tests/unit/sequences/test_sequence_sample.py

import torch

from gnn4colliders.data import EventMetadata, EventSample
from gnn4colliders.sequences import (
    SequenceDataLoader,
    SequenceSample,
    batch_sequence_samples,
    sequence_sample_from_event,
)


def _metadata(index: int) -> EventMetadata:
    return EventMetadata(fold=index, weight=1.0, sample_id=f"event:{index}")


def test_sequence_feature_builder_reuses_shared_event_features():
    event = EventSample(
        objects={
            "pt": [10.0, 20.0],
            "eta": [0.1, -0.2],
            "phi": [0.3, -0.4],
        },
        label=1,
        global_features=torch.empty(0),
        event_index=0,
        metadata=_metadata(0),
    )
    sample = sequence_sample_from_event(
        event,
        [["pt"], ["eta"], ["phi"], "CALC_E", [1.0], [0.0], "NODE_TYPE"],
        ["vector"],
        [1.0] * 7,
    )
    assert sample.tokens.shape == (2, 7)
    assert sample.label.item() == 1
    assert sample.metadata.sample_id == "event:0"


def test_sequence_batch_pads_and_masks_in_source_order():
    samples = [
        SequenceSample(
            tokens=torch.ones(2, 3),
            label=torch.tensor(1),
            global_features=None,
            metadata=_metadata(1),
        ),
        SequenceSample(
            tokens=torch.ones(1, 3),
            label=torch.tensor(0),
            global_features=None,
            metadata=_metadata(2),
        ),
    ]
    batch = batch_sequence_samples(samples)
    assert batch.tokens.shape == (2, 2, 3)
    assert batch.token_mask.tolist() == [[True, True], [True, False]]
    assert batch.metadata.sample_id == ("event:1", "event:2")

    loader = SequenceDataLoader(samples, batch_size=1, shuffle=True, seed=9)
    loader.set_epoch(2)
    assert len(list(loader)) == 2