"""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 flat_cos(u, v, eps=1e-12):
    return (u.flatten() @ v.flatten() / (u.norm() * v.norm() + eps)).item()


def check_residual_local_jacobian():
    """The three-factor rule must include the residual branch's alpha."""
    torch.manual_seed(7)
    batch = 32
    net = SDILNet([20, 16, 16, 16, 5], device="cpu", seed=4, residual=True)
    x = torch.randn(batch, 20)
    y = torch.randint(0, 5, (batch,))
    for weight in net.W:
        weight.requires_grad_(True)
    fwd = net.forward(x)
    loss = torch.nn.functional.cross_entropy(fwd["h"][-1], y)
    hidden_grads = torch.autograd.grad(loss, fwd["h"][1:-1], retain_graph=True)
    weight_grads = torch.autograd.grad(loss, net.W)
    for weight in net.W:
        weight.requires_grad_(False)

    print("CHECK0 residual local-rule Jacobian:")
    for l in range(net.L - 1):
        # autograd's hidden gradient includes the mean over the batch; recover
        # per-example errors before applying the same minibatch mean as SDIL.
        error = -hidden_grads[l] * batch
        delta = error * net.act_prime(fwd["u"][l])
        if l >= 1:
            delta = net.res_alpha * delta
        local_update = delta.t() @ fwd["h"][l] / batch
        exact_update = -weight_grads[l]
        cosine = flat_cos(local_update, exact_update)
        norm_ratio = (local_update.norm() / exact_update.norm()).item()
        print(f"  layer {l}: cos={cosine:.6f} norm_ratio={norm_ratio:.6f}")
        assert cosine > 0.99999
        assert abs(norm_ratio - 1.0) < 1e-5


def main():
    torch.manual_seed(0)
    dev = "cpu"
    check_residual_local_jacobian()
    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)
    initial_alignment = probes.alignment_report(net, x, y, yoh, cfg)
    initial_mean_cos = sum(initial_alignment["cos_r_negg"]) / len(initial_alignment["cos_r_negg"])
    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}"
    final_mean_cos = sum(al["cos_r_negg"]) / len(al["cos_r_negg"])
    assert final_mean_cos > initial_mean_cos + 0.1, (
        f"trained A should improve alignment: {initial_mean_cos:.3f} -> {final_mean_cos:.3f}")

    # ---- CHECK 3: report DFA as a sanity comparator. On a single repeatedly
    # trained batch, ordinary feedback alignment can exceed learned A, so no
    # universal ordering is asserted here; CHECK2 tests that A actually learns.
    print("\nCHECK3 alignment sanity comparison: 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()