Update prototype.py

fbe2b18 verified 7 months ago

12.5 kB

"""
SVAE v1 — Clean SVD Autoencoder
==================================
The version that works. 0.071 MSE at 384:1 compression.
Image → encoder MLP → matrix (32×32) → real SVD → keep 8 → decode → image
Spectral concentration pushes energy into the kept 8.
No extra machinery. The SVD IS the bottleneck.
"""

import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
import torchvision.transforms as T
import math


# ── Data ──

def get_cifar10(batch_size=256):
    transform = T.Compose([
        T.ToTensor(),
        T.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616)),
    ])
    train_ds = torchvision.datasets.CIFAR10(root='./data', train=True, download=True, transform=transform)
    test_ds = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=transform)
    train_loader = torch.utils.data.DataLoader(train_ds, batch_size=batch_size, shuffle=True, num_workers=2)
    test_loader = torch.utils.data.DataLoader(test_ds, batch_size=batch_size, shuffle=False, num_workers=2)
    return train_loader, test_loader


# ── SVAE ──

class SVAE(nn.Module):
    def __init__(self, matrix_h=32, matrix_k=32, keep_k=8):
        super().__init__()
        self.matrix_h = matrix_h
        self.matrix_k = matrix_k
        self.keep_k = keep_k
        self.img_dim = 3 * 32 * 32
        self.mat_dim = matrix_h * matrix_k

        self.encoder = nn.Sequential(
            nn.Linear(self.img_dim, 512),
            nn.GELU(),
            nn.Linear(512, 512),
            nn.GELU(),
            nn.Linear(512, self.mat_dim),
        )
        self.decoder = nn.Sequential(
            nn.Linear(self.mat_dim, 512),
            nn.GELU(),
            nn.Linear(512, 512),
            nn.GELU(),
            nn.Linear(512, self.img_dim),
        )

    def encode(self, images):
        B = images.shape[0]
        M = self.encoder(images.reshape(B, -1)).reshape(B, self.matrix_h, self.matrix_k)
        U, S, Vt = torch.linalg.svd(M, full_matrices=False)
        k = self.keep_k
        return {
            'U': U[:, :, :k], 'S': S[:, :k], 'Vt': Vt[:, :k, :],
            'S_full': S, 'M': M,
        }

    def decode_from_svd(self, U, S, Vt):
        B = U.shape[0]
        M_hat = torch.bmm(U * S.unsqueeze(1), Vt)
        return self.decoder(M_hat.reshape(B, -1)).reshape(B, 3, 32, 32)

    def forward(self, images):
        svd = self.encode(images)
        recon = self.decode_from_svd(svd['U'], svd['S'], svd['Vt'])
        return {'recon': recon, 'svd': svd}

    @staticmethod
    def effective_rank(S):
        p = S / (S.sum(-1, keepdim=True) + 1e-8)
        p = p.clamp(min=1e-8)
        return (-(p * p.log()).sum(-1)).exp()

    @staticmethod
    def energy_ratio(S_kept, S_full):
        return (S_kept ** 2).sum(-1) / ((S_full ** 2).sum(-1) + 1e-8)

    @staticmethod
    def spectral_concentration_loss(S_full, keep_k):
        tail = (S_full[:, keep_k:] ** 2).sum(-1)
        head = (S_full[:, :keep_k] ** 2).sum(-1)
        return (tail / (head + 1e-8)).mean()


# ── Training ──

def train(epochs=50, lr=1e-3, keep_k=8, conc_weight=0.5, device='cuda'):
    device = torch.device(device if torch.cuda.is_available() else 'cpu')
    train_loader, test_loader = get_cifar10(batch_size=256)

    model = SVAE(matrix_h=64, matrix_k=64, keep_k=keep_k).to(device)
    opt = torch.optim.Adam(model.parameters(), lr=lr)
    sched = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=epochs)

    total_params = sum(p.numel() for p in model.parameters())
    max_rank = min(model.matrix_h, model.matrix_k)
    print(f"SVAE v1 — Clean SVD Autoencoder")
    print(f"  Matrix: ({model.matrix_h}, {model.matrix_k}) → max rank {max_rank}, keep {keep_k}")
    print(f"  Compression: {model.img_dim} → {keep_k} singular values ({model.img_dim // keep_k}:1)")
    print(f"  Params: {total_params:,}")
    print(f"  Device: {device}")
    print("=" * 85)
    print(f"{'ep':>3} | {'loss':>7} {'recon':>7} {'conc':>7} | "
          f"{'t_recon':>7} | "
          f"{'erank':>6} {'energy':>6} {'S0':>7} {'Sk-1':>7} {'tail':>7}")
    print("-" * 85)

    for epoch in range(1, epochs + 1):
        model.train()
        total_loss, total_recon, n = 0, 0, 0

        for images, labels in train_loader:
            images = images.to(device)
            opt.zero_grad()
            out = model(images)

            recon_loss = F.mse_loss(out['recon'], images)
            conc_loss = model.spectral_concentration_loss(out['svd']['S_full'], keep_k)

            loss = recon_loss + conc_weight * conc_loss
            loss.backward()
            opt.step()

            total_loss += loss.item() * len(images)
            total_recon += recon_loss.item() * len(images)
            n += len(images)

        sched.step()

        if epoch % 2 == 0 or epoch <= 3:
            model.eval()
            test_recon, test_n = 0, 0
            test_erank, test_energy = 0, 0
            test_S = None
            nb = 0

            with torch.no_grad():
                for images, labels in test_loader:
                    images = images.to(device)
                    out = model(images)
                    test_recon += F.mse_loss(out['recon'], images).item() * len(images)
                    test_n += len(images)
                    test_erank += model.effective_rank(out['svd']['S_full']).mean().item()
                    test_energy += model.energy_ratio(out['svd']['S'], out['svd']['S_full']).mean().item()
                    if test_S is None:
                        test_S = out['svd']['S_full'].mean(0).cpu()
                    else:
                        test_S += out['svd']['S_full'].mean(0).cpu()
                    nb += 1

            test_erank /= nb
            test_energy /= nb
            test_S /= nb

            print(f"{epoch:3d} | {total_loss/n:7.4f} {total_recon/n:7.4f} "
                  f"{conc_loss.item():7.4f} | "
                  f"{test_recon/test_n:7.4f} | "
                  f"{test_erank:6.2f} {test_energy:6.3f} "
                  f"{test_S[0]:7.3f} {test_S[keep_k-1]:7.3f} {test_S[keep_k]:7.4f}")

    # ── Final Analysis ──
    print()
    print("=" * 85)
    print("FINAL ANALYSIS")
    print("=" * 85)

    model.eval()
    all_S_full, all_S_kept, all_erank, all_energy = [], [], [], []
    all_recon_err, all_labels = [], []

    with torch.no_grad():
        for images, labels in test_loader:
            images = images.to(device)
            out = model(images)

            all_S_full.append(out['svd']['S_full'].cpu())
            all_S_kept.append(out['svd']['S'].cpu())
            all_erank.append(model.effective_rank(out['svd']['S_full']).cpu())
            all_energy.append(model.energy_ratio(out['svd']['S'], out['svd']['S_full']).cpu())
            all_recon_err.append(
                F.mse_loss(out['recon'], images, reduction='none')
                .mean(dim=(1, 2, 3)).cpu())
            all_labels.append(labels.cpu())

    all_S_full = torch.cat(all_S_full)
    all_S_kept = torch.cat(all_S_kept)
    all_erank = torch.cat(all_erank)
    all_energy = torch.cat(all_energy)
    all_recon_err = torch.cat(all_recon_err)
    all_labels = torch.cat(all_labels)

    print(f"\n  Bottleneck: {keep_k} / {min(model.matrix_h, model.matrix_k)} singular values")
    print(f"  Energy captured: {all_energy.mean():.3f} ± {all_energy.std():.3f}")
    print(f"  Effective rank (full): {all_erank.mean():.2f} ± {all_erank.std():.2f}")
    print(f"  Recon MSE: {all_recon_err.mean():.6f} ± {all_recon_err.std():.6f}")

    # Singular value profile
    S_mean = all_S_full.mean(0)
    total_energy = (S_mean ** 2).sum()
    print(f"\n  Singular value profile:")
    cumulative = 0
    for i in range(min(32, len(S_mean))):
        e = (S_mean[i] ** 2).item()
        cumulative += e
        pct = cumulative / total_energy * 100
        bar = "█" * int(S_mean[i].item() * 30 / (S_mean[0].item() + 1e-8))
        marker = " ← k" if i == keep_k - 1 else ("   ┄ tail" if i == keep_k else "")
        print(f"    S[{i:2d}]: {S_mean[i]:8.3f}  cum_energy={pct:5.1f}%  {bar}{marker}")

    # Per-class
    cifar_names = ['plane', 'car', 'bird', 'cat', 'deer',
                   'dog', 'frog', 'horse', 'ship', 'truck']
    print(f"\n  Per-class:")
    print(f"    {'class':>6}  {'erank':>6}  {'energy':>6}  {'recon':>8}  {'S0':>7}  {'S1':>7}")
    for c in range(10):
        mask = all_labels == c
        er = all_erank[mask].mean().item()
        en = all_energy[mask].mean().item()
        rc = all_recon_err[mask].mean().item()
        s0 = all_S_full[mask, 0].mean().item()
        s1 = all_S_full[mask, 1].mean().item()
        print(f"    {cifar_names[c]:>6}  {er:6.2f}  {en:6.3f}  {rc:8.6f}  {s0:7.3f}  {s1:7.3f}")

    # Recon by energy quartile
    print(f"\n  Recon quality by energy quartile:")
    q25, q50, q75 = all_energy.quantile(torch.tensor([0.25, 0.50, 0.75]))
    for label, lo, hi in [("Q1 (low energy)", 0, q25),
                           ("Q2", q25, q50),
                           ("Q3", q50, q75),
                           ("Q4 (high energy)", q75, 1.1)]:
        mask = (all_energy >= lo) & (all_energy < hi)
        if mask.sum() > 0:
            recon = all_recon_err[mask].mean().item()
            er = all_erank[mask].mean().item()
            print(f"    {label:>17}: n={mask.sum():5d} recon={recon:.6f} erank={er:.2f}")

    # ── Save reconstruction grid ──
    print(f"\n  Saving reconstruction grid...")
    import matplotlib
    matplotlib.use('Agg')
    import matplotlib.pyplot as plt

    # CIFAR-10 denormalization
    mean = torch.tensor([0.4914, 0.4822, 0.4465]).reshape(1, 3, 1, 1).to(device)
    std = torch.tensor([0.2470, 0.2435, 0.2616]).reshape(1, 3, 1, 1).to(device)

    model.eval()
    with torch.no_grad():
        # Grab one batch
        images, labels = next(iter(test_loader))
        images = images.to(device)
        out = model(images)

        # Pick 2 samples per class = 20 images
        selected_idx = []
        for c in range(10):
            class_idx = (labels == c).nonzero(as_tuple=True)[0]
            selected_idx.extend(class_idx[:2].tolist())

        orig = images[selected_idx]
        recon_full = out['recon'][selected_idx]

        # Progressive reconstructions: 1, 4, 8, 16 modes
        U = out['svd']['U'][selected_idx]
        S = out['svd']['S'][selected_idx]
        Vt = out['svd']['Vt'][selected_idx]

        mode_counts = [1, 4, 8, keep_k]
        # deduplicate if keep_k is already in the list
        mode_counts = list(dict.fromkeys(mode_counts))
        prog_recons = []
        for n_modes in mode_counts:
            n_modes = min(n_modes, S.shape[1])
            U_n = U[:, :, :n_modes]
            S_n = S[:, :n_modes]
            Vt_n = Vt[:, :n_modes, :]
            r = model.decode_from_svd(U_n, S_n, Vt_n)
            prog_recons.append(r)

    def denorm(t):
        return (t * std + mean).clamp(0, 1).cpu()

    n_samples = len(selected_idx)
    n_cols = 2 + len(mode_counts)  # original + progressives + error
    fig, axes = plt.subplots(n_samples, n_cols, figsize=(n_cols * 1.5, n_samples * 1.5))

    col_titles = ['Original'] + [f'{m} mode{"s" if m > 1 else ""}' for m in mode_counts] + ['|Error|×5']

    for i in range(n_samples):
        # Original
        img_orig = denorm(orig[i:i+1])[0].permute(1, 2, 0).numpy()
        axes[i, 0].imshow(img_orig)

        # Progressive
        for j, r in enumerate(prog_recons):
            img_r = denorm(r[i:i+1])[0].permute(1, 2, 0).numpy()
            axes[i, j+1].imshow(img_r)

        # Error map (amplified 5×)
        err_col = 1 + len(prog_recons)
        diff = (denorm(orig[i:i+1]) - denorm(prog_recons[-1][i:i+1])).abs() * 5
        diff = diff.clamp(0, 1)[0].permute(1, 2, 0).numpy()
        axes[i, err_col].imshow(diff)

        # Class label
        c = labels[selected_idx[i]].item()
        axes[i, 0].set_ylabel(cifar_names[c], fontsize=8, rotation=0, labelpad=35)

    for j, title in enumerate(col_titles):
        axes[0, j].set_title(title, fontsize=8)

    for ax in axes.flat:
        ax.axis('off')

    plt.tight_layout()
    plt.savefig('/content/svae_recon_grid.png', dpi=200, bbox_inches='tight')
    print(f"  Saved to /content/svae_recon_grid.png")
    try:
        plt.show()
    except:
        pass
    plt.close()


if __name__ == "__main__":
    train(keep_k=16)

	"""
	SVAE v1 — Clean SVD Autoencoder
	==================================
	The version that works. 0.071 MSE at 384:1 compression.

	Image → encoder MLP → matrix (32×32) → real SVD → keep 8 → decode → image
	Spectral concentration pushes energy into the kept 8.
	No extra machinery. The SVD IS the bottleneck.
	"""

	import torch
	import torch.nn as nn
	import torch.nn.functional as F
	import torchvision
	import torchvision.transforms as T
	import math


	# ── Data ──

	def get_cifar10(batch_size=256):
	transform = T.Compose([
	T.ToTensor(),
	T.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616)),
	])
	train_ds = torchvision.datasets.CIFAR10(root='./data', train=True, download=True, transform=transform)
	test_ds = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=transform)
	train_loader = torch.utils.data.DataLoader(train_ds, batch_size=batch_size, shuffle=True, num_workers=2)
	test_loader = torch.utils.data.DataLoader(test_ds, batch_size=batch_size, shuffle=False, num_workers=2)
	return train_loader, test_loader


	# ── SVAE ──

	class SVAE(nn.Module):
	def __init__(self, matrix_h=32, matrix_k=32, keep_k=8):
	super().__init__()
	self.matrix_h = matrix_h
	self.matrix_k = matrix_k
	self.keep_k = keep_k
	self.img_dim = 3 * 32 * 32
	self.mat_dim = matrix_h * matrix_k

	self.encoder = nn.Sequential(
	nn.Linear(self.img_dim, 512),
	nn.GELU(),
	nn.Linear(512, 512),
	nn.GELU(),
	nn.Linear(512, self.mat_dim),
	)
	self.decoder = nn.Sequential(
	nn.Linear(self.mat_dim, 512),
	nn.GELU(),
	nn.Linear(512, 512),
	nn.GELU(),
	nn.Linear(512, self.img_dim),
	)

	def encode(self, images):
	B = images.shape[0]
	M = self.encoder(images.reshape(B, -1)).reshape(B, self.matrix_h, self.matrix_k)
	U, S, Vt = torch.linalg.svd(M, full_matrices=False)
	k = self.keep_k
	return {
	'U': U[:, :, :k], 'S': S[:, :k], 'Vt': Vt[:, :k, :],
	'S_full': S, 'M': M,
	}

	def decode_from_svd(self, U, S, Vt):
	B = U.shape[0]
	M_hat = torch.bmm(U * S.unsqueeze(1), Vt)
	return self.decoder(M_hat.reshape(B, -1)).reshape(B, 3, 32, 32)

	def forward(self, images):
	svd = self.encode(images)
	recon = self.decode_from_svd(svd['U'], svd['S'], svd['Vt'])
	return {'recon': recon, 'svd': svd}

	@staticmethod
	def effective_rank(S):
	p = S / (S.sum(-1, keepdim=True) + 1e-8)
	p = p.clamp(min=1e-8)
	return (-(p * p.log()).sum(-1)).exp()

	@staticmethod
	def energy_ratio(S_kept, S_full):
	return (S_kept 2).sum(-1) / ((S_full 2).sum(-1) + 1e-8)

	@staticmethod
	def spectral_concentration_loss(S_full, keep_k):
	tail = (S_full[:, keep_k:] ** 2).sum(-1)
	head = (S_full[:, :keep_k] ** 2).sum(-1)
	return (tail / (head + 1e-8)).mean()


	# ── Training ──

	def train(epochs=50, lr=1e-3, keep_k=8, conc_weight=0.5, device='cuda'):
	device = torch.device(device if torch.cuda.is_available() else 'cpu')
	train_loader, test_loader = get_cifar10(batch_size=256)

	model = SVAE(matrix_h=64, matrix_k=64, keep_k=keep_k).to(device)
	opt = torch.optim.Adam(model.parameters(), lr=lr)
	sched = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=epochs)

	total_params = sum(p.numel() for p in model.parameters())
	max_rank = min(model.matrix_h, model.matrix_k)
	print(f"SVAE v1 — Clean SVD Autoencoder")
	print(f" Matrix: ({model.matrix_h}, {model.matrix_k}) → max rank {max_rank}, keep {keep_k}")
	print(f" Compression: {model.img_dim} → {keep_k} singular values ({model.img_dim // keep_k}:1)")
	print(f" Params: {total_params:,}")
	print(f" Device: {device}")
	print("=" * 85)
	print(f"{'ep':>3} \| {'loss':>7} {'recon':>7} {'conc':>7} \| "
	f"{'t_recon':>7} \| "
	f"{'erank':>6} {'energy':>6} {'S0':>7} {'Sk-1':>7} {'tail':>7}")
	print("-" * 85)

	for epoch in range(1, epochs + 1):
	model.train()
	total_loss, total_recon, n = 0, 0, 0

	for images, labels in train_loader:
	images = images.to(device)
	opt.zero_grad()
	out = model(images)

	recon_loss = F.mse_loss(out['recon'], images)
	conc_loss = model.spectral_concentration_loss(out['svd']['S_full'], keep_k)

	loss = recon_loss + conc_weight * conc_loss
	loss.backward()
	opt.step()

	total_loss += loss.item() * len(images)
	total_recon += recon_loss.item() * len(images)
	n += len(images)

	sched.step()

	if epoch % 2 == 0 or epoch <= 3:
	model.eval()
	test_recon, test_n = 0, 0
	test_erank, test_energy = 0, 0
	test_S = None
	nb = 0

	with torch.no_grad():
	for images, labels in test_loader:
	images = images.to(device)
	out = model(images)
	test_recon += F.mse_loss(out['recon'], images).item() * len(images)
	test_n += len(images)
	test_erank += model.effective_rank(out['svd']['S_full']).mean().item()
	test_energy += model.energy_ratio(out['svd']['S'], out['svd']['S_full']).mean().item()
	if test_S is None:
	test_S = out['svd']['S_full'].mean(0).cpu()
	else:
	test_S += out['svd']['S_full'].mean(0).cpu()
	nb += 1

	test_erank /= nb
	test_energy /= nb
	test_S /= nb

	print(f"{epoch:3d} \| {total_loss/n:7.4f} {total_recon/n:7.4f} "
	f"{conc_loss.item():7.4f} \| "
	f"{test_recon/test_n:7.4f} \| "
	f"{test_erank:6.2f} {test_energy:6.3f} "
	f"{test_S[0]:7.3f} {test_S[keep_k-1]:7.3f} {test_S[keep_k]:7.4f}")

	# ── Final Analysis ──
	print()
	print("=" * 85)
	print("FINAL ANALYSIS")
	print("=" * 85)

	model.eval()
	all_S_full, all_S_kept, all_erank, all_energy = [], [], [], []
	all_recon_err, all_labels = [], []

	with torch.no_grad():
	for images, labels in test_loader:
	images = images.to(device)
	out = model(images)

	all_S_full.append(out['svd']['S_full'].cpu())
	all_S_kept.append(out['svd']['S'].cpu())
	all_erank.append(model.effective_rank(out['svd']['S_full']).cpu())
	all_energy.append(model.energy_ratio(out['svd']['S'], out['svd']['S_full']).cpu())
	all_recon_err.append(
	F.mse_loss(out['recon'], images, reduction='none')
	.mean(dim=(1, 2, 3)).cpu())
	all_labels.append(labels.cpu())

	all_S_full = torch.cat(all_S_full)
	all_S_kept = torch.cat(all_S_kept)
	all_erank = torch.cat(all_erank)
	all_energy = torch.cat(all_energy)
	all_recon_err = torch.cat(all_recon_err)
	all_labels = torch.cat(all_labels)

	print(f"\n Bottleneck: {keep_k} / {min(model.matrix_h, model.matrix_k)} singular values")
	print(f" Energy captured: {all_energy.mean():.3f} ± {all_energy.std():.3f}")
	print(f" Effective rank (full): {all_erank.mean():.2f} ± {all_erank.std():.2f}")
	print(f" Recon MSE: {all_recon_err.mean():.6f} ± {all_recon_err.std():.6f}")

	# Singular value profile
	S_mean = all_S_full.mean(0)
	total_energy = (S_mean ** 2).sum()
	print(f"\n Singular value profile:")
	cumulative = 0
	for i in range(min(32, len(S_mean))):
	e = (S_mean[i] ** 2).item()
	cumulative += e
	pct = cumulative / total_energy * 100
	bar = "█" * int(S_mean[i].item() * 30 / (S_mean[0].item() + 1e-8))
	marker = " ← k" if i == keep_k - 1 else (" ┄ tail" if i == keep_k else "")
	print(f" S[{i:2d}]: {S_mean[i]:8.3f} cum_energy={pct:5.1f}% {bar}{marker}")

	# Per-class
	cifar_names = ['plane', 'car', 'bird', 'cat', 'deer',
	'dog', 'frog', 'horse', 'ship', 'truck']
	print(f"\n Per-class:")
	print(f" {'class':>6} {'erank':>6} {'energy':>6} {'recon':>8} {'S0':>7} {'S1':>7}")
	for c in range(10):
	mask = all_labels == c
	er = all_erank[mask].mean().item()
	en = all_energy[mask].mean().item()
	rc = all_recon_err[mask].mean().item()
	s0 = all_S_full[mask, 0].mean().item()
	s1 = all_S_full[mask, 1].mean().item()
	print(f" {cifar_names[c]:>6} {er:6.2f} {en:6.3f} {rc:8.6f} {s0:7.3f} {s1:7.3f}")

	# Recon by energy quartile
	print(f"\n Recon quality by energy quartile:")
	q25, q50, q75 = all_energy.quantile(torch.tensor([0.25, 0.50, 0.75]))
	for label, lo, hi in [("Q1 (low energy)", 0, q25),
	("Q2", q25, q50),
	("Q3", q50, q75),
	("Q4 (high energy)", q75, 1.1)]:
	mask = (all_energy >= lo) & (all_energy < hi)
	if mask.sum() > 0:
	recon = all_recon_err[mask].mean().item()
	er = all_erank[mask].mean().item()
	print(f" {label:>17}: n={mask.sum():5d} recon={recon:.6f} erank={er:.2f}")

	# ── Save reconstruction grid ──
	print(f"\n Saving reconstruction grid...")
	import matplotlib
	matplotlib.use('Agg')
	import matplotlib.pyplot as plt

	# CIFAR-10 denormalization
	mean = torch.tensor([0.4914, 0.4822, 0.4465]).reshape(1, 3, 1, 1).to(device)
	std = torch.tensor([0.2470, 0.2435, 0.2616]).reshape(1, 3, 1, 1).to(device)

	model.eval()
	with torch.no_grad():
	# Grab one batch
	images, labels = next(iter(test_loader))
	images = images.to(device)
	out = model(images)

	# Pick 2 samples per class = 20 images
	selected_idx = []
	for c in range(10):
	class_idx = (labels == c).nonzero(as_tuple=True)[0]
	selected_idx.extend(class_idx[:2].tolist())

	orig = images[selected_idx]
	recon_full = out['recon'][selected_idx]

	# Progressive reconstructions: 1, 4, 8, 16 modes
	U = out['svd']['U'][selected_idx]
	S = out['svd']['S'][selected_idx]
	Vt = out['svd']['Vt'][selected_idx]

	mode_counts = [1, 4, 8, keep_k]
	# deduplicate if keep_k is already in the list
	mode_counts = list(dict.fromkeys(mode_counts))
	prog_recons = []
	for n_modes in mode_counts:
	n_modes = min(n_modes, S.shape[1])
	U_n = U[:, :, :n_modes]
	S_n = S[:, :n_modes]
	Vt_n = Vt[:, :n_modes, :]
	r = model.decode_from_svd(U_n, S_n, Vt_n)
	prog_recons.append(r)

	def denorm(t):
	return (t * std + mean).clamp(0, 1).cpu()

	n_samples = len(selected_idx)
	n_cols = 2 + len(mode_counts) # original + progressives + error
	fig, axes = plt.subplots(n_samples, n_cols, figsize=(n_cols * 1.5, n_samples * 1.5))

	col_titles = ['Original'] + [f'{m} mode{"s" if m > 1 else ""}' for m in mode_counts] + ['\|Error\|×5']

	for i in range(n_samples):
	# Original
	img_orig = denorm(orig[i:i+1])[0].permute(1, 2, 0).numpy()
	axes[i, 0].imshow(img_orig)

	# Progressive
	for j, r in enumerate(prog_recons):
	img_r = denorm(r[i:i+1])[0].permute(1, 2, 0).numpy()
	axes[i, j+1].imshow(img_r)

	# Error map (amplified 5×)
	err_col = 1 + len(prog_recons)
	diff = (denorm(orig[i:i+1]) - denorm(prog_recons[-1][i:i+1])).abs() * 5
	diff = diff.clamp(0, 1)[0].permute(1, 2, 0).numpy()
	axes[i, err_col].imshow(diff)

	# Class label
	c = labels[selected_idx[i]].item()
	axes[i, 0].set_ylabel(cifar_names[c], fontsize=8, rotation=0, labelpad=35)

	for j, title in enumerate(col_titles):
	axes[0, j].set_title(title, fontsize=8)

	for ax in axes.flat:
	ax.axis('off')

	plt.tight_layout()
	plt.savefig('/content/svae_recon_grid.png', dpi=200, bbox_inches='tight')
	print(f" Saved to /content/svae_recon_grid.png")
	try:
	plt.show()
	except:
	pass
	plt.close()


	if __name__ == "__main__":
	train(keep_k=16)