"""
Phase 2: CIFAR-10 Depth Scan.
Find the "Goldilocks regime" where Credit Bridge outperforms DFA.

Sweep: L in {2, 4, 6, 8, 12}, d in {256, 512}
Methods: DFA (3 seeds), Credit Bridge (3 seeds), BP (1 seed as reference)

Reuses training logic from cifar_resmlp.py.
"""
import os
import sys
import json
import argparse
import time
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 ValueNet, create_ema_model, update_ema
from models.state_bridge import StateBridgeNet
from metrics.credit_metrics import (
    cosine_similarity_batch, perturbation_correlation, nudging_test,
    offline_bp_cosine, feature_drift
)


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)
    train_loader = DataLoader(trainset, batch_size=batch_size, shuffle=True, num_workers=4, pin_memory=True)
    test_loader = DataLoader(testset, batch_size=batch_size, shuffle=False, num_workers=4, pin_memory=True)
    return train_loader, test_loader, 32 * 32 * 3, 10


def evaluate(model, test_loader, device):
    model.eval()
    correct, total = 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)
            logits = model(x)
            correct += (logits.argmax(1) == y).sum().item()
            total += x.size(0)
    return correct / total


# =============================================================================
# Training methods (adapted from cifar_resmlp.py)
# =============================================================================
def train_bp(model, train_loader, test_loader, device, args):
    optimizer = optim.AdamW(model.parameters(), lr=args.lr, weight_decay=args.wd)
    scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=args.epochs)
    log = {'train_loss': [], 'train_acc': [], 'test_acc': []}

    for epoch in range(1, args.epochs + 1):
        model.train()
        total_loss, correct, total = 0, 0, 0
        for x, y in train_loader:
            x = x.view(x.size(0), -1).to(device)
            y = y.to(device)
            logits = model(x)
            loss = F.cross_entropy(logits, y)
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
            total_loss += loss.item() * x.size(0)
            correct += (logits.argmax(1) == y).sum().item()
            total += x.size(0)
        scheduler.step()
        log['train_loss'].append(total_loss / total)
        log['train_acc'].append(correct / total)
        log['test_acc'].append(evaluate(model, test_loader, device))
        if epoch % 10 == 0 or epoch == 1:
            print(f"    [BP] Ep {epoch}: loss={log['train_loss'][-1]:.4f} "
                  f"train={log['train_acc'][-1]:.4f} test={log['test_acc'][-1]:.4f}")
    return log


def train_dfa(model, train_loader, test_loader, device, args):
    d = model.d_hidden
    L = model.num_blocks
    C = 10

    Bs = [torch.randn(d, C, device=device) / np.sqrt(C) for _ in range(L)]
    block_opts = [optim.AdamW(block.parameters(), lr=args.lr, weight_decay=args.wd)
                  for block in model.blocks]
    embed_opt = optim.AdamW(model.embed.parameters(), lr=args.lr, weight_decay=args.wd)
    head_opt = optim.AdamW(
        list(model.out_head.parameters()) + list(model.out_ln.parameters()),
        lr=args.lr, weight_decay=args.wd
    )
    all_schedulers = ([optim.lr_scheduler.CosineAnnealingLR(o, T_max=args.epochs) for o in block_opts]
                      + [optim.lr_scheduler.CosineAnnealingLR(embed_opt, T_max=args.epochs),
                         optim.lr_scheduler.CosineAnnealingLR(head_opt, T_max=args.epochs)])

    log = {'train_loss': [], 'train_acc': [], 'test_acc': []}

    for epoch in range(1, args.epochs + 1):
        model.train()
        total_loss, correct, total = 0, 0, 0
        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():
                logits, hiddens = model(x, return_hidden=True)
                loss_val = F.cross_entropy(logits, y)
                e_T = logits.softmax(dim=-1)
                e_T[torch.arange(batch), y] -= 1

            hL_det = hiddens[-1].detach()
            logits_out = model.out_head(model.out_ln(hL_det))
            loss_out = F.cross_entropy(logits_out, y)
            head_opt.zero_grad()
            loss_out.backward()
            head_opt.step()

            for l in range(L):
                h_l = hiddens[l].detach()
                a_dfa = (e_T @ Bs[l].T).detach()
                rms = (a_dfa ** 2).mean(dim=-1, keepdim=True).sqrt() + 1e-6
                a_norm = a_dfa / rms
                f_l = model.blocks[l](h_l)
                local_loss = (f_l * a_norm).sum(dim=-1).mean()
                block_opts[l].zero_grad()
                local_loss.backward()
                torch.nn.utils.clip_grad_norm_(model.blocks[l].parameters(), 1.0)
                block_opts[l].step()

            a_0_dfa = (e_T @ Bs[0].T).detach()
            rms_0 = (a_0_dfa ** 2).mean(dim=-1, keepdim=True).sqrt() + 1e-6
            a_0_norm = a_0_dfa / rms_0
            h0 = model.embed(x)
            embed_loss = (h0 * a_0_norm).sum(dim=-1).mean()
            embed_opt.zero_grad()
            embed_loss.backward()
            embed_opt.step()

            total_loss += loss_val.item() * batch
            correct += (logits.argmax(1) == y).sum().item()
            total += batch

        for s in all_schedulers:
            s.step()
        log['train_loss'].append(total_loss / total)
        log['train_acc'].append(correct / total)
        log['test_acc'].append(evaluate(model, test_loader, device))
        if epoch % 10 == 0 or epoch == 1:
            print(f"    [DFA] Ep {epoch}: loss={log['train_loss'][-1]:.4f} "
                  f"train={log['train_acc'][-1]:.4f} test={log['test_acc'][-1]:.4f}")
    return log, Bs


def train_credit_bridge(model, train_loader, test_loader, device, args):
    d = model.d_hidden
    L = model.num_blocks
    C = 10
    warmup_epochs = max(1, args.epochs // 5)

    value_net = ValueNet(d_hidden=d, s_dim=C, time_embed_dim=32,
                         hidden_dim=256, num_layers=3).to(device)
    value_net_ema = create_ema_model(value_net)

    Bs_fallback = [torch.randn(d, C, device=device) / np.sqrt(C) for _ in range(L)]

    block_opts = [optim.AdamW(block.parameters(), lr=args.lr, weight_decay=args.wd)
                  for block in model.blocks]
    embed_opt = optim.AdamW(model.embed.parameters(), lr=args.lr, weight_decay=args.wd)
    head_opt = optim.AdamW(
        list(model.out_head.parameters()) + list(model.out_ln.parameters()),
        lr=args.lr, weight_decay=args.wd
    )
    value_opt = optim.Adam(value_net.parameters(), lr=args.lr_fb)

    all_schedulers = ([optim.lr_scheduler.CosineAnnealingLR(o, T_max=args.epochs) for o in block_opts]
                      + [optim.lr_scheduler.CosineAnnealingLR(embed_opt, T_max=args.epochs),
                         optim.lr_scheduler.CosineAnnealingLR(head_opt, T_max=args.epochs)])

    lam = args.lam
    K_samples = args.K
    sigma_bridge = args.sigma_bridge
    ema_momentum = args.ema_momentum
    term_grad_weight = args.term_grad_weight

    log = {'train_loss': [], 'train_acc': [], 'test_acc': [], 'value_loss': []}

    for epoch in range(1, args.epochs + 1):
        model.train()
        value_net.train()
        total_loss, correct, total = 0, 0, 0
        total_vloss = 0

        if epoch <= warmup_epochs:
            credit_blend = 0.0
        else:
            credit_blend = min(1.0, (epoch - warmup_epochs) / max(1, warmup_epochs))

        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():
                logits, hiddens = model(x, return_hidden=True)
                loss_val = F.cross_entropy(logits, y)
                e_T = logits.softmax(dim=-1)
                e_T[torch.arange(batch), y] -= 1
                s = e_T.detach()
                true_loss = F.cross_entropy(logits, y, reduction='none').detach()

            hL_det = hiddens[-1].detach()

            # Train value net
            t_L = torch.ones(batch, device=device)
            V_terminal = value_net(hL_det, t_L, s)
            loss_term = ((V_terminal - true_loss) ** 2).mean()

            loss_tgrad = torch.tensor(0.0, device=device)
            if term_grad_weight > 0:
                hL_req = hL_det.clone().requires_grad_(True)
                V_at_L = value_net(hL_req, t_L, s)
                grad_V_L = torch.autograd.grad(V_at_L.sum(), hL_req, create_graph=True)[0]
                hL_req2 = hL_det.clone().requires_grad_(True)
                logits_tgt = model.out_head(model.out_ln(hL_req2))
                ce_loss = F.cross_entropy(logits_tgt, y, reduction='sum')
                a_L_exact = torch.autograd.grad(ce_loss, hL_req2, create_graph=False)[0].detach()
                loss_tgrad = ((grad_V_L - a_L_exact) ** 2).sum(dim=-1).mean()

            loss_bridge = 0.0
            for l in range(L):
                h_l_det = hiddens[l].detach()
                t_l = torch.full((batch,), l / L, device=device)
                t_l_next = torch.full((batch,), (l + 1) / L, device=device)
                V_l = value_net(h_l_det, t_l, s)
                with torch.no_grad():
                    h_next_det = hiddens[l + 1].detach()
                    log_terms = []
                    for k in range(K_samples):
                        noise = sigma_bridge * torch.randn_like(h_next_det)
                        V_next = value_net_ema(h_next_det + noise, t_l_next, s)
                        log_terms.append(-V_next / lam)
                    log_stack = torch.stack(log_terms, dim=-1)
                    V_target = -lam * (torch.logsumexp(log_stack, dim=-1) - np.log(K_samples))
                loss_bridge = loss_bridge + ((V_l - V_target.detach()) ** 2).mean()
            loss_bridge = loss_bridge / L

            value_loss = loss_term + loss_bridge + term_grad_weight * loss_tgrad
            value_opt.zero_grad()
            value_loss.backward()
            torch.nn.utils.clip_grad_norm_(value_net.parameters(), 1.0)
            value_opt.step()
            update_ema(value_net, value_net_ema, ema_momentum)
            total_vloss += value_loss.item() * batch

            # Compute credits
            cb_credits = []
            for l in range(L):
                h_l_det = hiddens[l].detach().requires_grad_(True)
                t_l = torch.full((batch,), l / L, device=device)
                V_l = value_net(h_l_det, t_l, s)
                a_l = torch.autograd.grad(V_l.sum(), h_l_det, create_graph=False)[0]
                cb_credits.append(a_l.detach())

            dfa_credits = [(e_T @ Bs_fallback[l].T).detach() for l in range(L)]

            credits = []
            for l in range(L):
                if credit_blend >= 1.0:
                    a = cb_credits[l]
                elif credit_blend <= 0.0:
                    a = dfa_credits[l]
                else:
                    cb_rms = (cb_credits[l] ** 2).mean(dim=-1, keepdim=True).sqrt() + 1e-6
                    dfa_rms = (dfa_credits[l] ** 2).mean(dim=-1, keepdim=True).sqrt() + 1e-6
                    a = credit_blend * (cb_credits[l] / cb_rms) + (1 - credit_blend) * (dfa_credits[l] / dfa_rms)
                credits.append(a)

            # Update output head
            logits_out = model.out_head(model.out_ln(hL_det))
            loss_out = F.cross_entropy(logits_out, y)
            head_opt.zero_grad()
            loss_out.backward()
            head_opt.step()

            # Update blocks
            for l in range(L):
                h_l = hiddens[l].detach()
                a = credits[l]
                rms = (a ** 2).mean(dim=-1, keepdim=True).sqrt() + 1e-6
                a_norm = a / rms
                f_l = model.blocks[l](h_l)
                local_loss = (f_l * a_norm).sum(dim=-1).mean()
                block_opts[l].zero_grad()
                local_loss.backward()
                torch.nn.utils.clip_grad_norm_(model.blocks[l].parameters(), 1.0)
                block_opts[l].step()

            # Update embedding
            a_0 = credits[0]
            rms_0 = (a_0 ** 2).mean(dim=-1, keepdim=True).sqrt() + 1e-6
            a_0_norm = a_0 / rms_0
            h0 = model.embed(x)
            embed_loss = (h0 * a_0_norm).sum(dim=-1).mean()
            embed_opt.zero_grad()
            embed_loss.backward()
            embed_opt.step()

            total_loss += loss_val.item() * batch
            correct += (logits.argmax(1) == y).sum().item()
            total += batch

        for sch in all_schedulers:
            sch.step()

        log['train_loss'].append(total_loss / total)
        log['train_acc'].append(correct / total)
        log['test_acc'].append(evaluate(model, test_loader, device))
        log['value_loss'].append(total_vloss / total)
        if epoch % 10 == 0 or epoch == 1:
            phase = "warmup" if epoch <= warmup_epochs else f"blend={credit_blend:.2f}"
            print(f"    [CB] Ep {epoch} ({phase}): loss={log['train_loss'][-1]:.4f} "
                  f"train={log['train_acc'][-1]:.4f} test={log['test_acc'][-1]:.4f} "
                  f"vloss={log['value_loss'][-1]:.6f}")
    return log, value_net, value_net_ema


# =============================================================================
# Diagnostics
# =============================================================================
def compute_diagnostics(model, method_name, test_loader, device, args,
                        value_net=None, dfa_Bs=None):
    model.eval()
    if value_net is not None:
        value_net.eval()

    d = model.d_hidden
    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 gradients via manual forward
    logits_bp, hiddens_bp = model(x, return_hidden=True)
    for l in range(L + 1):
        hiddens_bp[l].retain_grad()
    loss_bp = F.cross_entropy(logits_bp, y)
    loss_bp.backward()
    bp_grads = {l: hiddens_bp[l].grad.detach().clone() for l in range(L + 1)}

    with torch.no_grad():
        logits, hiddens = model(x, return_hidden=True)
        e_T = logits.softmax(dim=-1)
        e_T[torch.arange(batch), y] -= 1
        s = e_T.detach()

    results = {
        'bp_cosine': [],
        'perturbation_rho': [],
        'nudging': {'0.001': [], '0.003': [], '0.01': []},
    }

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

        if method_name == 'bp':
            a_l = bp_grads[l]
        elif method_name == 'dfa':
            a_l = (e_T @ dfa_Bs[l].T).detach()
        elif method_name == 'credit_bridge':
            h_l_req = h_l.clone().requires_grad_(True)
            V_l = value_net(h_l_req, t_l, s)
            a_l = torch.autograd.grad(V_l.sum(), h_l_req, create_graph=False)[0].detach()
        else:
            raise ValueError(f"Unknown method: {method_name}")

        bp_cos = cosine_similarity_batch(a_l, bp_grads[l])
        results['bp_cosine'].append(bp_cos)

        def make_fwd_fn(start_l):
            def fwd_fn(h):
                with torch.no_grad():
                    curr = h
                    for i in range(start_l, L):
                        curr = curr + model.blocks[i](curr)
                    out = model.out_head(model.out_ln(curr))
                    return F.cross_entropy(out, y, reduction='none')
            return fwd_fn

        fwd_fn = make_fwd_fn(l)
        rho = perturbation_correlation(h_l, a_l, fwd_fn, epsilon=1e-3, M=16)
        results['perturbation_rho'].append(rho)

        for eta in [0.001, 0.003, 0.01]:
            nud = nudging_test(h_l, a_l, fwd_fn, eta=eta)
            results['nudging'][str(eta)].append(nud)

    return results


# =============================================================================
# Main
# =============================================================================
def serialize(obj):
    if isinstance(obj, dict):
        return {str(k): serialize(v) for k, v in obj.items()}
    elif isinstance(obj, list):
        return [serialize(v) for v in obj]
    elif isinstance(obj, (np.floating, np.integer)):
        return float(obj)
    elif isinstance(obj, np.ndarray):
        return obj.tolist()
    elif isinstance(obj, torch.Tensor):
        return obj.cpu().numpy().tolist()
    return obj


def run_config(d_hidden, num_blocks, seed, methods, args, device):
    """Run specified methods for a single (d, L, seed) config."""
    input_dim = 32 * 32 * 3
    num_classes = 10
    args.num_classes = num_classes

    train_loader, test_loader, _, _ = get_cifar10(args.batch_size)

    results = {}

    for method in methods:
        torch.manual_seed(seed)
        np.random.seed(seed)
        torch.cuda.manual_seed_all(seed)

        model = ResidualMLP(input_dim, d_hidden, num_classes, num_blocks).to(device)
        init_params = {n: p.clone().detach() for n, p in model.named_parameters()}

        print(f"\n  d={d_hidden}, L={num_blocks}, seed={seed}, method={method}")
        t0 = time.time()

        if method == 'bp':
            log = train_bp(model, train_loader, test_loader, device, args)
            diag = compute_diagnostics(model, 'bp', test_loader, device, args)
            results['bp'] = {'log': log, 'diagnostics': diag}

        elif method == 'dfa':
            log, Bs = train_dfa(model, train_loader, test_loader, device, args)
            diag = compute_diagnostics(model, 'dfa', test_loader, device, args, dfa_Bs=Bs)
            results['dfa'] = {'log': log, 'diagnostics': diag}

        elif method == 'credit_bridge':
            log, vnet, _ = train_credit_bridge(model, train_loader, test_loader, device, args)
            diag = compute_diagnostics(model, 'credit_bridge', test_loader, device, args,
                                       value_net=vnet)
            results['credit_bridge'] = {'log': log, 'diagnostics': diag}

        drift = feature_drift(init_params, {n: p.detach() for n, p in model.named_parameters()})
        results[method]['drift'] = drift

        elapsed = time.time() - t0
        test_acc = log['test_acc'][-1]
        mean_gamma = np.mean(diag['bp_cosine'])
        mean_rho = np.mean(diag['perturbation_rho'])
        print(f"    Done in {elapsed:.0f}s: acc={test_acc:.4f} Gamma={mean_gamma:.4f} rho={mean_rho:.4f}")

    return results


def main():
    parser = argparse.ArgumentParser(description='CIFAR-10 Depth Scan')
    parser.add_argument('--depths', type=int, nargs='+', default=[2, 4, 6, 8, 12])
    parser.add_argument('--widths', type=int, nargs='+', default=[256, 512])
    parser.add_argument('--seeds', type=int, nargs='+', default=[42, 123, 456])
    parser.add_argument('--bp_seeds', type=int, nargs='+', default=[42],
                        help='Seeds for BP (reference only)')
    parser.add_argument('--methods', type=str, nargs='+',
                        default=['bp', 'dfa', 'credit_bridge'])
    parser.add_argument('--batch_size', type=int, default=128)
    parser.add_argument('--epochs', type=int, default=100)
    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('--lam', type=float, default=0.1)
    parser.add_argument('--K', type=int, default=4)
    parser.add_argument('--sigma_bridge', type=float, default=0.05)
    parser.add_argument('--ema_momentum', type=float, default=0.995)
    parser.add_argument('--term_grad_weight', type=float, default=1.0)
    parser.add_argument('--gpu', type=int, default=1)
    parser.add_argument('--output_dir', type=str, default='results/cifar_depth_scan')
    args = parser.parse_args()

    device = torch.device(f'cuda:{args.gpu}' if torch.cuda.is_available() else 'cpu')
    print(f"Device: {device}")
    print(f"Depths: {args.depths}, Widths: {args.widths}")
    print(f"Seeds: {args.seeds}, BP seeds: {args.bp_seeds}")
    os.makedirs(args.output_dir, exist_ok=True)

    all_summary = {}

    for d_hidden in args.widths:
        for num_blocks in args.depths:
            for seed in args.seeds:
                key = f"d{d_hidden}_L{num_blocks}_s{seed}"

                # Determine which methods to run for this seed
                methods = []
                for m in args.methods:
                    if m == 'bp' and seed not in args.bp_seeds:
                        continue
                    methods.append(m)

                if not methods:
                    continue

                print(f"\n{'='*60}")
                print(f"Config: {key}, methods: {methods}")
                print(f"{'='*60}")

                result = run_config(d_hidden, num_blocks, seed, methods, args, device)

                # Save per-config result
                out_path = os.path.join(args.output_dir, f'{key}.json')
                with open(out_path, 'w') as f:
                    json.dump(serialize(result), f, indent=2)

                # Summary
                summary = {}
                for method in result:
                    diag = result[method]['diagnostics']
                    summary[method] = {
                        'test_acc': result[method]['log']['test_acc'][-1],
                        'mean_bp_cosine': float(np.mean(diag['bp_cosine'])),
                        'mean_rho': float(np.mean(diag['perturbation_rho'])),
                        'mean_nudge_01': float(np.mean(diag['nudging']['0.01'])),
                        'bp_cosine_per_layer': [float(x) for x in diag['bp_cosine']],
                        'rho_per_layer': [float(x) for x in diag['perturbation_rho']],
                    }
                all_summary[key] = summary

    # Save full summary
    summary_path = os.path.join(args.output_dir, 'summary.json')
    with open(summary_path, 'w') as f:
        json.dump(all_summary, f, indent=2)
    print(f"\nSummary saved to {summary_path}")

    # Print table
    print("\n" + "=" * 90)
    print("CIFAR-10 DEPTH SCAN SUMMARY")
    print("=" * 90)
    print(f"{'Config':<25} {'Method':<18} {'Acc':>8} {'Gamma':>8} {'rho':>8} {'nudge':>10}")
    print("-" * 90)
    for key in sorted(all_summary.keys()):
        for method in sorted(all_summary[key].keys()):
            s = all_summary[key][method]
            print(f"{key:<25} {method:<18} {s['test_acc']:>8.4f} {s['mean_bp_cosine']:>8.4f} "
                  f"{s['mean_rho']:>8.4f} {s['mean_nudge_01']:>10.6f}")
        print()


if __name__ == '__main__':
    main()