File size: 4,271 Bytes
3f316da | 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 | """Run a small CPU check using one downloaded public asset; no paper scores."""
import argparse
import json
import math
import sys
import tempfile
from pathlib import Path
import numpy as np
import torch
ROOT = Path(__file__).resolve().parent
sys.path.insert(0, str(ROOT / "experiments/tbd_revision/scripts"))
from run_architecture_task_coverage import MODEL_ORDER, build_encoder
from run_representation_task_suite import load_normalization, task_loss
from train_downstream import load_npz
from train_representation_baselines import (
ReconstructionPretrainer,
build_method,
completed_result,
result_path_for,
run_one,
)
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--data-dir", type=Path, default=ROOT.parent / "deidentified_50stock_panel")
args = parser.parse_args()
torch.set_num_threads(2)
normalization = load_normalization(args.data_dir, "asset_001")
with np.load(args.data_dir / "asset_001.npz", allow_pickle=False) as original:
parts = [original[key][:count] for key, count in
[("natural_train_idx", 12), ("val_idx", 6), ("test_idx", 6)]]
selected = np.concatenate(parts)
assert len(set(selected.tolist())) == 24, "smoke partitions overlap or are too small"
arrays = {key: original[key][selected] for key in
["x", "y", "future_returns", "relative_day_id", "seconds_since_midnight"]}
assert arrays["x"].shape == (24, 100, 40)
assert np.isfinite(arrays["x"]).all()
arrays.update(natural_train_idx=np.arange(12), val_idx=np.arange(12, 18),
test_idx=np.arange(18, 24))
with tempfile.TemporaryDirectory(prefix="lobench-smoke-") as directory:
work = Path(directory)
np.savez_compressed(work / "asset_001.npz", **arrays)
(work / "manifest.json").write_text(json.dumps({"assets": [
{"alias": "asset_001", "normalization_training_only": normalization}
]}))
x, _, splits = load_npz(work / "asset_001.npz")
assert len(splits["train"]) == 12
config = {
"processed_dir": str(work), "dataset_manifest": str(work / "manifest.json"),
"data_file_template": "{stock}.npz", "results_dir": str(work / "results"),
"pretrain_scope": "local", "pretrain_index_key": "natural_train_idx",
"downstream_index_key": "natural_train_idx",
"training": {"batch_size": 6, "pretrain_epochs": 1, "pretrain_patience": 0,
"downstream_epochs": 1, "downstream_patience": 0,
"learning_rate": 3e-4, "weight_decay": 1e-5,
"latent_dim": 16, "d_model": 16, "nhead": 4,
"mask_ratio": 0.4, "device": "cpu", "num_workers": 0,
"class_weights": "balanced", "log_epochs": False},
}
settings = ["from_scratch", "frozen_probe", "fine_tune"]
for method in ["timae", "lobgrid_timae", "simlob2026"]:
run_one(config, "asset_001", method, 2026, 1.0, settings)
for setting in settings:
output = result_path_for(config, "asset_001", method, setting, 2026, 1.0)
assert completed_result(output), output
result = json.loads(output.read_text())
assert math.isfinite(result["test"]["macro_f1"])
model = ReconstructionPretrainer(build_method(method, 16, 16), d_model=16)
for task in ["reconstruction", "imputation"]:
loss = task_loss(model, x[:2], task, method, structure_weight=0)
assert torch.isfinite(loss), (method, task)
loss.backward()
model.zero_grad()
for name in MODEL_ORDER:
encoder = build_encoder(name, latent_dim=16, hidden_dim=16).eval()
with torch.no_grad():
output = encoder(x[:2])
assert output.shape == (2, 16) and torch.isfinite(output).all(), name
print("PASS: public data loading; 3 encoders × 3 prediction settings; recovery losses; 9 architecture adapters.")
print("This smoke test uses 24 windows and one epoch; it does not reproduce paper scores.")
if __name__ == "__main__":
main()
|