8f43927575
examples/ is based on KasperskyLab/knp at 55470ed6 with local changes to the mnist-learn altai/blifat models, cuda-backend CMake and model_loader. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
57 lines
2.4 KiB
Python
57 lines
2.4 KiB
Python
"""Choose inference threshold on validation only using the existing AltAI simulator.
|
|
|
|
Weights come from real KNP STDP training. This script performs no weight training.
|
|
The chosen configuration must be checked with real KNP before reporting test results.
|
|
"""
|
|
import argparse
|
|
import csv
|
|
import json
|
|
from pathlib import Path
|
|
import sys
|
|
|
|
import numpy as np
|
|
|
|
sys.path.insert(0, '/mnt/c/KNP/viz')
|
|
import altai_sim
|
|
|
|
altai_sim.PIX = 2048
|
|
altai_sim.CLASSES = 2
|
|
ap = argparse.ArgumentParser()
|
|
ap.add_argument('run', type=Path)
|
|
a = ap.parse_args()
|
|
data = Path('/home/int/run/cats-dogs/data')
|
|
w, original_thr = altai_sim.load_dump(a.run)
|
|
images = np.fromfile(data / 'val.bin', dtype=np.uint8).reshape(-1, 2048)
|
|
labels = np.loadtxt(data / 'val.target', dtype=int)
|
|
thresholds = np.array([10, 20, 40, 60, 80, 100, 150, 200, 255], dtype=float)
|
|
pot = np.zeros((len(thresholds), w.shape[1]))
|
|
prev_input = np.zeros_like(pot, dtype=bool)
|
|
prev_current = np.zeros(w.shape[1])
|
|
states = np.zeros(2048, dtype=np.float32)
|
|
correct = np.zeros(len(thresholds), dtype=int)
|
|
unanswered = correct.copy()
|
|
predictions = np.empty((len(images), len(thresholds)), dtype=int)
|
|
for i, (img, label) in enumerate(zip(images, labels)):
|
|
currents = altai_sim.raster(img, states) @ w
|
|
votes = np.zeros((len(thresholds), 2), dtype=int)
|
|
for t in range(15):
|
|
votes += prev_input.reshape(len(thresholds), 2, -1).any(axis=2)
|
|
pot = np.round(pot) + prev_current
|
|
pot -= 255 * (prev_input.sum(axis=1, keepdims=True) - prev_input)
|
|
fired = pot >= thresholds[:, None]
|
|
pot[fired] = 0
|
|
pot = np.maximum(pot, 0)
|
|
prev_input, prev_current = fired, currents[t]
|
|
pred = np.where(votes.max(axis=1) > 0, votes.argmax(axis=1), -1)
|
|
predictions[i] = pred
|
|
correct += pred == label
|
|
unanswered += pred == -1
|
|
best = int(np.argmax(correct)) # stable tie break: lower threshold
|
|
rows = [{'threshold': int(t), 'correct': int(c), 'unanswered': int(u),
|
|
'wrong': int(len(labels) - c - u), 'accuracy': float(c / len(labels))}
|
|
for t, c, u in zip(thresholds, correct, unanswered)]
|
|
result = {'method': 'validation-only threshold sweep; offline simulation, not KNP measurements',
|
|
'selected_threshold': int(thresholds[best]), 'rows': rows}
|
|
(a.run / 'validation-sweep.json').write_text(json.dumps(result, indent=2) + '\n')
|
|
np.save(a.run / 'validation-predictions.npy', predictions[:, best])
|
|
print(json.dumps(result), flush=True)
|