Datasets:
File size: 7,841 Bytes
c61008f | 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 | """Exact upstream splitting, lossless Parquet, and validated graph caches."""
import argparse
import gc
import hashlib
import json
from pathlib import Path
import pickle
import sys
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT))
import awkward as ak
import numpy as np
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from config import CONFIG
from data.preprocess_events import preprocess_and_split_events
from dataset import PhyGHTDataset
from download_data import digest
RUN = ROOT / 'runs/release-20261112'
PACKAGE = ROOT / 'release/artifacts/dataset'
SCHEMA = pa.schema([
('event_id', pa.int64()), ('pileup', pa.int32()),
('jets', pa.list_(pa.list_(pa.float64()))),
('trks', pa.list_(pa.list_(pa.float64()))),
('jet_trk_IDs', pa.list_(pa.list_(pa.int64()))),
])
def assert_equal(a, b):
# Arrow's nullable schema adds optional-type wrappers even when no value is
# null. Compare every nesting length and every numeric bit independently
# of those type wrappers; no precision or membership tolerance is allowed.
assert len(a) == len(b) and a.ndim == b.ndim, 'Changed event/list dimensions'
for axis in range(1, a.ndim):
assert ak.to_list(ak.num(a, axis=axis)) == ak.to_list(ak.num(b, axis=axis)), 'Changed list boundaries'
av = ak.to_numpy(ak.flatten(a, axis=None))
bv = ak.to_numpy(ak.flatten(b, axis=None))
assert av.dtype == bv.dtype and av.shape == bv.shape and av.tobytes() == bv.tobytes(), 'Changed numeric bits'
def main():
p = argparse.ArgumentParser()
p.add_argument('variant', choices=['mu200', 'mu60'])
args = p.parse_args()
variant = args.variant
status_path = RUN / f'{variant}_data_validated.json'
raw_report = json.loads((RUN / f'{variant}_raw_verified.json').read_text())
path = Path(raw_report['path'])
assert path.stat().st_size == raw_report['bytes']
assert digest(path, 'md5') == raw_report['md5'], 'Raw file changed since verification'
with open(path, 'rb') as stream:
raw = pickle.load(stream)
assert {'jets', 'trks', 'jet_trk_IDs'} == set(raw), set(raw)
assert all(len(x) == 10000 for x in raw.values())
assert bool(ak.all(ak.num(raw['jets'], axis=2) == 6))
assert bool(ak.all(ak.num(raw['trks'], axis=2) == 8))
assert bool(ak.all(ak.num(raw['jets'], axis=1) == ak.num(raw['jet_trk_IDs'], axis=1)))
for name in ['jets', 'trks']:
assert np.isfinite(ak.to_numpy(ak.flatten(raw[name], axis=None))).all(), name
tracks_per_event = ak.to_numpy(ak.num(raw['trks'], axis=1))
jets_per_event = ak.to_numpy(ak.num(raw['jets'], axis=1))
assert tracks_per_event.min() >= CONFIG['K_NEIGHBORS']
assert jets_per_event.min() > 0
shuffled = np.random.RandomState(42).permutation(10000)
indices = {'train': shuffled[:8000], 'validation': shuffled[8000:9000], 'test': shuffled[9000:]}
assert len(np.unique(np.concatenate(list(indices.values())))) == 10000
report = {
'variant': variant, 'raw': raw_report, 'seed': 42,
'split_algorithm': 'numpy.random.RandomState(42).permutation(10000); 80/10/10',
'counts': {k: len(v) for k, v in indices.items()},
'tracks_min_median_max': [int(tracks_per_event.min()), float(np.median(tracks_per_event)), int(tracks_per_event.max())],
'jets_min_median_max': [int(jets_per_event.min()), float(np.median(jets_per_event)), int(jets_per_event.max())],
'raw_awkward_types': {k: str(v.type) for k, v in raw.items()},
'parquet_files': [], 'equivalent_model_inputs': 0,
'missing_association_ids': 0, 'duplicate_track_ids': 0,
}
split_dir = RUN / 'processed_data'
split_dir.mkdir(exist_ok=True)
np.savez(split_dir / f'{variant}_event_indices_seed42.npz', **indices)
json_indices = PACKAGE / 'splits' / f'{variant}_seed42.json'
json_indices.parent.mkdir(parents=True, exist_ok=True)
json_indices.write_text(json.dumps({k: v.tolist() for k, v in indices.items()}))
preprocess_and_split_events(str(path), str(split_dir), f'{variant}_10k_events', seed=42)
CONFIG.update(DATASET_NAME=f'{variant}_10k_events', PROCESSED_DATA_DIR=str(split_dir))
graph_dir = RUN / 'graphs' / variant
graph_dir.mkdir(parents=True, exist_ok=True)
for split, event_ids in indices.items():
ds = PhyGHTDataset(split)
graphs = []
for i, original_id in enumerate(event_ids):
# Check exact event membership against the independently generated indices.
for name, actual in [('jets', ds.jets_ak[i]), ('trks', ds.tracks_ak[i]), ('jet_trk_IDs', ds.jet_trk_ids_ak[i])]:
assert_equal(raw[name][original_id], actual)
track_ids = ak.to_numpy(ds.tracks_ak[i][:, 6])
assert np.array_equal(track_ids, track_ids.astype(np.int64))
report['duplicate_track_ids'] += len(track_ids) - len(np.unique(track_ids))
assoc_ids = ak.to_numpy(ak.flatten(ds.jet_trk_ids_ak[i], axis=None))
report['missing_association_ids'] += int((~np.isin(assoc_ids, track_ids)).sum())
graph = ds[i]
graph.event_id = int(original_id)
graphs.append(graph)
graph_path = graph_dir / f'{split}.pt'
torch.save(graphs, graph_path)
for shard, start in enumerate(range(0, len(event_ids), 250)):
ids = event_ids[start:start + 250]
arrays = [pa.array(ids), pa.array(np.full(len(ids), int(variant[2:]), dtype=np.int32))]
arrays += [ak.to_arrow(raw[key][ids], extensionarray=False) for key in ['jets', 'trks', 'jet_trk_IDs']]
table = pa.Table.from_arrays(arrays, names=SCHEMA.names).cast(SCHEMA)
out = PACKAGE / variant / f'{split}-{shard:05d}.parquet'
out.parent.mkdir(parents=True, exist_ok=True)
pq.write_table(table, out, compression='zstd', row_group_size=25)
restored = pq.read_table(out)
assert restored.schema == SCHEMA
assert restored.column('event_id').to_pylist() == ids.tolist()
for key in ['jets', 'trks', 'jet_trk_IDs']:
assert_equal(raw[key][ids], ak.from_arrow(restored[key]))
# Exercise upstream graph construction on every converted event.
converted = PhyGHTDataset.__new__(PhyGHTDataset)
converted.jets_ak = ak.from_arrow(restored['jets'])
converted.tracks_ak = ak.from_arrow(restored['trks'])
converted.jet_trk_ids_ak = ak.from_arrow(restored['jet_trk_IDs'])
for local_id in range(len(ids)):
original_graph = graphs[start + local_id]
converted_graph = converted[local_id]
for node, attrs in [('track', ['x', 'label']), ('jet', ['x', 'y'])]:
for attr in attrs:
assert torch.equal(original_graph[node][attr], converted_graph[node][attr])
edge = ('jet', 'contains', 'track')
assert torch.equal(original_graph[edge].edge_index, converted_graph[edge].edge_index)
report['equivalent_model_inputs'] += 1
report['parquet_files'].append({'path': str(out.relative_to(PACKAGE)), 'sha256': digest(out), 'rows': len(ids)})
print(f'{variant} {split}: {len(graphs)} graphs and lossless Parquet verified', flush=True)
del ds, graphs
gc.collect()
assert report['duplicate_track_ids'] == 0, 'Ambiguous track identifiers'
assert report['equivalent_model_inputs'] == 10000
report['status'] = 'validated'
status_path.write_text(json.dumps(report, indent=2))
(PACKAGE / f'{variant}_validation.json').write_text(json.dumps(report, indent=2))
print(json.dumps({k:v for k,v in report.items() if k != 'parquet_files'}), flush=True)
if __name__ == '__main__':
main()
|