PLOT_STEPS = 55
x_ = np.linspace(-5, 5, 500)
y_ = np.linspace(-2.2, 2.2, 500)
X, Y = np.meshgrid(x_, y_)
Z = loss(X, Y)
fig = plt.determine(figsize=(16, 10), facecolor="#FAFAF8")
gs = GridSpec(2, 3, determine=fig, hspace=0.45, wspace=0.38,
left=0.07, proper=0.97, prime=0.88, backside=0.08)
COLORS = {
"gd": "#E05C4B",
"mom_good": "#3A7CA5",
"mom_large": "#F4A536",
"contour": "#D4C9B8",
"minima": "#2A9D5C",
"begin": "#444444",
}
PANEL_TITLES = [
"Vanilla Gradient DescentnOscillates, slow (185 steps to converge)",
"Momentum β = 0.90nSmooth, fast (159 steps to converge)",
"Momentum β = 0.99 (too large)nOvershoots -- never converges",
]
paths_plot = [
path_gd[:PLOT_STEPS+1],
path_mom_good[:PLOT_STEPS+1],
path_mom_large[:PLOT_STEPS+1],
]
colours = [COLORS["gd"], COLORS["mom_good"], COLORS["mom_large"]]
# prime row: trajectory panels
for col, (path, shade, title) in enumerate(zip(paths_plot, colours, PANEL_TITLES)):
ax = fig.add_subplot(gs[0, col])
ax.set_facecolor("#F5F3EE")
ranges = np.geomspace(0.005, 3.5, 28)
ax.contour(X, Y, Z, ranges=ranges, colours=COLORS["contour"],
linewidths=0.7, alpha=0.9)
ax.plot(path[:, 0], path[:, 1], shade=shade, lw=1.8, alpha=0.85, zorder=3)
ax.scatter(path[:, 0], path[:, 1], shade=shade, s=18, zorder=4, alpha=0.6)
ax.scatter(*path[0], marker="o", s=90, shade=COLORS["start"], zorder=5, label="begin")
ax.scatter(*path[-1], marker="*", s=120, shade=COLORS["minima"], zorder=5, label="finish")
ax.scatter(0, 0, marker="+", s=200, shade=COLORS["minima"], linewidths=2.5, zorder=6)
ax.set_xlim(-5, 5)
ax.set_ylim(-2.2, 2.2)
ax.set_title(title, fontsize=9.5, fontweight="daring", shade="#222", pad=7, loc="left")
ax.set_xlabel("θ₁ (sluggish route)", fontsize=8, shade="#666")
ax.set_ylabel("θ₂ (quick route)", fontsize=8, shade="#666")
ax.tick_params(labelsize=7, colours="#888")
for backbone in ax.spines.values():
backbone.set_edgecolor("#CCCCCC")
# bottom-left: loss curves (full 300 steps)
ax_loss = fig.add_subplot(gs[1, :2])
ax_loss.set_facecolor("#F5F3EE")
full_paths = [path_gd, path_mom_good, path_mom_large]
full_labels = ["Vanilla GD (185 steps)", "Momentum β=0.90 (159 steps)", "Momentum β=0.99 (diverges)"]
for path, shade, label in zip(full_paths, colours, full_labels):
losses = [loss(*p) for p in path]
steps_range = np.arange(len(path))
ax_loss.plot(steps_range, losses, shade=shade, lw=2, label=label, alpha=0.9)
ax_loss.axhline(0.001, shade="#999", lw=1, ls="--", alpha=0.6)
ax_loss.textual content(305, 0.001, "convergencenthreshold", fontsize=7, shade="#888", va="heart")
ax_loss.set_yscale("log")
ax_loss.set_xlim(0, STEPS)
ax_loss.set_title("Loss vs. Optimisation Step (log scale, 300 steps)",
fontsize=10.5, fontweight="daring", shade="#222", loc="left")
ax_loss.set_xlabel("Step", fontsize=9, shade="#666")
ax_loss.set_ylabel("Loss f(θ)", fontsize=9, shade="#666")
ax_loss.legend(fontsize=8.5, framealpha=0.6)
ax_loss.tick_params(labelsize=8, colours="#888")
for backbone in ax_loss.spines.values():
backbone.set_edgecolor("#CCCCCC")
# bottom-right: annotation panel
ax_ann = fig.add_subplot(gs[1, 2])
ax_ann.set_facecolor("#F5F3EE")
ax_ann.axis("off")
annotation = (
"Replace rulesnn"
"Vanilla GDn"
" θ ← θ − α·∇L(θ)nn"
"Momentum GDn"
" v ← β·v + (1−β)·∇L(θ)n"
" θ ← θ − α·vnn"
"Key intuitionn"
" v accumulates previous gradients.n"
" Vertical oscillations cancel out.n"
" Horizontal steps compound.nn"
"Hyperparameter βn"
" β → 0 : behaves like GDn"
" β = 0.9: typical candy spotn"
" β → 1 : overshoots / diverges"
)
ax_ann.textual content(0.05, 0.97, annotation, rework=ax_ann.transAxes,
fontsize=8.8, va="prime", ha="left",
fontfamily="monospace", shade="#333", linespacing=1.7)
fig.suptitle("Momentum in Gradient Descent",
fontsize=16, fontweight="daring", shade="#111", y=0.95)
plt.savefig("momentum_explainer.png", dpi=150, bbox_inches="tight",
facecolor=fig.get_facecolor())
plt.present()