Datasets:
File size: 14,032 Bytes
d922ac9 | 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 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 | """Check that offset-0 windows reproduce the published 5 km terrain table.
The whole claim behind the quarter-cell design is that a window composed from
additive quarter statistics is not an approximation of the 5 km cell it covers
but the *same number*. ``offset_id == 0`` covers exactly the cell it is
anchored on, so every column in :data:`quarter_cells.EXACT_COLUMNS` must come
back bit-identical to the published terrain table, and only the six
percentile-derived columns may differ at all.
Comparison is on raw bit patterns, not on ``==``. Two reasons, both learned
here: ``NaN == NaN`` is false, and Earth is mostly ocean, so a value-wise
comparison reports every all-NaN column as a mismatch (that is exactly how a
first pass at verifying terrain-v0.2 appeared to find 20 broken columns). A
bit comparison instead demands that a NaN payload match too, which is strictly
stronger than equality.
One column is expected to differ and is reported separately rather than as a
failure: ``roughness_detrended_std`` solves a plane from sums here and from a
second pass over residuals in the extractor. The two disagree only in the last
bits, and only where the roughness is small enough that the residual sum of
squares is itself near cancellation. The script measures where the difference
lives instead of asserting it is zero.
Approximate columns are reported as ABSOLUTE error. Relative error is
meaningless on flat ice, where the published value is itself near zero and any
denominator built from it explodes.
Usage:
python validate_windows.py --planet earth \
--windows <dir of band_group=*.parquet> \
--terrain <dir of band_group=*.parquet> [--json report.json]
"""
from __future__ import annotations
import argparse
import json
from pathlib import Path
import numpy as np
import pyarrow.parquet as pq
import quarter_cells as qc
# The two count columns of EXACT_COLUMNS, named here only so the comparison
# can say which ones are integers rather than float bit patterns.
COUNT_COLUMNS = ('n_native_samples', 'n_valid_native_samples')
# Below this roughness the plane-from-sums and plane-from-residuals paths are
# allowed to disagree in the last bits; above it they must not disagree at all.
ROUGHNESS_FLOOR = 0.01
def _bits(values: np.ndarray) -> np.ndarray:
"""A float array as the unsigned integer of the same width.
Lets NaN compare equal to an identical NaN, and unequal to a NaN with a
different payload, which plain ``==`` cannot express.
"""
values = np.asarray(values)
if values.dtype == np.float32:
return values.view(np.uint32)
if values.dtype == np.float64:
return values.view(np.uint64)
return values
def group_numbers(directory: Path) -> list[int]:
return sorted(int(p.stem.split('=')[1])
for p in Path(directory).glob('band_group=*.parquet'))
def _key(band: np.ndarray, column: np.ndarray) -> np.ndarray:
"""One integer per cell, so the join is a sort rather than a merge."""
return band.astype(np.int64) * (1 << 32) + column.astype(np.int64)
class Accumulator:
"""Running totals over all band groups."""
def __init__(self):
self.exact = {name: {'compared': 0, 'mismatched': 0, 'max_abs_diff': 0.0}
for name in qc.EXACT_COLUMNS}
self.errors = {name: [] for name in qc.APPROXIMATE_COLUMNS}
self.rough_above = {'compared': 0, 'mismatched': 0, 'max_abs_diff': 0.0}
self.rough_below = {'compared': 0, 'mismatched': 0, 'max_abs_diff': 0.0}
self.rows = 0
self.terrain_rows = 0
self.terrain_measured = 0
self.windows_without_cell = 0
self.measured_without_window = 0
self.groups = 0
# The cells behind rough_above['mismatched'], so a release can check
# that the exceptions it accepts are exactly the ones that exist.
self.rough_above_cells = []
def exact_column(self, name: str, left: np.ndarray, right: np.ndarray) -> None:
slot = self.exact[name]
slot['compared'] += int(left.size)
differs = _bits(left) != _bits(right)
slot['mismatched'] += int(differs.sum())
if differs.any():
slot['max_abs_diff'] = max(slot['max_abs_diff'],
_finite_max_abs(left, right, differs))
def roughness(self, left: np.ndarray, right: np.ndarray,
band: np.ndarray | None = None,
column: np.ndarray | None = None) -> None:
"""Split the one expected difference by the size of the value.
With ``band`` and ``column`` the cells that differ at or above the
floor are also recorded (at most MAX_CELLS_LISTED in all).
"""
differs = _bits(left) != _bits(right)
with np.errstate(invalid='ignore'):
small = np.nan_to_num(np.abs(right), nan=0.0) < ROUGHNESS_FLOOR
for mask, slot in ((~small, self.rough_above), (small, self.rough_below)):
slot['compared'] += int(mask.sum())
hit = differs & mask
slot['mismatched'] += int(hit.sum())
if hit.any():
slot['max_abs_diff'] = max(slot['max_abs_diff'],
_finite_max_abs(left, right, hit))
if band is not None and column is not None:
for i in np.flatnonzero(differs & ~small):
if len(self.rough_above_cells) >= MAX_CELLS_LISTED:
break
self.rough_above_cells.append({
'band': int(band[i]), 'column': int(column[i]),
'window': _json_float(left[i]), 'terrain': _json_float(right[i])})
MAX_CELLS_LISTED = 100
def _json_float(value) -> float | None:
"""A float JSON can carry: NaN and inf become null."""
value = float(value)
return value if np.isfinite(value) else None
def _finite_max_abs(left, right, mask) -> float:
"""Largest finite |difference| among the masked entries.
A NaN that differs only in payload has no meaningful magnitude, and a
value present on one side and NaN on the other has no finite difference
either; both are already counted as mismatches.
"""
a = np.asarray(left, dtype=np.float64)[mask]
b = np.asarray(right, dtype=np.float64)[mask]
diff = np.abs(a - b)
good = np.isfinite(diff)
return float(diff[good].max()) if good.any() else 0.0
def compare_group(window_path: Path, terrain_path: Path, acc: Accumulator) -> dict:
windows = pq.read_table(window_path)
terrain = pq.read_table(terrain_path)
offset = windows['offset_id'].to_numpy()
windows = windows.filter(offset == 0)
w_key = _key(windows['band'].to_numpy(), windows['column'].to_numpy())
t_key = _key(terrain['band'].to_numpy(), terrain['column'].to_numpy())
order = np.argsort(t_key)
sorted_key = t_key[order]
position = np.searchsorted(sorted_key, w_key)
position = np.clip(position, 0, sorted_key.size - 1)
found = sorted_key[position] == w_key
t_index = order[position]
acc.rows += int(found.sum())
acc.terrain_rows += terrain.num_rows
acc.windows_without_cell += int((~found).sum())
acc.groups += 1
# A cell the extractor measured must have an offset-0 window. A cell it
# did not measure has nothing to reproduce, so its absence is correct.
measured = np.asarray(terrain['n_valid_samples_200m'].to_numpy()) > 0
acc.terrain_measured += int(measured.sum())
matched = np.zeros(terrain.num_rows, dtype=bool)
matched[t_index[found]] = True
acc.measured_without_window += int((measured & ~matched).sum())
w_rows = np.flatnonzero(found)
t_rows = t_index[found]
for name in qc.EXACT_COLUMNS:
left = windows[name].to_numpy(zero_copy_only=False)[w_rows]
right = terrain[name].to_numpy(zero_copy_only=False)[t_rows]
if name == 'roughness_detrended_std':
acc.roughness(left, right,
windows['band'].to_numpy()[w_rows],
windows['column'].to_numpy()[w_rows])
else:
acc.exact_column(name, left, right)
for name in qc.APPROXIMATE_COLUMNS:
left = np.asarray(windows[name].to_numpy(zero_copy_only=False)[w_rows],
dtype=np.float64)
right = np.asarray(terrain[name].to_numpy(zero_copy_only=False)[t_rows],
dtype=np.float64)
diff = np.abs(left - right)
# Both NaN is agreement, not an error; one-sided NaN is counted by the
# coverage numbers above and cannot be summarised as a magnitude.
acc.errors[name].append(diff[np.isfinite(diff)].astype(np.float32))
return {'group': int(window_path.stem.split('=')[1]),
'windows': int(found.sum())}
def summarise(acc: Accumulator) -> dict:
exact = {}
for name, slot in acc.exact.items():
if name == 'roughness_detrended_std':
continue
exact[name] = dict(slot)
approximate = {}
for name, parts in acc.errors.items():
values = (np.concatenate(parts) if parts
else np.zeros(0, dtype=np.float32))
approximate[name] = {
'compared': int(values.size),
'max_abs_error': float(values.max()) if values.size else 0.0,
'p99_abs_error': (float(np.percentile(values, 99))
if values.size else 0.0),
}
clean = all(v['mismatched'] == 0 for v in exact.values())
rough_ok = acc.rough_above['mismatched'] == 0
return {
'groups': acc.groups,
'offset0_windows_compared': acc.rows,
'terrain_rows': acc.terrain_rows,
'terrain_rows_measured': acc.terrain_measured,
'measured_cells_without_a_window': acc.measured_without_window,
'windows_without_a_terrain_row': acc.windows_without_cell,
'exact_columns': exact,
'roughness_detrended_std': {
f'at_or_above_{ROUGHNESS_FLOOR}_m': dict(acc.rough_above),
f'below_{ROUGHNESS_FLOOR}_m': dict(acc.rough_below),
},
# Kept out of roughness_detrended_std above: report() reads every
# entry there as a {compared, mismatched, max_abs_diff} slot.
'roughness_above_floor_cells': list(acc.rough_above_cells),
'approximate_columns': approximate,
'verdict': ('PASS' if clean and rough_ok
and acc.measured_without_window == 0
and acc.windows_without_cell == 0 else 'FAIL'),
}
def report(summary: dict) -> None:
print(f"\ngroups compared {summary['groups']}")
print(f"offset-0 windows {summary['offset0_windows_compared']:,}")
print(f"terrain rows {summary['terrain_rows']:,} "
f"({summary['terrain_rows_measured']:,} with a valid sample)")
print(f"measured cells with no window "
f"{summary['measured_cells_without_a_window']:,}")
print(f"windows with no terrain row "
f"{summary['windows_without_a_terrain_row']:,}")
print('\nEXACT columns (bit-identical required)')
print(f" {'column':<28} {'compared':>12} {'mismatched':>11} {'max |diff|':>11}")
for name, slot in summary['exact_columns'].items():
flag = '' if slot['mismatched'] == 0 else ' <-- MISMATCH'
print(f" {name:<28} {slot['compared']:>12,} {slot['mismatched']:>11,}"
f" {slot['max_abs_diff']:>11.3e}{flag}")
print('\nroughness_detrended_std (the one understood exception)')
for where, slot in summary['roughness_detrended_std'].items():
print(f" {where:<28} {slot['compared']:>12,} {slot['mismatched']:>11,}"
f" {slot['max_abs_diff']:>11.3e}")
for cell in summary.get('roughness_above_floor_cells', [])[:10]:
print(f" band {cell['band']} col {cell['column']}: "
f"window {cell['window']} vs terrain {cell['terrain']}")
print('\nAPPROXIMATE columns (absolute error vs published)')
print(f" {'column':<28} {'compared':>12} {'max abs':>12} {'p99 abs':>12}")
for name, slot in summary['approximate_columns'].items():
print(f" {name:<28} {slot['compared']:>12,} "
f"{slot['max_abs_error']:>12.4g} {slot['p99_abs_error']:>12.4g}")
print(f"\nverdict: {summary['verdict']}")
def main(argv=None) -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--planet', choices=('mars', 'earth'), required=True)
parser.add_argument('--windows', type=Path, required=True)
parser.add_argument('--terrain', type=Path, required=True)
parser.add_argument('--groups', type=str, default=None,
help='inclusive START-END over band groups')
parser.add_argument('--json', type=Path, default=None)
args = parser.parse_args(argv)
available = group_numbers(args.windows)
have_terrain = set(group_numbers(args.terrain))
if args.groups:
low, high = (int(v) for v in args.groups.split('-'))
available = [g for g in available if low <= g <= high]
missing = [g for g in available if g not in have_terrain]
if missing:
print(f'no terrain table for group(s): {missing}')
return 2
if not available:
print('no window band groups found')
return 2
acc = Accumulator()
for group in available:
stats = compare_group(args.windows / f'band_group={group:03d}.parquet',
args.terrain / f'band_group={group:03d}.parquet',
acc)
print(f"group {stats['group']:03d}: {stats['windows']:,} offset-0 windows",
flush=True)
summary = summarise(acc)
report(summary)
if args.json:
args.json.parent.mkdir(parents=True, exist_ok=True)
args.json.write_text(json.dumps(summary, indent=2, sort_keys=True))
print(f'wrote {args.json}')
return 0 if summary['verdict'] == 'PASS' else 1
if __name__ == '__main__':
raise SystemExit(main())
|