GNN4Colliders / tests /test_data.py
ho22joshua's picture
Claude Opus 5.5
feat: PFN prior from the ATLAS Open Data 2025 MC, held-out processes and mixed classes
9076261
Raw History Blame Contribute Delete
8.66 kB
import hashlib
import awkward as ak
import numpy as np
import pytest
import torch
import uproot
from conftest import GAMGAM_FIXTURES, TTH_FIXTURE, fixture_config, random_events
from gnn4colliders import data
from gnn4colliders.data import (
apply_derived_branches,
build_object_features,
load_events,
parse_derived_branches,
prepare,
save_events,
)
from gnn4colliders.graphs import collate_graph_batch
from gnn4colliders.loader import EventLoader
BTAG_THRESHOLD = 0.8244273
def test_object_features_are_grouped_by_type_scaled_and_typed():
arrays = {
"jet_pt": ak.Array([[30.0, 20.0], []]),
"jet_eta": ak.Array([[0.5, -1.0], []]),
"met": np.array([10.0, 40.0]),
}
features, counts = build_object_features(
arrays,
2,
[["jet_pt", "met"], ["jet_eta", 0], "CALC_E", "NODE_TYPE"],
["vector", "single"],
[0.1, 1, 0.1, 1],
)
assert counts.tolist() == [3, 1]
torch.testing.assert_close(features[:, 0], torch.tensor([3.0, 2.0, 1.0, 4.0]))
energy = torch.tensor([30 * np.cosh(0.5), 20 * np.cosh(-1.0), 10, 40]) * 0.1
torch.testing.assert_close(features[:, 2], energy.float())
assert features[:, 3].tolist() == [0, 0, 1, 1]
def test_derived_branches_split_collections_and_threshold_scores():
arrays = ak.Array(
{
"lep_pt": [[1.0, 2.0, 3.0]],
"lep_type": [[11, 13, 11]],
"jet_score": [[0.9, 0.1]],
}
)
derived = parse_derived_branches(
{
"ele_pt": {"source": "lep_pt", "where": {"lep_type": 11}},
"jet_btag": {"source": "jet_score", "above": 0.5},
}
)
result = apply_derived_branches(arrays, derived)
assert result["ele_pt"].tolist() == [[1.0, 3.0]]
assert result["jet_btag"].tolist() == [[1.0, 0.0]]
with pytest.raises(ValueError, match="depends on a derived"):
parse_derived_branches(
{"a": {"source": "x"}, "b": {"source": "a"}},
)
def test_gamgam_preparation_follows_the_open_data_schema(gamgam):
events = load_events(gamgam.data.cache)
assert len(events) == 5 * 64
assert events.labels.tolist() == [label for label in range(5) for _ in range(64)]
assert torch.equal(events.weight, torch.ones(len(events)))
offset = 0
for path in GAMGAM_FIXTURES:
raw = uproot.open(path)["mini"].arrays(library="ak")
folds = events.fold[offset : offset + 64].tolist()
assert folds == (np.asarray(raw.eventNumber, dtype=np.int64) % 10).tolist()
for local in (0, 17, 63):
event = raw[local]
first = int(events.object_offsets[offset + local])
nodes = events.object_features[
first : int(events.object_offsets[offset + local + 1])
]
electron = np.asarray(event.lep_type) == 11
muon = np.asarray(event.lep_type) == 13
pt = np.concatenate(
[
np.asarray(event.jet_pt),
np.asarray(event.lep_pt)[electron],
np.asarray(event.lep_pt)[muon],
np.asarray(event.photon_pt),
[event.met_et],
]
)
sizes = [len(event.jet_pt), electron.sum(), muon.sum()]
sizes += [len(event.photon_pt), 1]
# MeV -> units of 10 GeV, as in pretraining.
np.testing.assert_allclose(nodes[:, 0], pt * 1e-4, rtol=1e-6)
assert nodes[:, 6].tolist() == np.repeat(np.arange(5), sizes).tolist()
jets = len(event.jet_pt)
btag = np.asarray(event.jet_MV2c10) > BTAG_THRESHOLD
assert nodes[:jets, 4].tolist() == btag.astype(float).tolist()
charge = np.concatenate(
[
np.asarray(event.lep_charge)[electron],
np.asarray(event.lep_charge)[muon],
]
)
leptons = nodes[jets : jets + len(charge), 5]
assert leptons.tolist() == charge.astype(float).tolist()
assert nodes[-1, 1] == 0.0 # MET has no pseudorapidity
offset += 64
def test_parallel_and_chunked_preparation_match_serial(gamgam, tmp_path, monkeypatch):
serial = load_events(gamgam.data.cache)
monkeypatch.setattr(data, "CHUNK_EVENTS", 24)
gamgam.data.cache = str(tmp_path / "parallel.pt")
parallel = load_events(prepare(gamgam.data, workers=2))
for name in ("object_features", "object_offsets", "labels", "fold", "weight"):
assert torch.equal(getattr(parallel, name), getattr(serial, name))
assert torch.equal(parallel.entry_index, serial.entry_index)
def test_preparation_reads_the_leading_events_of_each_file(gamgam, tmp_path):
full = load_events(gamgam.data.cache)
gamgam.data.cache, gamgam.data.max_events_per_file = str(tmp_path / "c.pt"), 5
capped = load_events(prepare(gamgam.data))
files = full.file_index.unique()
assert len(capped) == 5 * len(files)
assert (capped.entry_index < 5).all()
assert torch.equal(capped.file_index.unique(), files)
def test_delphes_preparation_reads_signed_weights_and_number_folds(tmp_path):
config = fixture_config("tth_cp_even_odd", [TTH_FIXTURE], [0], tmp_path)
events = load_events(prepare(config.data))
raw = uproot.open(TTH_FIXTURE)["output"].arrays(["Number", "weight"], library="np")
assert events.fold.tolist() == (raw["Number"].astype(np.int64) % 10).tolist()
np.testing.assert_array_equal(events.weight.numpy(), raw["weight"])
assert (events.weight < 0).any()
def test_huggingface_files_are_downloaded_at_the_revision_and_verified(
tmp_path, monkeypatch
):
root_file = tmp_path / "sample.root"
root_file.write_bytes(b"fixture")
requests = []
def download(**kwargs):
requests.append(kwargs)
return str(root_file)
monkeypatch.setattr("huggingface_hub.hf_hub_download", download)
config = fixture_config("gamgam", GAMGAM_FIXTURES[:1], [0], tmp_path).data
config.hf_repo, config.hf_revision = "org/data", "abc123"
good = hashlib.sha256(b"fixture").hexdigest()
config.files = [{"path": "sample.root", "label": 4, "sha256": good}]
assert data.resolve_files(config) == [(root_file, 4)]
assert requests[0]["revision"] == "abc123"
config.files[0]["sha256"] = hashlib.sha256(b"wrong").hexdigest()
with pytest.raises(ValueError, match="SHA-256 mismatch"):
data.resolve_files(config)
def test_cache_round_trip_and_version_check(tmp_path):
events = random_events()
loaded = load_events(save_events(events, tmp_path / "e.pt"))
assert torch.equal(loaded.object_features, events.object_features)
payload = torch.load(tmp_path / "e.pt")
payload["cache_schema_version"] = 1
torch.save(payload, tmp_path / "old.pt")
with pytest.raises(ValueError, match="cache version"):
load_events(tmp_path / "old.pt")
def test_shuffled_order_depends_only_on_seed_and_epoch():
events = random_events()
loader = EventLoader(events, collate_graph_batch, 8, shuffle=True, seed=3)
first = loader.order()
assert torch.equal(first, loader.order())
loader.set_epoch(1)
assert not torch.equal(first, loader.order())
assert sorted(loader.order().tolist()) == list(range(len(events)))
def test_ranks_split_training_with_padding_and_evaluation_without():
events = random_events(41)
def orders(shuffle):
return [
EventLoader(
events,
collate_graph_batch,
8,
shuffle=shuffle,
rank=rank,
world_size=2,
).order()
for rank in range(2)
]
train = orders(True)
assert len(train[0]) == len(train[1]) == 21
assert set(torch.cat(train).tolist()) == set(range(41))
evaluation = orders(False)
assert sorted(torch.cat(evaluation).tolist()) == list(range(41))
def test_loader_workers_yield_identical_pinned_batches():
events = random_events(seed=4)
serial = EventLoader(events, collate_graph_batch, 7, shuffle=True, seed=3)
workers = EventLoader(
events,
collate_graph_batch,
7,
shuffle=True,
seed=3,
num_workers=2,
pin_memory=torch.cuda.is_available(),
)
pairs = list(zip(serial, workers, strict=True))
assert len(pairs) == len(serial) == 6
for left, right in pairs:
assert torch.equal(left.graph.edge_features, right.graph.edge_features)
assert torch.equal(left.labels, right.labels)
assert torch.equal(left.event_index, right.event_index)