SPIRA v2: entropy-patch tokenization + hierarchical dopamine-gated spiking cortex + hippocampal one-shot memory with replay consolidation
032a961 verified | #!/usr/bin/env python3 | |
| """Probe: sharp sparse input drive -> token-discriminative spike patterns.""" | |
| 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=40_000, am_tokens=1000, epochs=1) | |
| model = SPIRA.build(data, cfg, log=lambda m: None) | |
| ids = model._ids_train | |
| N = 40_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}") | |
| variants = [ | |
| # (w_in_scale, in_fanout, w_rec_scale, rec_fanout, inh_frac) | |
| (3.0, 8, 0.35, 48, 0.50), | |
| (2.5, 12, 0.35, 48, 0.50), | |
| (4.0, 8, 0.35, 48, 0.50), | |
| (3.0, 8, 0.25, 24, 0.50), | |
| (5.0, 6, 0.35, 48, 0.50), | |
| ] | |
| def pattern_discrimination(core, neurons, n=400, stride=17): | |
| """Mean Jaccard overlap of spike patterns for tokens at distant positions | |
| with DIFFERENT patch ids; low overlap = high token information.""" | |
| core.reset() | |
| pats = [] | |
| id_seen = [] | |
| for t in range(0, n * stride, stride): | |
| core.step(neurons[t]) | |
| pats.append(set(np.nonzero(core.s)[0].tolist())) | |
| id_seen.append(int(neurons[t][0])) # proxy: first neuron differs | |
| overlaps = [] | |
| for i in range(len(pats) - 1): | |
| for j in range(i + 1, min(i + 5, len(pats))): | |
| a, b = pats[i], pats[j] | |
| if a or b: | |
| overlaps.append(len(a & b) / max(len(a | b), 1)) | |
| return float(np.mean(overlaps)) | |
| for w_in, fan, w_rec, recf, inh in variants: | |
| core = SpikingCore(n_out=cfg.vocab_size + 1, seed=cfg.seed, | |
| w_in_scale=w_in, in_fanout=fan, | |
| w_rec_scale=w_rec, rec_fanout=recf, inh_frac=inh, | |
| eta_r=0.002) | |
| core.reset() | |
| sp = 0 | |
| for t in range(2000): | |
| core.step(neurons[t]) | |
| sp += int(core.s.sum()) | |
| rate = sp / 2000 / core.n_res | |
| jac = pattern_discrimination(core, neurons) | |
| res = core.train_pass(neurons[: N - 1], cols[1:N], log_every=20_000, log_fn=lambda d: None) | |
| print(f"w_in={w_in:.1f} fan={fan} w_rec={w_rec:.2f} recf={recf} | rate={rate:.3f} " | |
| f"jac={jac:.3f} overall={res['accuracy']:.4f} in-vocab~={res['accuracy']/cov:.4f}") | |