Generative Models: Theory Notebook

This notebook makes generative-model objectives executable: autoregressive likelihood, VAE reparameterization, GAN losses, flow change of variables, diffusion noising, score updates, and FID intuition.

Code cell 2

import numpy as np
import matplotlib.pyplot as plt
import matplotlib as mpl

try:
    import seaborn as sns
    sns.set_theme(style="whitegrid", palette="colorblind")
    HAS_SNS = True
except ImportError:
    plt.style.use("seaborn-v0_8-whitegrid")
    HAS_SNS = False

mpl.rcParams.update({
    "figure.figsize":    (10, 6),
    "figure.dpi":         120,
    "font.size":           13,
    "axes.titlesize":      15,
    "axes.labelsize":      13,
    "xtick.labelsize":     11,
    "ytick.labelsize":     11,
    "legend.fontsize":     11,
    "legend.framealpha":   0.85,
    "lines.linewidth":      2.0,
    "axes.spines.top":     False,
    "axes.spines.right":   False,
    "savefig.bbox":       "tight",
    "savefig.dpi":         150,
})
np.random.seed(42)
print("Plot setup complete.")

1. Autoregressive log likelihood

Code cell 4

conditional_probs = np.array([0.8, 0.5, 0.25, 0.9])
logp = np.log(conditional_probs).sum()
print("log probability:", logp)
print("probability:", np.exp(logp))

2. VAE reparameterization

Code cell 6

rng = np.random.default_rng(0)
mu = np.array([0.5, -0.2])
logvar = np.array([0.0, -1.0])
eps = rng.normal(size=mu.shape)
z = mu + np.exp(0.5 * logvar) * eps
print("epsilon:", np.round(eps, 3))
print("z:", np.round(z, 3))

3. Gaussian KL to standard normal

Code cell 8

kl = 0.5 * np.sum(np.exp(logvar) + mu**2 - 1 - logvar)
print("KL(q(z|x) || N(0,I)):", kl)

4. ELBO pieces

Code cell 10

reconstruction_logprob = -1.25
elbo = reconstruction_logprob - kl
negative_elbo = -elbo
print("ELBO:", elbo)
print("training loss -ELBO:", negative_elbo)

5. GAN losses

Code cell 12

D_real = np.array([0.9, 0.8, 0.7])
D_fake = np.array([0.2, 0.3, 0.4])
D_loss = -(np.log(D_real).mean() + np.log(1 - D_fake).mean())
G_loss_saturating = np.log(1 - D_fake).mean()
G_loss_nonsat = -np.log(D_fake).mean()
print("D loss:", D_loss)
print("G saturating objective:", G_loss_saturating)
print("G non-saturating loss:", G_loss_nonsat)

6. 1D flow change of variables

Code cell 14

# x = a z + b, z=(x-b)/a, p_x(x)=p_z(z)*|dz/dx|
a, b = 2.0, -1.0
x = 1.0
z = (x - b) / a
log_pz = -0.5 * z**2 - 0.5 * np.log(2 * np.pi)
log_px = log_pz - np.log(abs(a))
print("z:", z)
print("log p_x:", log_px)

7. Diffusion noising step

Code cell 16

x0 = np.array([1.0, -0.5, 0.25])
alpha_bar = 0.7
eps = rng.normal(size=x0.shape)
xt = np.sqrt(alpha_bar) * x0 + np.sqrt(1 - alpha_bar) * eps
print("x_t:", np.round(xt, 3))

8. Denoising MSE

Code cell 18

eps_pred = eps + np.array([0.1, -0.2, 0.05])
loss = np.mean((eps - eps_pred) ** 2)
print("noise prediction MSE:", loss)

9. Score-based Langevin update

Code cell 20

x = np.array([2.0, -1.0])
score = -x  # score of standard normal
eta = 0.1
noise = rng.normal(size=x.shape)
x_next = x + eta * score + np.sqrt(2 * eta) * noise
print("x_next:", np.round(x_next, 3))

10. Simplified FID-style distance

Code cell 22

mu_r = np.array([0.0, 0.0])
mu_g = np.array([0.5, -0.25])
cov_r = np.eye(2)
cov_g = np.array([[1.2, 0.0], [0.0, 0.8]])
mean_term = np.sum((mu_r - mu_g) ** 2)
cov_term = np.trace(cov_r + cov_g - 2 * np.diag(np.sqrt(np.diag(cov_r) * np.diag(cov_g))))
print("FID-style diagonal distance:", mean_term + cov_term)

11. Diversity check

Code cell 24

samples = rng.normal(size=(100, 2))
collapsed = np.repeat(samples[:1], repeats=100, axis=0)
print("sample variance normal:", samples.var(axis=0))
print("sample variance collapsed:", collapsed.var(axis=0))

12. Denoising loss by timestep

Code cell 26

timesteps = np.arange(1, 21)
loss_by_t = 0.1 + 0.02 * timesteps + 0.05 * np.sin(timesteps / 2)
plt.plot(timesteps, loss_by_t, marker="o")
plt.title("Toy denoising loss by timestep")
plt.xlabel("timestep")
plt.ylabel("loss")
plt.tight_layout()
plt.show()

13. Final checklist

Code cell 28

checks = [
    "state whether the objective is likelihood, ELBO, adversarial loss, or denoising loss",
    "evaluate sample quality and diversity separately",
    "track likelihood only when it is meaningful for the model family",
    "inspect latent traversals or conditioning controls",
    "report sampling cost and number of generation steps",
    "check for mode collapse or timestep-specific failures",
]
for i, check in enumerate(checks, 1):
    print(f"{i}. {check}")