + th, ws:ws + tw, :]

    # valid (real, non-padded) region inside the 128x128 canvas
    valid_h, valid_w = min(H, th), min(W, tw)
    return (feat.astype(np.float32), mask.astype(np.float32),
            H, W, valid_h, valid_w, pad_h > 0 or pad_w > 0, H > th or W > tw)


# ─────────────────────────────────────────────────────────────────────────────
# Build model + load weights
# ─────────────────────────────────────────────────────────────────────────────
print("Building model…")
_MODEL_TYPE = os.environ.get("MODEL_NAME", "swin_unet")
if _MODEL_TYPE == "attention_unet":
    model = build_attention_unet(input_shape=INPUT_SHAPE, num_classes=1)
elif _MODEL_TYPE == "unet":
    model = build_unet(input_shape=INPUT_SHAPE, num_classes=1)
elif _MODEL_TYPE == "segformer":
    model = build_segformer(input_shape=INPUT_SHAPE, num_classes=1)
else:
    model = build_model(input_shape=INPUT_SHAPE, num_classes=1, num_heads=8)
print(f"[eval] model type: {_MODEL_TYPE}")
model.compile(loss=combined_loss, metrics=['accuracy', dice_coef])

if not os.path.exists(CKPT_PATH):
    raise FileNotFoundError(f"Checkpoint not found: {CKPT_PATH}. Train first (python train.py).")
model.load_weights(CKPT_PATH)
print("Weights loaded.")

# ─────────────────────────────────────────────────────────────────────────────
# Load + preprocess every test image
# ─────────────────────────────────────────────────────────────────────────────
# Detect .npy patches (new pipeline) or .tif files (legacy)
npy_files = sorted(glob.glob(os.path.join(TEST_FEAT, '*.npy')))
is_npy = len(npy_files) > 0

if is_npy:
    feat_files = npy_files
    mask_files = [os.path.join(TEST_MASK, os.path.basename(f)) for f in feat_files]
    for mf in mask_files:
        assert os.path.exists(mf), f"Missing mask: {mf}"
    print(f"\nLoading {len(feat_files)} test patches (.npy)…")
else:
    feat_files = sorted(glob.glob(os.path.join(TEST_FEAT, 'feature_*.tif')))
    mask_files = sorted(glob.glob(os.path.join(TEST_MASK, 'mask_*.tif')))
    assert len(feat_files) == len(mask_files) and len(feat_files) > 0, "Test feature/mask mismatch."
    print(f"\nLoading {len(feat_files)} test images (.tif)…")

X, Y, meta, ids = [], [], [], []
for fp, mp in zip(feat_files, mask_files):
    if is_npy:
        feat = np.load(fp).astype(np.float32)       # (H, W, num_bands)
        mask = np.load(mp).astype(np.float32)        # (H, W, 1)
        H, W = feat.shape[:2]
        feat = np.where((feat == -9999) | ~np.isfinite(feat), 0, feat)
        if BAND_INDICES is not None:
            feat = feat[..., BAND_INDICES]
        feat = (feat - NORM_MEAN) / (NORM_STD + 1e-7)
        meta.append((H, W, H, W, False, False))
        ids.append(os.path.basename(fp).replace('.npy', ''))
    else:
        feat, mask, H, W, vh, vw, padded, cropped = load_and_prep(fp, mp)
        meta.append((H, W, vh, vw, padded, cropped))
        ids.append(os.path.basename(fp).replace('feature_', '').replace('.tif', ''))
    X.append(feat)
    Y.append(mask)

X = np.array(X, dtype=np.float32)       # (N, 128, 128, 12)
Y = np.array(Y, dtype=np.float32)       # (N, 128, 128, 1)


def predict_tta(model, X, batch_size):
    """Average sigmoid probs over the 8 dihedral (D4) transforms."""
    acc = np.zeros((X.shape[0], X.shape[1], X.shape[2], 1), np.float32)
    n = 0
    for flip in (False, True):
        for k in (0, 1, 2, 3):
            Xt = X[:, :, ::-1, :] if flip else X
            Xt = np.ascontiguousarray(np.rot90(Xt, k, axes=(1, 2)))
            p = model.predict(Xt, batch_size=batch_size, verbose=0)
            p = np.rot90(p, -k, axes=(1, 2))
            if flip:
                p = p[:, :, ::-1, :]
            acc += p
            n += 1
    return acc / n


if USE_TTA:
    print("Predicting (8-way TTA: 4 rotations x 2 flips)…")
    Y_prob = predict_tta(model, X, BATCH_SIZE)                # (N, 128, 128, 1)
else:
    print("Predicting…")
    Y_prob = model.predict(X, batch_size=BATCH_SIZE, verbose=1)   # (N, 128, 128, 1)


# ─────────────────────────────────────────────────────────────────────────────
# Collect VALID-region pixels (drop zero-padding) for global + per-image metrics
# ─────────────────────────────────────────────────────────────────────────────
def metrics_from_cm(tp, fp, fn, tn):
    prec = tp / (tp + fp + 1e-9)
    rec  = tp / (tp + fn + 1e-9)
    f1   = 2 * prec * rec / (prec + rec + 1e-9)
    iou  = tp / (tp + fp + fn + 1e-9)
    dice = 2 * tp / (2 * tp + fp + fn + 1e-9)
    acc  = (tp + tn) / (tp + tn + fp + fn + 1e-9)
    spec = tn / (tn + fp + 1e-9)
    return prec, rec, f1, iou, dice, acc, spec


# ─────────────────────────────────────────────────────────────────────────────
# Per-image metrics @ t=0.5  (worst IoU first)
# ─────────────────────────────────────────────────────────────────────────────
all_true, all_prob = [], []
per_image = []
for i, iid in enumerate(ids):
    H, W, vh, vw, padded, cropped = meta[i]
    yt = Y[i, :vh, :vw, 0].flatten().astype(np.int32)
    yprob = Y_prob[i, :vh, :vw, 0].flatten()
    all_true.append(yt); all_prob.append(yprob)

    yp = (yprob >= 0.5).astype(np.int32)
    tn, fp, fn, tp = confusion_matrix(yt, yp, labels=[0, 1]).ravel()
    prec, rec, f1, iou, dice, acc, spec = metrics_from_cm(tp, fp, fn, tn)
    # Empty GT + empty prediction = perfect agreement (correct definition;
    # otherwise background-only tiles wrongly score IoU=0 and drag the mean down).
    if (tp + fp + fn) == 0:
        prec = rec = f1 = iou = dice = 1.0
    n_px = yt.size
    per_image.append({
        'image_id':       iid,
        'orig_H':         H,
        'orig_W':         W,
        'valid_px':       int(n_px),
        'was_padded':     bool(padded),
        'was_cropped':    bool(cropped),
        'glacier_px_gt':  int(yt.sum()),
        'glacier_frac_gt': round(float(yt.mean()), 4),
        'glacier_px_pred': int(yp.sum()),
        'TP': int(tp), 'FP': int(fp), 'FN': int(fn), 'TN': int(tn),
        'precision':      round(prec, 4),
        'recall':         round(rec, 4),
        'f1_score':       round(f1, 4),
        'iou':            round(iou, 4),
        'dice':           round(dice, 4),
        'accuracy':       round(acc, 4),
        'mean_pred_prob': round(float(yprob.mean()), 4),
        'max_pred_prob':  round(float(yprob.max()), 4),
    })

per_image_df = pd.DataFrame(per_image).sort_values('iou', ascending=True).reset_index(drop=True)
per_image_df.to_csv(os.path.join(OUT_DIR, 'per_image_metrics.csv'), index=False)

# ─────────────────────────────────────────────────────────────────────────────
# Global threshold sweep (valid pixels only)
# ─────────────────────────────────────────────────────────────────────────────
y_true_flat = np.concatenate(all_true)
y_prob_flat = np.concatenate(all_prob)

rows = []
for t in THRESHOLDS:
    yp = (y_prob_flat >= t).astype(np.int32)
    tn, fp, fn, tp = confusion_matrix(y_true_flat, yp, labels=[0, 1]).ravel()
    prec, rec, f1, iou, dice, acc, spec = metrics_from_cm(tp, fp, fn, tn)
    rows.append({
        'threshold': t,
        'TP': int(tp), 'FP': int(fp), 'TN': int(tn), 'FN': int(fn),
        'precision': round(prec, 6), 'recall': round(rec, 6),
        'specificity': round(spec, 6), 'f1_score': round(f1, 6),
        'iou': round(iou, 6), 'dice': round(dice, 6), 'accuracy': round(acc, 6),
    })
thresh_df = pd.DataFrame(rows)
thresh_df.to_csv(os.path.join(OUT_DIR, 'evaluation_metrics.csv'), index=False)

best_row = thresh_df.loc[thresh_df['iou'].idxmax()]
best_t = float(best_row['threshold'])

cm = confusion_matrix(y_true_flat, (y_prob_flat >= best_t).astype(int), labels=[0, 1])
pd.DataFrame(cm,
             index=['actual_background', 'actual_glacier'],
             columns=['pred_background', 'pred_glacier']
             ).to_csv(os.path.join(OUT_DIR, 'confusion_matrix.csv'))

# ─────────────────────────────────────────────────────────────────────────────
# Summary
# ─────────────────────────────────────────────────────────────────────────────
mean_iou = per_image_df['iou'].mean()
median_iou = per_image_df['iou'].median()
worst = per_image_df.head(10)[['image_id', 'orig_H', 'orig_W', 'glacier_frac_gt', 'iou', 'precision', 'recall']]

# Machine-readable metrics for results_table.py (reliable, no text parsing).
with open(os.path.join(OUT_DIR, 'metrics.json'), 'w') as f:
    json.dump({
        'run_name':            RUN_NAME,
        'n_bands':             int(INPUT_SHAPE[2]),
        'patch':               [int(_ph), int(_pw)],
        'best_threshold':      best_t,
        'test_precision':      round(float(best_row['precision']), 4),
        'test_recall':         round(float(best_row['recall']), 4),
        'test_f1':             round(float(best_row['f1_score']), 4),
        'test_iou':            round(float(best_row['iou']), 4),
        'test_dice':           round(float(best_row['dice']), 4),
        'per_image_mean_iou':  round(float(mean_iou), 4),
    }, f, indent=2)

lines = [
    "=" * 60,
    "  ROCKGLACIERNET — EVALUATION SUMMARY (baseline 128x128)",
    "=" * 60,
    "",
    f"Test images           : {len(ids)}",
    f"Best threshold by IoU : {best_t}",
    "",
    "GLOBAL (pixel-pooled) metrics @ best threshold:",
    f"  precision : {best_row['precision']:.4f}",
    f"  recall    : {best_row['recall']:.4f}",
    f"  f1        : {best_row['f1_score']:.4f}",
    f"  iou       : {best_row['iou']:.4f}",
    f"  dice      : {best_row['dice']:.4f}",
    "",
    "PER-IMAGE IoU (@ t=0.5):",
    f"  mean   : {mean_iou:.4f}",
    f"  median : {median_iou:.4f}",
    "",
    "Threshold sweep:",
    "  {:<8} {:<10} {:<10} {:<10} {:<10} {:<10}".format(
        "thresh", "precision", "recall", "f1", "iou", "dice"),
    "  " + "-" * 60,
]
for _, r in thresh_df.iterrows():
    tag = " <- best IoU" if r['threshold'] == best_t else ""
    lines.append("  {:<8} {:<10.4f} {:<10.4f} {:<10.4f} {:<10.4f} {:<10.4f}{}".format(
        r['threshold'], r['precision'], r['recall'], r['f1_score'], r['iou'], r['dice'], tag))

lines += ["", "10 WORST images by IoU (@ t=0.5):", worst.to_string(index=False)]
lines += [
    "",
    f"Output dir: {OUT_DIR}",
    "Files: per_image_metrics.csv, evaluation_metrics.csv, confusion_matrix.csv",
]
summary = "\n".join(lines)
print("\n" + summary)
with open(os.path.join(OUT_DIR, 'evaluation_summary.txt'), 'w') as f:
    f.write(summary)

# ─────────────────────────────────────────────────────────────────────────────
# Prediction comparison PNGs (10 worst + 10 best)
# ─────────────────────────────────────────────────────────────────────────────
try:
    import matplotlib
    matplotlib.use('Agg')
    import matplotlib.pyplot as plt

    id_to_idx = {iid: i for i, iid in enumerate(ids)}
    to_plot = list(per_image_df.head(10)['image_id']) + list(per_image_df.tail(10)['image_id'])
    for iid in to_plot:
        i = id_to_idx[iid]
        H, W, vh, vw, padded, cropped = meta[i]
        fig, axes = plt.subplots(1, 3, figsize=(9, 3))
        rgb = X[i, :vh, :vw, :][:, :, [2, 1, 0]]
        rgb = (rgb - rgb.min()) / (rgb.max() - rgb.min() + 1e-8)
        iou_val = float(per_image_df.loc[per_image_df['image_id'] == iid, 'iou'].iloc[0])
        axes[0].imshow(rgb);                                     axes[0].set_title(f'{iid} RGB'); axes[0].axis('off')
        axes[1].imshow(Y[i, :vh, :vw, 0], cmap='gray');          axes[1].set_title('GT');         axes[1].axis('off')
        axes[2].imshow((Y_prob[i, :vh, :vw, 0] >= best_t), cmap='gray')
        axes[2].set_title(f'Pred IoU={iou_val:.2f}');            axes[2].axis('off')
        plt.tight_layout()
        plt.savefig(os.path.join(OUT_DIR, f'pred_{iid}.png'), dpi=80)
        plt.close()
    print(f"Saved {len(to_plot)} comparison PNGs (10 worst + 10 best).")
except Exception as e:
    print(f"PNG save skipped: {e}")

print("\nEvaluation complete.")

	"""
	Evaluation script for RockGlacierNet — baseline 128x128 full-image model.

	Each test GeoTIFF is normalized and padded/cropped to 128x128 exactly like
	training, then predicted. Metrics are computed on the VALID (non-padded) region
	of each image so zero-padding does not inflate the scores.

	Outputs (outputs/predictions/baseline_model_output/)
	-----------------------------------------------------
	per_image_metrics.csv — ONE ROW PER TEST IMAGE, sorted worst IoU first
	(this is the file to inspect for problem images)
	evaluation_metrics.csv — global threshold sweep 0.30–0.70
	confusion_matrix.csv — aggregate CM at best IoU threshold
	evaluation_summary.txt — human-readable summary
	pred_*.png — RGB / GT / prediction comparison figures
	"""

	import os
	import glob
	import json
	import numpy as np
	np.Inf = np.inf
	import pandas as pd
	import rasterio
	import tensorflow as tf
	from sklearn.metrics import confusion_matrix

	from data_generator import GLOBAL_MEAN, GLOBAL_STD
	from models.swin_unet import build_model
	from models.attention_unet import build_attention_unet
	from models.unet import build_unet
	from models.segformer import build_segformer

	# ─────────────────────────────────────────────────────────────────────────────
	# Config
	# ─────────────────────────────────────────────────────────────────────────────
	import glob as _glob
	DATA_DIR = os.environ.get("DATA_DIR", "all_data/patches")
	TEST_FEAT = os.environ.get("TEST_FEAT", f"{DATA_DIR}/test/features")
	TEST_MASK = os.environ.get("TEST_MASK", f"{DATA_DIR}/test/masks")

	# Resolve the run name -> checkpoint/result dirs (mirrors train.py).
	RUN_NAME = os.environ.get("RUN_NAME")
	if RUN_NAME is None:
	_sel = os.environ.get("BANDS", "all")
	RUN_NAME = "all" if _sel == "all" else _sel + "band"
	CKPT_DIR = os.environ.get("CKPT_DIR",
	"all_data/checkpoint" if RUN_NAME == "all" else f"all_data/checkpoint_{RUN_NAME}")
	CKPT_PATH = os.environ.get("CKPT_PATH", f"{CKPT_DIR}/best_model.weights.h5")
	OUT_DIR = os.environ.get("OUT_DIR",
	"all_data/result" if RUN_NAME == "all" else f"all_data/result_{RUN_NAME}")

	# Bands MUST match training. Prefer run_config.json written by train.py;
	# else BAND_LIST env; else all bands. This makes a mismatch impossible.
	_band_names = json.load(open(f"{DATA_DIR}/metadata/norm_stats.json"))["band_names"]
	_rc = os.path.join(CKPT_DIR, "run_config.json")
	if os.environ.get("BAND_LIST"):
	_bands = [b.strip() for b in os.environ["BAND_LIST"].split(",") if b.strip()]
	elif os.path.exists(_rc):
	_bands = json.load(open(_rc)).get("bands")
	else:
	_bands = None
	if _bands and list(_bands) != list(_band_names):
	BAND_INDICES = [_band_names.index(b) for b in _bands]
	else:
	BAND_INDICES = None

	# Patch size auto-detected from a sample test patch.
	_ph, _pw = np.load(sorted(_glob.glob(f"{TEST_FEAT}/*.npy"))[0]).shape[:2]
	INPUT_SHAPE = (_ph, _pw, len(BAND_INDICES) if BAND_INDICES else len(_band_names))
	TARGET = INPUT_SHAPE[:2]
	print(f"[eval] run={RUN_NAME} bands={INPUT_SHAPE[2]} patch={_ph}x{_pw} ckpt={CKPT_PATH}")
	BATCH_SIZE = 8
	THRESHOLDS = [0.30, 0.35, 0.40, 0.45, 0.50, 0.55, 0.60, 0.65, 0.70]
	USE_TTA = True # 8-way test-time augmentation (4 rotations x 2 flips)

	os.makedirs(OUT_DIR, exist_ok=True)


	# ─────────────────────────────────────────────────────────────────────────────
	# Normalization: load the exact stats the checkpoint was trained with
	# (checkpoint_dataset/norm_stats.json, written by train.py). Falls back to the
	# built-in GLOBAL_MEAN/STD when absent so OLD checkpoints stay consistent.
	# ─────────────────────────────────────────────────────────────────────────────
	def _load_eval_norm():
	cand = os.path.join(os.path.dirname(CKPT_PATH), "norm_stats.json")
	if os.path.exists(cand):
	st = json.load(open(cand))
	m = np.array(st["mean"], np.float32).reshape(1, 1, -1)
	s = np.array(st["std"], np.float32).reshape(1, 1, -1)
	print(f"[eval] normalization: loaded {cand} (matches training)")
	return m, s
	print("[eval] normalization: built-in GLOBAL_MEAN/STD (no saved stats in checkpoint dir)")
	return GLOBAL_MEAN, GLOBAL_STD


	NORM_MEAN, NORM_STD = _load_eval_norm()


	# ─────────────────────────────────────────────────────────────────────────────
	# Loss / metric helpers (needed for model.compile)
	# ─────────────────────────────────────────────────────────────────────────────
	def dice_loss(y_true, y_pred):
	smooth = 1e-6
	yt = tf.cast(y_true, tf.float32)
	i = tf.reduce_sum(yt * y_pred, axis=[1, 2, 3])
	u = tf.reduce_sum(yt, axis=[1, 2, 3]) + tf.reduce_sum(y_pred, axis=[1, 2, 3])
	return 1.0 - tf.reduce_mean((2.0 * i + smooth) / (u + smooth))


	def combined_loss(y_true, y_pred):
	bce = tf.reduce_mean(tf.keras.losses.binary_crossentropy(y_true, y_pred)) * 2.0
	return bce + dice_loss(y_true, y_pred)


	def dice_coef(y_true, y_pred):
	yp = tf.cast(y_pred > 0.5, tf.float32); yt = tf.cast(y_true, tf.float32)
	return (2 * tf.reduce_sum(yt * yp) + 1e-6) / (tf.reduce_sum(yt) + tf.reduce_sum(yp) + 1e-6)


	# ─────────────────────────────────────────────────────────────────────────────
	# Preprocess one image exactly like the (non-augment) data generator.
	# Returns the padded/cropped 128x128 feature + mask AND the valid region size.
	# ─────────────────────────────────────────────────────────────────────────────
	def load_and_prep(f_path, m_path):
	with rasterio.open(f_path) as s:
	feat = np.moveaxis(s.read(), 0, -1) # (H, W, 12)
	with rasterio.open(m_path) as s:
	mask = np.expand_dims(s.read(1), axis=-1) # (H, W, 1)

	# Force mask to match feature shape (legacy .tif misalignment fix)
	H_f, W_f = feat.shape[:2]
	mask = mask[:H_f, :W_f, :]
	pad_mh, pad_mw = max(0, H_f - mask.shape[0]), max(0, W_f - mask.shape[1])
	if pad_mh > 0 or pad_mw > 0:
	mask = np.pad(mask, ((0, pad_mh), (0, pad_mw), (0, 0)), mode='constant')

	H, W = feat.shape[:2]
	feat = np.where((feat == -9999) \| ~np.isfinite(feat), 0, feat).astype(np.float32)
	feat = (feat - NORM_MEAN) / (NORM_STD + 1e-7)

	th, tw = TARGET
	pad_h, pad_w = max(0, th - H), max(0, tw - W)
	if pad_h or pad_w:
	feat = np.pad(feat, ((0, pad_h), (0, pad_w), (0, 0)), mode='constant')
	mask = np.pad(mask, ((0, pad_h), (0, pad_w), (0, 0)), mode='constant')

	h, w = feat.shape[:2]
	if h > th or w > tw: # center-crop large images
	hs = (h - th) // 2 if h > th else 0
	ws = (w - tw) // 2 if w > tw else 0
	feat = feat[hs:hs + th, ws:ws + tw, :]
	mask = mask[hs:hs + th, ws:ws + tw, :]

	# valid (real, non-padded) region inside the 128x128 canvas
	valid_h, valid_w = min(H, th), min(W, tw)
	return (feat.astype(np.float32), mask.astype(np.float32),
	H, W, valid_h, valid_w, pad_h > 0 or pad_w > 0, H > th or W > tw)


	# ─────────────────────────────────────────────────────────────────────────────
	# Build model + load weights
	# ─────────────────────────────────────────────────────────────────────────────
	print("Building model…")
	_MODEL_TYPE = os.environ.get("MODEL_NAME", "swin_unet")
	if _MODEL_TYPE == "attention_unet":
	model = build_attention_unet(input_shape=INPUT_SHAPE, num_classes=1)
	elif _MODEL_TYPE == "unet":
	model = build_unet(input_shape=INPUT_SHAPE, num_classes=1)
	elif _MODEL_TYPE == "segformer":
	model = build_segformer(input_shape=INPUT_SHAPE, num_classes=1)
	else:
	model = build_model(input_shape=INPUT_SHAPE, num_classes=1, num_heads=8)
	print(f"[eval] model type: {_MODEL_TYPE}")
	model.compile(loss=combined_loss, metrics=['accuracy', dice_coef])

	if not os.path.exists(CKPT_PATH):
	raise FileNotFoundError(f"Checkpoint not found: {CKPT_PATH}. Train first (python train.py).")
	model.load_weights(CKPT_PATH)
	print("Weights loaded.")

	# ─────────────────────────────────────────────────────────────────────────────
	# Load + preprocess every test image
	# ─────────────────────────────────────────────────────────────────────────────
	# Detect .npy patches (new pipeline) or .tif files (legacy)
	npy_files = sorted(glob.glob(os.path.join(TEST_FEAT, '*.npy')))
	is_npy = len(npy_files) > 0

	if is_npy:
	feat_files = npy_files
	mask_files = [os.path.join(TEST_MASK, os.path.basename(f)) for f in feat_files]
	for mf in mask_files:
	assert os.path.exists(mf), f"Missing mask: {mf}"
	print(f"\nLoading {len(feat_files)} test patches (.npy)…")
	else:
	feat_files = sorted(glob.glob(os.path.join(TEST_FEAT, 'feature_*.tif')))
	mask_files = sorted(glob.glob(os.path.join(TEST_MASK, 'mask_*.tif')))
	assert len(feat_files) == len(mask_files) and len(feat_files) > 0, "Test feature/mask mismatch."
	print(f"\nLoading {len(feat_files)} test images (.tif)…")

	X, Y, meta, ids = [], [], [], []
	for fp, mp in zip(feat_files, mask_files):
	if is_npy:
	feat = np.load(fp).astype(np.float32) # (H, W, num_bands)
	mask = np.load(mp).astype(np.float32) # (H, W, 1)
	H, W = feat.shape[:2]
	feat = np.where((feat == -9999) \| ~np.isfinite(feat), 0, feat)
	if BAND_INDICES is not None:
	feat = feat[..., BAND_INDICES]
	feat = (feat - NORM_MEAN) / (NORM_STD + 1e-7)
	meta.append((H, W, H, W, False, False))
	ids.append(os.path.basename(fp).replace('.npy', ''))
	else:
	feat, mask, H, W, vh, vw, padded, cropped = load_and_prep(fp, mp)
	meta.append((H, W, vh, vw, padded, cropped))
	ids.append(os.path.basename(fp).replace('feature_', '').replace('.tif', ''))
	X.append(feat)
	Y.append(mask)

	X = np.array(X, dtype=np.float32) # (N, 128, 128, 12)
	Y = np.array(Y, dtype=np.float32) # (N, 128, 128, 1)


	def predict_tta(model, X, batch_size):
	"""Average sigmoid probs over the 8 dihedral (D4) transforms."""
	acc = np.zeros((X.shape[0], X.shape[1], X.shape[2], 1), np.float32)
	n = 0
	for flip in (False, True):
	for k in (0, 1, 2, 3):
	Xt = X[:, :, ::-1, :] if flip else X
	Xt = np.ascontiguousarray(np.rot90(Xt, k, axes=(1, 2)))
	p = model.predict(Xt, batch_size=batch_size, verbose=0)
	p = np.rot90(p, -k, axes=(1, 2))
	if flip:
	p = p[:, :, ::-1, :]
	acc += p
	n += 1
	return acc / n


	if USE_TTA:
	print("Predicting (8-way TTA: 4 rotations x 2 flips)…")
	Y_prob = predict_tta(model, X, BATCH_SIZE) # (N, 128, 128, 1)
	else:
	print("Predicting…")
	Y_prob = model.predict(X, batch_size=BATCH_SIZE, verbose=1) # (N, 128, 128, 1)


	# ─────────────────────────────────────────────────────────────────────────────
	# Collect VALID-region pixels (drop zero-padding) for global + per-image metrics
	# ─────────────────────────────────────────────────────────────────────────────
	def metrics_from_cm(tp, fp, fn, tn):
	prec = tp / (tp + fp + 1e-9)
	rec = tp / (tp + fn + 1e-9)
	f1 = 2 * prec * rec / (prec + rec + 1e-9)
	iou = tp / (tp + fp + fn + 1e-9)
	dice = 2 * tp / (2 * tp + fp + fn + 1e-9)
	acc = (tp + tn) / (tp + tn + fp + fn + 1e-9)
	spec = tn / (tn + fp + 1e-9)
	return prec, rec, f1, iou, dice, acc, spec


	# ─────────────────────────────────────────────────────────────────────────────
	# Per-image metrics @ t=0.5 (worst IoU first)
	# ─────────────────────────────────────────────────────────────────────────────
	all_true, all_prob = [], []
	per_image = []
	for i, iid in enumerate(ids):
	H, W, vh, vw, padded, cropped = meta[i]
	yt = Y[i, :vh, :vw, 0].flatten().astype(np.int32)
	yprob = Y_prob[i, :vh, :vw, 0].flatten()
	all_true.append(yt); all_prob.append(yprob)

	yp = (yprob >= 0.5).astype(np.int32)
	tn, fp, fn, tp = confusion_matrix(yt, yp, labels=[0, 1]).ravel()
	prec, rec, f1, iou, dice, acc, spec = metrics_from_cm(tp, fp, fn, tn)
	# Empty GT + empty prediction = perfect agreement (correct definition;
	# otherwise background-only tiles wrongly score IoU=0 and drag the mean down).
	if (tp + fp + fn) == 0:
	prec = rec = f1 = iou = dice = 1.0
	n_px = yt.size
	per_image.append({
	'image_id': iid,
	'orig_H': H,
	'orig_W': W,
	'valid_px': int(n_px),
	'was_padded': bool(padded),
	'was_cropped': bool(cropped),
	'glacier_px_gt': int(yt.sum()),
	'glacier_frac_gt': round(float(yt.mean()), 4),
	'glacier_px_pred': int(yp.sum()),
	'TP': int(tp), 'FP': int(fp), 'FN': int(fn), 'TN': int(tn),
	'precision': round(prec, 4),
	'recall': round(rec, 4),
	'f1_score': round(f1, 4),
	'iou': round(iou, 4),
	'dice': round(dice, 4),
	'accuracy': round(acc, 4),
	'mean_pred_prob': round(float(yprob.mean()), 4),
	'max_pred_prob': round(float(yprob.max()), 4),
	})

	per_image_df = pd.DataFrame(per_image).sort_values('iou', ascending=True).reset_index(drop=True)
	per_image_df.to_csv(os.path.join(OUT_DIR, 'per_image_metrics.csv'), index=False)

	# ─────────────────────────────────────────────────────────────────────────────
	# Global threshold sweep (valid pixels only)
	# ─────────────────────────────────────────────────────────────────────────────
	y_true_flat = np.concatenate(all_true)
	y_prob_flat = np.concatenate(all_prob)

	rows = []
	for t in THRESHOLDS:
	yp = (y_prob_flat >= t).astype(np.int32)
	tn, fp, fn, tp = confusion_matrix(y_true_flat, yp, labels=[0, 1]).ravel()
	prec, rec, f1, iou, dice, acc, spec = metrics_from_cm(tp, fp, fn, tn)
	rows.append({
	'threshold': t,
	'TP': int(tp), 'FP': int(fp), 'TN': int(tn), 'FN': int(fn),
	'precision': round(prec, 6), 'recall': round(rec, 6),
	'specificity': round(spec, 6), 'f1_score': round(f1, 6),
	'iou': round(iou, 6), 'dice': round(dice, 6), 'accuracy': round(acc, 6),
	})
	thresh_df = pd.DataFrame(rows)
	thresh_df.to_csv(os.path.join(OUT_DIR, 'evaluation_metrics.csv'), index=False)

	best_row = thresh_df.loc[thresh_df['iou'].idxmax()]
	best_t = float(best_row['threshold'])

	cm = confusion_matrix(y_true_flat, (y_prob_flat >= best_t).astype(int), labels=[0, 1])
	pd.DataFrame(cm,
	index=['actual_background', 'actual_glacier'],
	columns=['pred_background', 'pred_glacier']
	).to_csv(os.path.join(OUT_DIR, 'confusion_matrix.csv'))

	# ─────────────────────────────────────────────────────────────────────────────
	# Summary
	# ─────────────────────────────────────────────────────────────────────────────
	mean_iou = per_image_df['iou'].mean()
	median_iou = per_image_df['iou'].median()
	worst = per_image_df.head(10)[['image_id', 'orig_H', 'orig_W', 'glacier_frac_gt', 'iou', 'precision', 'recall']]

	# Machine-readable metrics for results_table.py (reliable, no text parsing).
	with open(os.path.join(OUT_DIR, 'metrics.json'), 'w') as f:
	json.dump({
	'run_name': RUN_NAME,
	'n_bands': int(INPUT_SHAPE[2]),
	'patch': [int(_ph), int(_pw)],
	'best_threshold': best_t,
	'test_precision': round(float(best_row['precision']), 4),
	'test_recall': round(float(best_row['recall']), 4),
	'test_f1': round(float(best_row['f1_score']), 4),
	'test_iou': round(float(best_row['iou']), 4),
	'test_dice': round(float(best_row['dice']), 4),
	'per_image_mean_iou': round(float(mean_iou), 4),
	}, f, indent=2)

	lines = [
	"=" * 60,
	" ROCKGLACIERNET — EVALUATION SUMMARY (baseline 128x128)",
	"=" * 60,
	"",
	f"Test images : {len(ids)}",
	f"Best threshold by IoU : {best_t}",
	"",
	"GLOBAL (pixel-pooled) metrics @ best threshold:",
	f" precision : {best_row['precision']:.4f}",
	f" recall : {best_row['recall']:.4f}",
	f" f1 : {best_row['f1_score']:.4f}",
	f" iou : {best_row['iou']:.4f}",
	f" dice : {best_row['dice']:.4f}",
	"",
	"PER-IMAGE IoU (@ t=0.5):",
	f" mean : {mean_iou:.4f}",
	f" median : {median_iou:.4f}",
	"",
	"Threshold sweep:",
	" {:<8} {:<10} {:<10} {:<10} {:<10} {:<10}".format(
	"thresh", "precision", "recall", "f1", "iou", "dice"),
	" " + "-" * 60,
	]
	for _, r in thresh_df.iterrows():
	tag = " <- best IoU" if r['threshold'] == best_t else ""
	lines.append(" {:<8} {:<10.4f} {:<10.4f} {:<10.4f} {:<10.4f} {:<10.4f}{}".format(
	r['threshold'], r['precision'], r['recall'], r['f1_score'], r['iou'], r['dice'], tag))

	lines += ["", "10 WORST images by IoU (@ t=0.5):", worst.to_string(index=False)]
	lines += [
	"",
	f"Output dir: {OUT_DIR}",
	"Files: per_image_metrics.csv, evaluation_metrics.csv, confusion_matrix.csv",
	]
	summary = "\n".join(lines)
	print("\n" + summary)
	with open(os.path.join(OUT_DIR, 'evaluation_summary.txt'), 'w') as f:
	f.write(summary)

	# ─────────────────────────────────────────────────────────────────────────────
	# Prediction comparison PNGs (10 worst + 10 best)
	# ─────────────────────────────────────────────────────────────────────────────
	try:
	import matplotlib
	matplotlib.use('Agg')
	import matplotlib.pyplot as plt

	id_to_idx = {iid: i for i, iid in enumerate(ids)}
	to_plot = list(per_image_df.head(10)['image_id']) + list(per_image_df.tail(10)['image_id'])
	for iid in to_plot:
	i = id_to_idx[iid]
	H, W, vh, vw, padded, cropped = meta[i]
	fig, axes = plt.subplots(1, 3, figsize=(9, 3))
	rgb = X[i, :vh, :vw, :][:, :, [2, 1, 0]]
	rgb = (rgb - rgb.min()) / (rgb.max() - rgb.min() + 1e-8)
	iou_val = float(per_image_df.loc[per_image_df['image_id'] == iid, 'iou'].iloc[0])
	axes[0].imshow(rgb); axes[0].set_title(f'{iid} RGB'); axes[0].axis('off')
	axes[1].imshow(Y[i, :vh, :vw, 0], cmap='gray'); axes[1].set_title('GT'); axes[1].axis('off')
	axes[2].imshow((Y_prob[i, :vh, :vw, 0] >= best_t), cmap='gray')
	axes[2].set_title(f'Pred IoU={iou_val:.2f}'); axes[2].axis('off')
	plt.tight_layout()
	plt.savefig(os.path.join(OUT_DIR, f'pred_{iid}.png'), dpi=80)
	plt.close()
	print(f"Saved {len(to_plot)} comparison PNGs (10 worst + 10 best).")
	except Exception as e:
	print(f"PNG save skipped: {e}")

	print("\nEvaluation complete.")