SPIRA v2: entropy-patch tokenization + hierarchical dopamine-gated spiking cortex + hippocampal one-shot memory with replay consolidation
032a961 verified | #!/usr/bin/env python3 | |
| """A/B test: ORIGINAL reservoir params (rich recurrent states) with honest | |
| vocab-masked metrics, with/without trace centering, various eta_r.""" | |
| import sys | |
| import numpy as np | |
| sys.path.insert(0, "/home/z/my-project/download/spira") | |
| from spira import SPIRA, SPIRAConfig | |
| from spira.spiking import SpikingCore | |
| data = open("/home/z/my-project/data/train.bin", "rb").read() | |
| cfg = SPIRAConfig(snn_tokens=60_000, am_tokens=1000, epochs=1) | |
| model = SPIRA.build(data, cfg, log=lambda m: None) | |
| ids = model._ids_train | |
| N = 60_000 | |
| neurons = SPIRA.input_neurons_stream(ids[:N], cfg.d_in, cfg.k_in) | |
| cols = model.cols_of(ids[: N + 1]) | |
| cov = (cols[1:] != cfg.vocab_size)[: N - 1].mean() | |
| print(f"coverage: {cov:.3f}") | |
| # ORIGINAL params that reached EMA 0.52-0.82 (with the UNK shortcut metric) | |
| variants = [ | |
| ("sharp+delta e=.70", dict(w_in_scale=3.0, in_fanout=8, w_rec_scale=0.25, rec_fanout=24, inh_frac=0.5, eta_delta=0.70), True), | |
| ("sharp+delta e=1.0", dict(w_in_scale=3.0, in_fanout=8, w_rec_scale=0.25, rec_fanout=24, inh_frac=0.5, eta_delta=1.0), True), | |
| ("sharp+delta e=1.5", dict(w_in_scale=3.0, in_fanout=8, w_rec_scale=0.25, rec_fanout=24, inh_frac=0.5, eta_delta=1.5), True), | |
| ("sharp+delta e=1.0 tau.3", dict(w_in_scale=3.0, in_fanout=8, w_rec_scale=0.25, rec_fanout=24, inh_frac=0.5, eta_delta=1.0, tau_trace=0.35), True), | |
| ] | |
| for name, kw, center in variants: | |
| core = SpikingCore(n_out=cfg.vocab_size + 1, seed=cfg.seed, **kw) | |
| if not center: | |
| # disable centering by pinning x_mean to zero | |
| class NoCenter: | |
| pass | |
| core._center = False | |
| orig_step = core.step | |
| import types | |
| # monkeypatch: after step, reset x_mean to 0 | |
| def step_noc(self, input_idx, _orig=orig_step): | |
| x = _orig(input_idx) | |
| self.x_mean = np.zeros_like(self.x_mean) | |
| return x | |
| core.step = types.MethodType(step_noc, core) | |
| res = core.train_pass(neurons[: N - 1], cols[1:N], log_every=30_000, log_fn=lambda d: None) | |
| print(f"{name:28s} overall={res['accuracy']:.4f} in-vocab~={res['accuracy']/cov:.4f} " | |
| f"ema={res['acc_ema_final']:.4f}") | |