File size: 3,927 Bytes
42c70d5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 | """Re-score a Blue Marble Clouds model on CloudSEN12+ (L1C, official high-quality test split: 975 images, 4 classes).
pip install numpy onnxruntime rasterio huggingface_hub
python evaluate.py --model medium/model_fp32.onnx # or medium/model_int8.onnx
Data: CloudSEN12+ test_509_high.mlstac, about 2.7 GB, downloaded once."""
import numpy as np
# Sentinel-2 L1C, 13 bands (B01, B02, B03, B04, B05, B06, B07, B08, B8A, B09, B10, B11, B12)
MEAN = np.array([2357.089, 2137.385, 2018.788, 2082.986, 2295.651, 2854.537, 3122.849, 3040.560, 3306.481, 1473.847, 506.070, 2472.825, 1838.929], np.float32)
STD = np.array([1624.683, 1675.806, 1557.708, 1833.702, 1823.738, 1733.977, 1732.131, 1679.732, 1727.26, 1024.687, 442.165, 1331.411, 1160.419], np.float32)
def normalize(dn, bands=None):
"""dn: [C, H, W] digital numbers for the 13 bands, or for the band indices in `bands`. Missing bands are 0 after normalisation."""
out = np.zeros((13, *dn.shape[1:]), np.float32)
idx = list(range(13)) if bands is None else list(bands)
out[idx] = (dn.astype(np.float32) - MEAN[idx, None, None]) / STD[idx, None, None]
return out
def session(path, threads=0):
import onnxruntime as ort
so = ort.SessionOptions()
if threads: so.intra_op_num_threads = threads
prov = [p for p in ("CUDAExecutionProvider", "CPUExecutionProvider") if p in ort.get_available_providers()]
return ort.InferenceSession(path, so, providers=prov)
def predict(sess, x):
"""Whole image in one pass: pad H and W to multiples of 16 by edge replication, crop the logits back."""
h, w = x.shape[-2:]
xp = np.pad(x, ((0, 0), (0, (-h) % 16), (0, (-w) % 16)), mode="edge")[None].astype(np.float32)
return sess.run(None, {sess.get_inputs()[0].name: xp})[0][0, :, :h, :w].argmax(0)
class MIoU:
def __init__(self, classes):
self.classes, n = classes, len(classes)
self.cm = np.zeros((n, n), np.int64)
def add(self, y, p):
n = len(self.classes); m = y >= 0
self.cm += np.bincount(n * y[m].astype(np.int64) + p[m], minlength=n * n).reshape(n, n)
def report(self, title):
inter = np.diag(self.cm).astype(float); union = self.cm.sum(0) + self.cm.sum(1) - inter
iou = inter / np.maximum(union, 1)
print(f"{title}: mIoU {100 * iou.mean():.2f} (" + ", ".join(f"{c} {100 * v:.2f}" for c, v in zip(self.classes, iou)) + ")", flush=True)
return float(iou.mean())
import argparse, json, os, struct
import numpy as np
from rasterio.io import MemoryFile
def fetch(data):
p = f"{data}/test_509_high.mlstac"
if os.path.exists(p): return p
from huggingface_hub import hf_hub_download
return hf_hub_download("isp-uv-es/CloudSEN12Plus", "test/test_509_high.mlstac", repo_type="dataset", local_dir=data)
def images(path):
"""Container: b'#y', uint64 index length, JSON index {key: [offset, length]}, then one GeoTIFF per image
(bands 1-13: L1C B01-B12, band 14: expert label 0 clear, 1 thick cloud, 2 thin cloud, 3 cloud shadow)."""
with open(path, "rb") as f:
assert f.read(2) == b"#y"
n = struct.unpack("<Q", f.read(8))[0]; idx = json.loads(f.read(n)); off = 10 + n
for k in sorted(idx):
start, length = idx[k][:2]
f.seek(off + start)
with MemoryFile(f.read(length)) as mf, mf.open() as s: a = s.read()
yield a[:13].astype(np.uint16), a[13].astype(np.int64)
def main():
ap = argparse.ArgumentParser(); ap.add_argument("--model", required=True); ap.add_argument("--data", default="data")
a = ap.parse_args(); os.makedirs(a.data, exist_ok=True)
sess = session(a.model); m = MIoU(["clear", "thick cloud", "thin cloud", "cloud shadow"]); n = 0
for dn, y in images(fetch(a.data)):
m.add(y, predict(sess, normalize(dn))); n += 1
m.report(f"CloudSEN12+ test ({n} images)")
if __name__ == "__main__": main()
|