File size: 1,757 Bytes
7e9cfd1 4cbf907 7e9cfd1 4cbf907 7e9cfd1 | 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 | #!/usr/bin/env python
"""Stacking-ensemble predictions.
A per-class linear stacking head over the ensemble's per-item probability outputs.
Given the bundled probability features (probs/), it produces the label vector and
checks it against the bundled prediction file.
python predict.py --track t1
python predict.py --track t2
"""
import argparse
import json
import os
import numpy as np
HERE = os.path.dirname(os.path.abspath(__file__))
def predict(track):
spec = json.load(open(os.path.join(HERE, f"{track}_stacker.json")))
lab = spec["label_order"]
cols = spec["npy_column_order"] #
perm = [cols.index(c) for c in lab]
n = spec["n_items"]
feats = []
for name in spec["sources"]: # fixed source order
a = np.load(os.path.join(HERE, "probs", name)).astype(np.float64)[:, perm]
a = a / a.sum(1, keepdims=True)
feats.append(a)
X = np.stack(feats).transpose(1, 0, 2).reshape(n, -1) # n x 3J, matches training
W = np.array(spec["coef"]); b = np.array(spec["intercept"])
logits = X @ W.T + b
pred = [lab[i] for i in logits.argmax(1)]
ref_path = os.path.join(HERE, f"{track}_perclass_stack.txt")
ref = open(ref_path).read().split()
ok = pred == ref
out = os.path.join(HERE, f"{track}_perclass_stack_out.txt")
open(out, "w").write("\n".join(pred) + "\n")
print(f"[{track}] {len(pred)} labels -> {out} | matches bundled predictions: {ok}")
if not ok:
raise SystemExit(f"[{track}] mismatch vs bundled predictions")
if __name__ == "__main__":
ap = argparse.ArgumentParser()
ap.add_argument("--track", choices=["t1", "t2"], required=True)
predict(ap.parse_args().track)
|