SPIRA v2: entropy-patch tokenization + hierarchical dopamine-gated spiking cortex + hippocampal one-shot memory with replay consolidation
032a961 verified | #!/usr/bin/env python3 | |
| """Rebuild the corpus hippocampal AM with the FIXED save format. | |
| The pre-fix checkpoints packed/rebuilt AM entries at wrong addresses | |
| (load wrote at k[lane]%size instead of the derivable primary/LOO slots), | |
| making the loaded corpus AM invisible to queries. This script resets the AM, | |
| re-runs the corpus write pass in-memory, and saves with the corrected | |
| packing (dedup at derivable slots). | |
| """ | |
| import sys | |
| import time | |
| sys.path.insert(0, "/home/z/my-project/download/spira") | |
| import numpy as np | |
| from spira import SPIRA | |
| from spira.hippocampus import Hippocampus | |
| BASE = "/home/z/my-project" | |
| CKPT = f"{BASE}/download/spira/artifacts/spira_model.npz" | |
| m = SPIRA.load(CKPT) | |
| print(f"loaded (old AM: {m.am.stats()})") | |
| m.am = Hippocampus(m=m.cfg.am_m, size=m.cfg.am_size, min_lanes=m.cfg.am_min_lanes) | |
| data = open(f"{BASE}/data/train.bin", "rb").read() | |
| patches = m.ept.segment(data) | |
| ids = np.fromiter((__import__("spira.hashing", fromlist=["fnv1a64"]).fnv1a64(p) | |
| for p in patches), dtype=np.uint64, count=len(patches)) | |
| t0 = time.perf_counter() | |
| taught = m.teach_ids(ids[:1_200_000], episodic=False) | |
| print(f"corpus AM write: {taught['associations']} associations in " | |
| f"{time.perf_counter()-t0:.0f}s -> {m.am.stats()}") | |
| t0 = time.perf_counter() | |
| m.save(CKPT) | |
| print(f"saved (fixed format) in {time.perf_counter()-t0:.0f}s") | |
| # roundtrip validation | |
| m2 = SPIRA.load(CKPT) | |
| st = m2.am.stats() | |
| print(f"reloaded AM: {st}") | |
| v, w = m2.am.query(m2.am.lane_keys(ids[100:104])) | |
| print(f"query ctx[100:104] -> match={v == ids[104]} (w={w:.2f})") | |
| hits = 0 | |
| for i in range(500, 1000): | |
| v, w = m2.am.query(m2.am.lane_keys(ids[i : i + 4])) | |
| hits += int(v == ids[i + 4]) | |
| print(f"roundtrip retrieval on 500 corpus contexts: {hits/5:.1f}%") | |