Add resumable SigLIP2 Giant formal training
CI / test (push) Successful in 13s

This commit is contained in:
2026-10-01 23:40:40 +08:00
parent 521dfb3d57
commit bfb96b76c7
11 changed files with 341 additions and 27 deletions
+21 -11
View File
@@ -90,19 +90,24 @@ def summarize(counts: dict[str, int]) -> dict[str, float]:
}
def null_slot_counts(
def slot_type_counts(
predictions: dict[str, torch.Tensor], targets: list[torch.Tensor]
) -> tuple[int, int]:
) -> tuple[int, int, int, int]:
matching = hungarian_matching_loss(predictions, targets)
predicted_types = predictions["type_logits"].argmax(-1)
correct = 0
total = predicted_types.numel()
type_correct = 0
type_total = predicted_types.numel()
null_correct = 0
null_total = 0
for batch_index, (slots, target_indices) in enumerate(matching.assignments):
expected = torch.full_like(predicted_types[batch_index], NULL_TYPE)
if len(slots):
expected[slots] = targets[batch_index].to(expected.device)[target_indices, 0]
correct += int((predicted_types[batch_index] == expected).sum())
return correct, total
type_correct += int((predicted_types[batch_index] == expected).sum())
null_positions = expected == NULL_TYPE
null_correct += int((predicted_types[batch_index][null_positions] == NULL_TYPE).sum())
null_total += int(null_positions.sum())
return null_correct, null_total, type_correct, type_total
@torch.no_grad()
@@ -118,6 +123,8 @@ def evaluate(
}
null_correct = 0
null_total = 0
type_correct = 0
type_total = 0
for batch in loader:
outputs = model(batch["images"].to(device), batch["rules_text"])
requests = (
@@ -134,11 +141,14 @@ def evaluate(
("visual", "visual_targets"),
("text", "text_targets"),
):
correct, count = null_slot_counts(outputs[modality], batch[target_name])
null_correct += correct
null_total += count
counts = slot_type_counts(outputs[modality], batch[target_name])
null_correct += counts[0]
null_total += counts[1]
type_correct += counts[2]
type_total += counts[3]
result = {name: summarize(counts) for name, counts in totals.items()}
result["null_slot_accuracy"] = null_correct / max(1, null_total)
result["slot_type_accuracy"] = type_correct / max(1, type_total)
return result
@@ -146,14 +156,14 @@ def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--config", type=Path, default=Path("configs/mvp.yaml"))
parser.add_argument("--checkpoint", type=Path, required=True)
parser.add_argument("--backbone", choices=("tiny", "clip"), default=None)
parser.add_argument("--backbone", choices=("tiny", "clip", "siglip2"), default=None)
parser.add_argument("--split", default="test_id")
args = parser.parse_args()
config = yaml.safe_load(args.config.read_text(encoding="utf-8"))
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = EvidenceExtractorModel(config, backbone_name=args.backbone).to(device)
state = torch.load(args.checkpoint, map_location=device, weights_only=True)
model.load_state_dict(state["model"])
model.load_state_dict(state["model"], strict=False)
dataset = EvidenceDataset(
Path(config["data"]["root"]) / f"{args.split}.jsonl",
config["data"]["image_size"],