path: root/experiments/freeze_with_decay.py

"""
Phase 10A.8A: Freeze with Alpha Decay.

Core question: After freezing Vec, can linearly decaying alpha (fading out the
frozen Vec and returning to pure DFA) recover or improve over a fixed-alpha frozen blend?

8 branches from the same DFA checkpoint at t0=5:
1. continue_DFA                    — pure DFA baseline
2. blend_random_trainable_alpha075 — standard reference (always trainable, alpha=0.75)
3. freeze_after_1_fixed075         — train Vec 1 epoch, freeze, keep alpha=0.75
4. freeze_after_5_fixed075         — train Vec 5 epochs, freeze, keep alpha=0.75
5. freeze_after_1_decay_to_025     — train Vec 1 epoch, freeze, then decay alpha 0.75->0.25 over 5 epochs
6. freeze_after_5_decay_to_025     — train Vec 5 epochs, freeze, then decay alpha 0.75->0.25 over 5 epochs
7. freeze_after_1_decay_to_000     — train Vec 1 epoch, freeze, then decay alpha 0.75->0.0 over 5 epochs
8. freeze_after_5_decay_to_000     — train Vec 5 epochs, freeze, then decay alpha 0.75->0.0 over 5 epochs
"""
import os
import sys
import json
import argparse
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
import torchvision
import torchvision.transforms as transforms
import copy

sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))

from models.residual_mlp import ResidualMLP
from models.value_net import SinusoidalTimeEmbed
from metrics.credit_metrics import cosine_similarity_batch, perturbation_correlation


# ---------------------------------------------------------------------------
# Auxiliary network
# ---------------------------------------------------------------------------

class VectorCreditNet(nn.Module):
    """Standard Vec: takes (h, t, s) -> d_hidden credit vector."""
    def __init__(self, d_hidden, s_dim, time_embed_dim=32, hidden_dim=256, num_layers=3):
        super().__init__()
        self.ln = nn.LayerNorm(d_hidden)
        self.time_embed = SinusoidalTimeEmbed(time_embed_dim)
        input_dim = d_hidden + time_embed_dim + s_dim
        layers = []
        for i in range(num_layers):
            in_d = input_dim if i == 0 else hidden_dim
            layers.append(nn.Linear(in_d, hidden_dim))
            layers.append(nn.GELU())
        layers.append(nn.Linear(hidden_dim, d_hidden))
        self.net = nn.Sequential(*layers)

    def forward(self, h, t, s):
        return self.net(torch.cat([self.ln(h), self.time_embed(t), s], dim=-1))


# ---------------------------------------------------------------------------
# Alpha schedule helpers
# ---------------------------------------------------------------------------

def make_alpha_schedule(freeze_epoch, initial_alpha, target_alpha, decay_window):
    """
    Returns a function alpha_fn(epoch, t0) -> current alpha.

    Before freeze_epoch training epochs have passed, alpha = initial_alpha.
    After freeze_epoch training epochs, linearly decay from initial_alpha to
    target_alpha over decay_window epochs, then stay at target_alpha.

    epoch is the absolute epoch number; t0 is the DFA checkpoint epoch.
    Training epochs elapsed since handoff = epoch - t0.
    """
    def alpha_fn(epoch, t0):
        elapsed = epoch - t0  # epochs since handoff (1-indexed)
        if elapsed <= freeze_epoch:
            return initial_alpha
        # epochs after freeze
        after_freeze = elapsed - freeze_epoch
        if decay_window <= 0 or target_alpha == initial_alpha:
            return target_alpha
        progress = min(after_freeze / decay_window, 1.0)
        return initial_alpha + (target_alpha - initial_alpha) * progress
    return alpha_fn


# ---------------------------------------------------------------------------
# Data
# ---------------------------------------------------------------------------

def get_cifar10(batch_size=128):
    transform_train = transforms.Compose([
        transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
        transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616))])
    transform_test = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616))])
    trainset = torchvision.datasets.CIFAR10(
        root='./data', train=True, download=True, transform=transform_train)
    testset = torchvision.datasets.CIFAR10(
        root='./data', train=False, download=True, transform=transform_test)
    return (DataLoader(trainset, batch_size=batch_size, shuffle=True,
                       num_workers=4, pin_memory=True),
            DataLoader(testset, batch_size=batch_size, shuffle=False,
                       num_workers=4, pin_memory=True))


# ---------------------------------------------------------------------------
# Evaluation helpers
# ---------------------------------------------------------------------------

def evaluate(model, test_loader, device):
    model.eval(); c, t = 0, 0
    with torch.no_grad():
        for x, y in test_loader:
            x = x.view(x.size(0), -1).to(device); y = y.to(device)
            c += (model(x).argmax(1) == y).sum().item(); t += x.size(0)
    return c / t


def compute_diagnostics(model, aux_net, Bs, test_loader, device, credit_mode, alpha=0.75):
    """Compute mean Gamma (BP cosine) and mean rho (perturbation correlation)."""
    model.eval()
    if aux_net is not None:
        aux_net.eval()
    L = model.num_blocks

    for x, y in test_loader:
        x = x.view(x.size(0), -1).to(device); y = y.to(device); break
    batch = x.size(0)

    # BP pass for hidden gradients (offline eval only, not used for training)
    was_frozen = not next(model.parameters()).requires_grad
    if was_frozen:
        for p in model.parameters(): p.requires_grad_(True)
    model.zero_grad()
    lo, hbp = model(x, return_hidden=True)
    for l in range(L + 1): hbp[l].retain_grad()
    F.cross_entropy(lo, y).backward()
    bp = {l: hbp[l].grad.detach().clone() for l in range(L + 1)}
    if was_frozen:
        for p in model.parameters(): p.requires_grad_(False)

    with torch.no_grad():
        lo2, hi = model(x, return_hidden=True)
        eT = lo2.softmax(-1); eT[torch.arange(batch), y] -= 1; s = eT.detach()

    gammas, rhos = [], []
    for l in range(L):
        h_l = hi[l].detach()
        t_l = torch.full((batch,), l / L, device=device)

        if credit_mode == 'dfa':
            a_l = (s @ Bs[l].T).detach()
        elif credit_mode == 'blend' and aux_net is not None:
            a_dfa = (s @ Bs[l].T).detach()
            a_aux = aux_net(h_l, t_l, s).detach()
            rd = (a_dfa ** 2).mean(-1, keepdim=True).sqrt() + 1e-6
            rv = (a_aux ** 2).mean(-1, keepdim=True).sqrt() + 1e-6
            a_l = alpha * a_aux / rv + (1 - alpha) * a_dfa / rd
        else:
            a_l = (s @ Bs[l].T).detach()

        gammas.append(cosine_similarity_batch(a_l, bp[l]))

        def make_fwd(sl):
            def f(h):
                with torch.no_grad():
                    c = h
                    for i in range(sl, L):
                        c = c + model.blocks[i](c)
                    return F.cross_entropy(
                        model.out_head(model.out_ln(c)), y, reduction='none')
            return f

        rhos.append(perturbation_correlation(h_l, a_l, make_fwd(l), epsilon=1e-3, M=16))

    return float(np.mean(gammas)), float(np.mean(rhos))


# ---------------------------------------------------------------------------
# DFA training + checkpoint
# ---------------------------------------------------------------------------

def train_dfa_get_checkpoint(model, train_loader, test_loader, device,
                              total_epochs, t0, lr, wd):
    d = model.d_hidden; L = model.num_blocks
    Bs = [torch.randn(d, 10, device=device) / np.sqrt(10) for _ in range(L)]
    block_opts = [optim.AdamW(b.parameters(), lr=lr, weight_decay=wd)
                  for b in model.blocks]
    embed_opt = optim.AdamW(model.embed.parameters(), lr=lr, weight_decay=wd)
    head_opt = optim.AdamW(
        list(model.out_head.parameters()) + list(model.out_ln.parameters()),
        lr=lr, weight_decay=wd)
    scheds = ([optim.lr_scheduler.CosineAnnealingLR(o, T_max=total_epochs)
               for o in block_opts] +
              [optim.lr_scheduler.CosineAnnealingLR(embed_opt, T_max=total_epochs),
               optim.lr_scheduler.CosineAnnealingLR(head_opt, T_max=total_epochs)])
    ckpt = None
    for epoch in range(1, total_epochs + 1):
        model.train(); tl, c, t = 0, 0, 0
        for x, y in train_loader:
            x = x.view(x.size(0), -1).to(device); y = y.to(device); b = x.size(0)
            with torch.no_grad():
                lo, hi = model(x, return_hidden=True); lv = F.cross_entropy(lo, y)
                eT = lo.softmax(-1); eT[torch.arange(b), y] -= 1
            hL = hi[-1].detach()
            lo2 = F.cross_entropy(model.out_head(model.out_ln(hL)), y)
            head_opt.zero_grad(); lo2.backward(); head_opt.step()
            for l in range(L):
                a = (eT @ Bs[l].T).detach()
                rm = (a ** 2).mean(-1, keepdim=True).sqrt() + 1e-6
                f = model.blocks[l](hi[l].detach())
                ll = (f * (a / rm)).sum(-1).mean()
                block_opts[l].zero_grad(); ll.backward()
                torch.nn.utils.clip_grad_norm_(model.blocks[l].parameters(), 1.0)
                block_opts[l].step()
            a0 = (eT @ Bs[0].T).detach()
            r0 = (a0 ** 2).mean(-1, keepdim=True).sqrt() + 1e-6
            el = (model.embed(x) * (a0 / r0)).sum(-1).mean()
            embed_opt.zero_grad(); el.backward(); embed_opt.step()
            tl += lv.item() * b; c += (lo.argmax(1) == y).sum().item(); t += b
        for s in scheds: s.step()
        if epoch == t0:
            acc = evaluate(model, test_loader, device)
            ckpt = {'model': copy.deepcopy(model.state_dict()),
                    'Bs': [B.clone() for B in Bs], 'acc': acc}
            print(f"  [DFA] Checkpoint at epoch {t0}: acc={acc:.4f}")
        if epoch % 10 == 0:
            print(f"  [DFA] Epoch {epoch}: acc={evaluate(model, test_loader, device):.4f}")
    return Bs, ckpt


# ---------------------------------------------------------------------------
# Branch runner
# ---------------------------------------------------------------------------

def run_branch(model, aux_net, Bs, train_loader, test_loader, device,
               t0, total_epochs, branch_type, alpha_schedule_fn,
               lr, lr_fb, wd, M, branch_name='', freeze_epoch=None):
    """
    Run a training branch from a loaded checkpoint.

    branch_type:
      'dfa'          — pure DFA, no aux
      'blend'        — blend DFA + Vec; aux_net trained online if vec_opt active
      'blend_frozen' — blend DFA + frozen Vec; Vec trained for freeze_epoch epochs then frozen

    alpha_schedule_fn(epoch, t0) -> float: returns alpha at each absolute epoch.
    freeze_epoch: int — for 'blend_frozen', number of epochs to train Vec before freezing.
    """
    d = model.d_hidden; L = model.num_blocks; eps_pert = 1e-3

    block_opts = [optim.AdamW(b.parameters(), lr=lr, weight_decay=wd)
                  for b in model.blocks]
    embed_opt = optim.AdamW(model.embed.parameters(), lr=lr, weight_decay=wd)
    head_opt = optim.AdamW(
        list(model.out_head.parameters()) + list(model.out_ln.parameters()),
        lr=lr, weight_decay=wd)

    if branch_type != 'dfa' and aux_net is not None:
        vec_opt = optim.Adam(aux_net.parameters(), lr=lr_fb)
    else:
        vec_opt = None

    scheds = ([optim.lr_scheduler.CosineAnnealingLR(o, T_max=total_epochs)
               for o in block_opts] +
              [optim.lr_scheduler.CosineAnnealingLR(embed_opt, T_max=total_epochs),
               optim.lr_scheduler.CosineAnnealingLR(head_opt, T_max=total_epochs)])
    # Advance schedulers to match checkpoint epoch
    for _ in range(t0):
        for s in scheds: s.step()

    log = {'test_acc': [], 'train_loss': [], 'gamma': [], 'rho': [], 'alpha_eff': []}
    diag_epochs = set(
        list(range(t0 + 1, min(t0 + 6, total_epochs + 1))) +
        [t0 + 8, t0 + 10, t0 + 15, t0 + 20] +
        list(range(t0 + 10, total_epochs + 1, 10)) +
        [total_epochs])

    vec_frozen = False  # whether Vec has been frozen

    for epoch in range(t0 + 1, total_epochs + 1):
        # Handle freeze: freeze Vec after freeze_epoch training epochs
        if (branch_type == 'blend_frozen' and freeze_epoch is not None
                and not vec_frozen):
            elapsed = epoch - t0  # training epochs since handoff (1-indexed)
            if elapsed > freeze_epoch:
                if aux_net is not None:
                    aux_net.requires_grad_(False)
                    aux_net.eval()
                vec_opt = None
                vec_frozen = True
                print(f"    [{branch_name}] Freezing Vec at epoch {epoch} "
                      f"(after {freeze_epoch} training epochs)")

        # Compute alpha for this epoch
        cur_alpha = alpha_schedule_fn(epoch, t0)

        model.train()
        if aux_net is not None:
            if vec_opt is not None:
                aux_net.train()
            else:
                aux_net.eval()

        tl, c, t = 0, 0, 0
        epoch_aux_norms, epoch_dfa_norms = [], []

        for x, y in train_loader:
            x = x.view(x.size(0), -1).to(device); y = y.to(device); batch = x.size(0)
            with torch.no_grad():
                lo, hi = model(x, return_hidden=True); lv = F.cross_entropy(lo, y)
                eT = lo.softmax(-1); eT[torch.arange(batch), y] -= 1; s = eT.detach()
            hL = hi[-1].detach()

            # ----------------------------------------------------------------
            # Train Vec with standard perturbation targets (if applicable)
            # ----------------------------------------------------------------
            if vec_opt is not None and aux_net is not None:
                t_L = torch.ones(batch, device=device)
                a_term = aux_net(hL, t_L, s)
                hL_req = hL.clone().requires_grad_(True)
                ce = F.cross_entropy(
                    model.out_head(model.out_ln(hL_req)), y, reduction='sum')
                dL = torch.autograd.grad(ce, hL_req)[0].detach()
                loss_term = ((a_term - dL) ** 2).sum(-1).mean()
                lt = np.random.randint(0, L)
                h_l = hi[lt].detach()
                t_l = torch.full((batch,), lt / L, device=device)
                a_l = aux_net(h_l, t_l, s)
                lp2 = torch.tensor(0.0, device=device)
                for _ in range(M):
                    v = torch.randn_like(h_l)
                    v = v / (v.norm(-1, keepdim=True) + 1e-8)
                    with torch.no_grad():
                        lp = F.cross_entropy(
                            model.forward_from_layer(h_l + eps_pert * v, lt),
                            y, reduction='none')
                        lm = F.cross_entropy(
                            model.forward_from_layer(h_l - eps_pert * v, lt),
                            y, reduction='none')
                        gj = (lp - lm) / (2 * eps_pert)
                    lp2 = lp2 + (((a_l * v).sum(-1) - gj.detach()) ** 2).mean()
                lp2 /= M
                vl = loss_term + lp2
                vec_opt.zero_grad(); vl.backward()
                torch.nn.utils.clip_grad_norm_(aux_net.parameters(), 1.0)
                vec_opt.step()

            # ----------------------------------------------------------------
            # Compute credits for each block
            # ----------------------------------------------------------------
            credits = []
            for l in range(L):
                a_dfa = (eT @ Bs[l].T).detach()
                rms_d = (a_dfa ** 2).mean(-1, keepdim=True).sqrt() + 1e-6

                if branch_type == 'dfa' or aux_net is None or cur_alpha == 0.0:
                    credits.append(a_dfa / rms_d)
                    epoch_aux_norms.append(0.0)
                    epoch_dfa_norms.append((a_dfa / rms_d).norm().item())
                else:
                    h_l = hi[l].detach()
                    t_l = torch.full((batch,), l / L, device=device)
                    with torch.no_grad():
                        a_aux = aux_net(h_l, t_l, s).detach()
                    rms_v = (a_aux ** 2).mean(-1, keepdim=True).sqrt() + 1e-6
                    a_blend = cur_alpha * a_aux / rms_v + (1 - cur_alpha) * a_dfa / rms_d
                    credits.append(a_blend)
                    epoch_aux_norms.append((cur_alpha * a_aux / rms_v).norm().item())
                    epoch_dfa_norms.append(((1 - cur_alpha) * a_dfa / rms_d).norm().item())

            # ----------------------------------------------------------------
            # Update output head
            # ----------------------------------------------------------------
            lo2 = F.cross_entropy(model.out_head(model.out_ln(hL)), y)
            head_opt.zero_grad(); lo2.backward(); head_opt.step()

            # ----------------------------------------------------------------
            # Update blocks with local surrogate
            # ----------------------------------------------------------------
            for l in range(L):
                a = credits[l]
                rm = (a ** 2).mean(-1, keepdim=True).sqrt() + 1e-6
                f = model.blocks[l](hi[l].detach())
                ll = (f * (a / rm)).sum(-1).mean()
                block_opts[l].zero_grad(); ll.backward()
                torch.nn.utils.clip_grad_norm_(model.blocks[l].parameters(), 1.0)
                block_opts[l].step()

            # Update embedding with block-0 credit
            a0 = credits[0]
            r0 = (a0 ** 2).mean(-1, keepdim=True).sqrt() + 1e-6
            el = (model.embed(x) * (a0 / r0)).sum(-1).mean()
            embed_opt.zero_grad(); el.backward(); embed_opt.step()

            tl += lv.item() * batch; c += (lo.argmax(1) == y).sum().item(); t += batch

        for sch in scheds: sch.step()
        ta = evaluate(model, test_loader, device)
        log['test_acc'].append(ta); log['train_loss'].append(tl / t)

        mean_aux = np.mean(epoch_aux_norms) if epoch_aux_norms else 0.0
        mean_dfa = np.mean(epoch_dfa_norms) if epoch_dfa_norms else 1.0
        aeff = mean_aux / (mean_aux + mean_dfa + 1e-12)
        log['alpha_eff'].append((epoch, aeff))

        if epoch in diag_epochs:
            cm = 'blend' if (branch_type != 'dfa' and aux_net is not None
                             and cur_alpha > 0.0) else 'dfa'
            diag_aux = aux_net if cm == 'blend' else None
            gamma, rho = compute_diagnostics(
                model, diag_aux, Bs, test_loader, device, cm, cur_alpha)
            log['gamma'].append((epoch, gamma)); log['rho'].append((epoch, rho))
            if epoch <= t0 + 15 or epoch % 20 == 0 or epoch == total_epochs:
                frozen_str = ' [FROZEN]' if vec_frozen else ''
                print(f"    [{branch_name}]{frozen_str} Ep {epoch}: acc={ta:.4f}, "
                      f"G={gamma:.4f}, r={rho:.4f}, aeff={aeff:.3f}, alpha={cur_alpha:.3f}")
        elif epoch % 10 == 0 or epoch == total_epochs:
            frozen_str = ' [FROZEN]' if vec_frozen else ''
            print(f"    [{branch_name}]{frozen_str} Ep {epoch}: acc={ta:.4f}, "
                  f"alpha={cur_alpha:.3f}")

    return log


# ---------------------------------------------------------------------------
# Main experiment
# ---------------------------------------------------------------------------

def run_experiment(args):
    device = torch.device(f'cuda:{args.gpu}' if torch.cuda.is_available() else 'cpu')
    print(f"Using device: {device}")
    os.makedirs(args.output_dir, exist_ok=True)
    torch.manual_seed(args.seed); np.random.seed(args.seed)
    torch.cuda.manual_seed_all(args.seed)

    train_loader, test_loader = get_cifar10(args.batch_size)
    input_dim = 32 * 32 * 3; L = args.num_blocks; d = args.d_hidden

    # ----------------------------------------------------------------
    # Step 1: Train DFA and capture checkpoint at t0
    # ----------------------------------------------------------------
    print(f"\n{'='*60}\nTraining DFA baseline (checkpoint at t0={args.t0})\n{'='*60}")
    model_dfa = ResidualMLP(input_dim, d, 10, L).to(device)
    Bs, ckpt = train_dfa_get_checkpoint(
        model_dfa, train_loader, test_loader, device,
        args.epochs, args.t0, args.lr, args.wd)
    print(f"  Checkpoint acc at t0={args.t0}: {ckpt['acc']:.4f}")

    # ----------------------------------------------------------------
    # Step 2: Define branches
    # ----------------------------------------------------------------
    VEC_SEED = args.seed + 7777
    DECAY_WINDOW = 5

    def make_vec():
        torch.manual_seed(VEC_SEED)
        return VectorCreditNet(d_hidden=d, s_dim=10).to(device)

    # constant alpha
    def fixed_alpha(a):
        return lambda epoch, t0: a

    # (name, branch_type, aux_factory, freeze_epoch, alpha_schedule_fn)
    branches = [
        ('continue_DFA',
         'dfa', lambda: None, None,
         fixed_alpha(0.0)),

        ('blend_random_trainable_alpha075',
         'blend', make_vec, None,
         fixed_alpha(0.75)),

        ('freeze_after_1_fixed075',
         'blend_frozen', make_vec, 1,
         fixed_alpha(0.75)),

        ('freeze_after_5_fixed075',
         'blend_frozen', make_vec, 5,
         fixed_alpha(0.75)),

        ('freeze_after_1_decay_to_025',
         'blend_frozen', make_vec, 1,
         make_alpha_schedule(freeze_epoch=1, initial_alpha=0.75,
                             target_alpha=0.25, decay_window=DECAY_WINDOW)),

        ('freeze_after_5_decay_to_025',
         'blend_frozen', make_vec, 5,
         make_alpha_schedule(freeze_epoch=5, initial_alpha=0.75,
                             target_alpha=0.25, decay_window=DECAY_WINDOW)),

        ('freeze_after_1_decay_to_000',
         'blend_frozen', make_vec, 1,
         make_alpha_schedule(freeze_epoch=1, initial_alpha=0.75,
                             target_alpha=0.0, decay_window=DECAY_WINDOW)),

        ('freeze_after_5_decay_to_000',
         'blend_frozen', make_vec, 5,
         make_alpha_schedule(freeze_epoch=5, initial_alpha=0.75,
                             target_alpha=0.0, decay_window=DECAY_WINDOW)),
    ]

    # ----------------------------------------------------------------
    # Step 3: Run all branches
    # ----------------------------------------------------------------
    all_results = {}
    for bname, btype, aux_factory, freeze_ep, alpha_fn in branches:
        print(f"\n{'='*60}\n{bname}\n{'='*60}")
        model_b = ResidualMLP(input_dim, d, 10, L).to(device)
        model_b.load_state_dict(ckpt['model'])
        aux_net_b = aux_factory()

        log = run_branch(
            model_b, aux_net_b, ckpt['Bs'],
            train_loader, test_loader, device,
            args.t0, args.epochs, btype,
            alpha_fn, args.lr, args.lr_fb, args.wd, args.M,
            branch_name=bname,
            freeze_epoch=freeze_ep)
        all_results[bname] = log
        print(f"  {bname} final acc: {log['test_acc'][-1]:.4f}")

    # ----------------------------------------------------------------
    # Step 4: Summary table
    # ----------------------------------------------------------------
    dfa_final = all_results['continue_DFA']['test_acc'][-1]

    print(f"\n{'='*95}")
    print("SUMMARY — Phase 10A.8A: Freeze with Alpha Decay")
    print(f"{'='*95}")
    print(f"{'Branch':<38} {'@20':>6} {'final':>7} {'diff':>7} "
          f"{'mG_5:15':>9} {'mr_5:15':>9} {'aeff':>7}")
    print("-" * 85)

    for bname, log in all_results.items():
        accs = log['test_acc']
        idx20 = max(0, 20 - args.t0 - 1)
        acc20 = accs[idx20] if len(accs) > idx20 else accs[-1]
        final = accs[-1]
        diff = final - dfa_final
        gammas_e = [g for e, g in log['gamma'] if args.t0 < e <= args.t0 + 15]
        rhos_e = [r for e, r in log['rho'] if args.t0 < e <= args.t0 + 15]
        aeffs_e = [a for e, a in log['alpha_eff'] if args.t0 < e <= args.t0 + 15]
        mg = float(np.mean(gammas_e)) if gammas_e else float('nan')
        mr = float(np.mean(rhos_e)) if rhos_e else float('nan')
        mae = float(np.mean(aeffs_e)) if aeffs_e else float('nan')
        print(f"{bname:<38} {acc20:>6.4f} {final:>7.4f} {diff:>+7.4f} "
              f"{mg:>9.4f} {mr:>9.4f} {mae:>7.3f}")

    # ----------------------------------------------------------------
    # Step 5: Save results
    # ----------------------------------------------------------------
    save_data = {'args': vars(args), 'dfa_ckpt_acc': float(ckpt['acc'])}
    for bname, log in all_results.items():
        save_data[bname] = {
            'test_acc': log['test_acc'],
            'train_loss': log['train_loss'],
            'gamma': log['gamma'],
            'rho': log['rho'],
            'alpha_eff': log['alpha_eff'],
        }
    out_path = os.path.join(args.output_dir,
                            f'freeze_with_decay_t{args.t0}_s{args.seed}.json')
    with open(out_path, 'w') as f:
        json.dump(save_data, f, indent=2, default=float)
    print(f"\nSaved to {out_path}")

    # ----------------------------------------------------------------
    # Step 6: Judgment
    # ----------------------------------------------------------------
    print(f"\n{'='*60}\nJUDGMENT\n{'='*60}")
    r = {bname: log['test_acc'][-1] for bname, log in all_results.items()}
    dfa  = r['continue_DFA']
    ref  = r.get('blend_random_trainable_alpha075', float('nan'))
    f1   = r.get('freeze_after_1_fixed075', float('nan'))
    f5   = r.get('freeze_after_5_fixed075', float('nan'))
    f1d25 = r.get('freeze_after_1_decay_to_025', float('nan'))
    f5d25 = r.get('freeze_after_5_decay_to_025', float('nan'))
    f1d00 = r.get('freeze_after_1_decay_to_000', float('nan'))
    f5d00 = r.get('freeze_after_5_decay_to_000', float('nan'))

    print(f"  DFA={dfa:.4f}  ref={ref:.4f}")
    print(f"  freeze1_fixed={f1:.4f}  freeze5_fixed={f5:.4f}")
    print(f"  freeze1_to025={f1d25:.4f}  freeze5_to025={f5d25:.4f}")
    print(f"  freeze1_to000={f1d00:.4f}  freeze5_to000={f5d00:.4f}")

    thr = 0.003

    # Fixed vs trainable reference
    best_fixed = max(f1, f5)
    if best_fixed > ref - thr:
        print(f"\n  -> Best frozen-fixed ({best_fixed:.4f}) ≈ trainable reference: "
              "freezing early is sufficient; ongoing Vec training adds no value")
    elif ref > best_fixed + thr:
        print(f"\n  -> Trainable reference ({ref:.4f}) > best frozen-fixed ({best_fixed:.4f}): "
              "continuous Vec adaptation helps")

    # Effect of more training before freeze
    if f5 > f1 + thr:
        print(f"  -> More Vec training before freeze helps: "
              f"5ep ({f5:.4f}) > 1ep ({f1:.4f})")
    else:
        print(f"  -> Freeze timing (1 vs 5 epochs) makes little difference: "
              f"f1={f1:.4f}  f5={f5:.4f}")

    # Effect of decay on fixed-freeze branches
    print(f"\n  Decay effect (vs fixed075):")
    for label, fixed_v, d25, d00 in [
            ('freeze_after_1', f1, f1d25, f1d00),
            ('freeze_after_5', f5, f5d25, f5d00)]:
        print(f"    {label}: fixed={fixed_v:.4f}  ->0.25={d25:.4f}  ->0.00={d00:.4f}")
        if d25 > fixed_v + thr:
            print(f"      -> decay to 0.25 helps vs fixed ({d25-fixed_v:+.4f})")
        if d00 > fixed_v + thr:
            print(f"      -> decay to 0.0 (full DFA) helps vs fixed ({d00-fixed_v:+.4f})")
        if d00 > d25 + thr:
            print(f"      -> faster decay (to 0) better than partial ({d00-d25:+.4f})")
        elif d25 > d00 + thr:
            print(f"      -> partial decay (to 0.25) better than full decay ({d25-d00:+.4f})")

    # Overall winner
    best_name = max(r, key=r.get)
    print(f"\n  Best branch: {best_name} = {r[best_name]:.4f}  "
          f"(+{r[best_name]-dfa:+.4f} vs DFA)")


def main():
    parser = argparse.ArgumentParser(
        description='Phase 10A.8A: Freeze with Alpha Decay')
    parser.add_argument('--num_blocks', type=int, default=4)
    parser.add_argument('--d_hidden', type=int, default=256)
    parser.add_argument('--batch_size', type=int, default=128)
    parser.add_argument('--epochs', type=int, default=100)
    parser.add_argument('--t0', type=int, default=5)
    parser.add_argument('--alpha', type=float, default=0.75)
    parser.add_argument('--lr', type=float, default=1e-3)
    parser.add_argument('--lr_fb', type=float, default=1e-3)
    parser.add_argument('--wd', type=float, default=0.01)
    parser.add_argument('--M', type=int, default=4)
    parser.add_argument('--seed', type=int, default=42)
    parser.add_argument('--gpu', type=int, default=2)
    parser.add_argument('--output_dir', type=str, default='results/freeze_with_decay')
    args = parser.parse_args()
    run_experiment(args)


if __name__ == '__main__':
    main()