File size: 2,604 Bytes
fcf4209 | 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 | """Real-checkpoint parity for the weight-free candidate selection stage."""
from pathlib import Path
import torch
from gliner2 import AutoExtractor, Schema
from gliner2.training.trainer import ExtractorCollator
from extraction_pool import select_candidates
SOURCE = (
Path.home()
/ ".cache/huggingface/hub/models--fastino--gliner2.5-multi-v1/snapshots/a221b77a8baf4a613b8f8652661d41fa10a5641e"
)
def test_candidate_pool_matches_native_with_real_weights():
torch.set_num_threads(4)
native = AutoExtractor.from_pretrained(str(SOURCE), map_location="cpu").eval()
fixtures = [
("Alice founded Acme in Toronto in 2020.", ["person", "organization", "location"]),
("Apple acquired Beats for three billion dollars.", ["company", "product", "money"]),
]
for text, labels in fixtures:
schema = Schema().entities(labels)
batch = ExtractorCollator(native.processor, is_training=False, max_len=128, architecture=native.architecture)(
[(text, schema.build())]
)
with torch.no_grad():
core = native._encode_core(batch)
head = native.boundary_head
encoded = head.boundary_encoder(core["text_states"], core["text_mask"])
marginals = head.boundary_query_head(
encoded.states,
encoded.mask,
core["text_states"],
core["text_mask"],
core["query_states"],
core["query_mask"],
)
expected = head.shared_pool_builder(
encoded.states,
encoded.mask,
core["query_mask"],
marginals.start_logits,
marginals.end_logits,
)
actual = select_candidates(
head.shared_pool_builder.start_projection(encoded.states),
head.shared_pool_builder.end_projection(encoded.states),
encoded.mask,
core["query_mask"],
marginals.start_logits,
marginals.end_logits,
boundary_top_k=head.shared_pool_builder.pool_boundary_top_k,
pool_size=head.shared_pool_builder.pool_size,
min_pool_per_query=head.shared_pool_builder.min_pool_per_query,
)
assert torch.equal(actual.indices, expected.indices)
assert torch.equal(actual.mask, expected.mask)
assert torch.allclose(actual.compat_logits, expected.compat_logits, atol=1e-5)
assert torch.allclose(actual.proposal_logits, expected.proposal_logits, atol=1e-5)
|