diff options
Diffstat (limited to 'experiments/smoke.py')
| -rw-r--r-- | experiments/smoke.py | 86 |
1 files changed, 86 insertions, 0 deletions
diff --git a/experiments/smoke.py b/experiments/smoke.py new file mode 100644 index 0000000..f2efa87 --- /dev/null +++ b/experiments/smoke.py @@ -0,0 +1,86 @@ +"""Fast CPU correctness checks for SDIL mechanics and signs. +Run: python experiments/smoke.py +""" +import os +import sys +import torch + +sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) +from sdil.core import SDILNet, SDILConfig, sdil_step, node_perturbation_targets +from sdil.baselines import dfa_config +from sdil import probes +from sdil.data import get_dataset, onehot + + +def row_cos(u, v, eps=1e-12): + return ((u * v).sum(1) / (u.norm(dim=1) * v.norm(dim=1) + eps)).mean().item() + + +def main(): + torch.manual_seed(0) + dev = "cpu" + print("loading MNIST subset...") + tr, te, n_in, n_out = get_dataset("mnist", batch_size=128, device=dev) + xb, yb = next(iter(tr)) + x = xb[:256] + y = yb[:256] + yoh = onehot(y, 10) + + sizes = [784, 64, 64, 64, 10] + + # ---- CHECK 1: node-perturbation q estimates the descent direction -g ---- + net = SDILNet(sizes, device=dev, seed=1) + grads, _ = probes.true_hidden_grads(net, x, y) + for ndirs in (1, 8, 32): + qs = node_perturbation_targets(net, x, y, sigma=1e-2, n_dirs=ndirs) + cs = [row_cos(qs[l], -grads[l]) for l in range(len(qs))] + print(f"CHECK1 node-pert n_dirs={ndirs:2d} cos(q,-g) per layer = " + + " ".join(f"{c:+.3f}" for c in cs)) + assert all(c > 0 for c in cs), "node perturbation q should align with -grad" + + # ---- CHECK 2: SDIL overfits a fixed batch and alignment climbs ---- + net = SDILNet(sizes, device=dev, seed=2) + cfg = SDILConfig(eta=0.1, eta_A=0.05, eta_P=0.005, pert_every=2, pert_ndirs=8, + momentum=0.9) + print("\nCHECK2 SDIL overfit fixed batch:") + for step in range(401): + loss, _ = sdil_step(net, x, y, yoh, cfg, step) + if step % 50 == 0: + al = probes.alignment_report(net, x, y, yoh, cfg) + mc = sum(al["cos_r_negg"]) / len(al["cos_r_negg"]) + mca = sum(al["cos_apical_negg"]) / len(al["cos_apical_negg"]) + print(f" step {step:3d} loss {loss:.4f} mean_cos(r,-g) {mc:+.3f} " + f"mean_cos(apical,-g) {mca:+.3f} per-layer_r {['%.2f'%v for v in al['cos_r_negg']]}") + assert loss < 1.5, f"SDIL should reduce loss on fixed batch, got {loss}" + + # ---- CHECK 3: trained-A SDIL beats fixed-random-A (DFA) on alignment ---- + print("\nCHECK3 alignment: SDIL(trained A) vs DFA(fixed A):") + net_dfa = SDILNet(sizes, device=dev, seed=3) + cfg_dfa = dfa_config(eta=0.1, momentum=0.9) + for step in range(401): + sdil_step(net_dfa, x, y, yoh, cfg_dfa, step) + al_dfa = probes.alignment_report(net_dfa, x, y, yoh, cfg_dfa) + al_sdil = probes.alignment_report(net, x, y, yoh, cfg) + print(f" DFA mean_cos(r,-g) = {sum(al_dfa['cos_r_negg'])/len(al_dfa['cos_r_negg']):+.3f}") + print(f" SDIL mean_cos(r,-g) = {sum(al_sdil['cos_r_negg'])/len(al_sdil['cos_r_negg']):+.3f}") + + # ---- CHECK 4: nuisance -> residual preserves alignment, raw apical degrades ---- + print("\nCHECK4 residualization under soma-predictable nuisance (rho=2.0):") + net_n = SDILNet(sizes, device=dev, seed=4, nuis_rho=2.0) + cfg_n = SDILConfig(eta=0.1, eta_A=0.05, eta_P=0.02, pert_every=2, pert_ndirs=8, momentum=0.9) + for step in range(401): + sdil_step(net_n, x, y, yoh, cfg_n, step) + al_n = probes.alignment_report(net_n, x, y, yoh, cfg_n) + print(f" residual cos(r,-g) = {sum(al_n['cos_r_negg'])/len(al_n['cos_r_negg']):+.3f}") + print(f" raw apical cos(a,-g) = {sum(al_n['cos_apical_negg'])/len(al_n['cos_apical_negg']):+.3f}") + print(f" pure A c cos(Ac,-g) = {sum(al_n['cos_Ac_negg'])/len(al_n['cos_Ac_negg']):+.3f}") + + # ---- CHECK 5: single-step loss-decrease ratio vs exact GD ---- + print("\nCHECK5 single-step loss-decrease ratio vs exact GD:") + ldr = probes.loss_decrease_ratio(net, x, y, yoh, cfg, step=100) + print(f" dL_sdil={ldr['dL_sdil']:+.4f} dL_bp={ldr['dL_bp']:+.4f} ratio={ldr['ratio']:+.3f}") + print("\nALL SMOKE CHECKS PASSED") + + +if __name__ == "__main__": + main() |