mini batch and stochastic
ipynb6 kBPython · 3 cellsExtra material of chapter 3
# -*- coding: utf-8 -*- mini Batch
"""
Mini-batch Gradient Descent on f(x)=x^2 — Smooth Video Animation (NO SAVING)
"""
import numpy as np
import matplotlib
matplotlib.use("Qt5Agg") # IMPORTANT for real animations
import matplotlib.pyplot as plt
# Target function
def f(x):
return x**2
# Visualization grid
xx = np.linspace(-100, 100, 800)
yy = f(xx)
# Hyperparameters
x = 80.0 # starting point
eta = 0.01 # learning rate
steps = 600 # number of iterations
m = 32 # mini-batch size
rng = np.random.default_rng(0)
# Enable interactive mode
plt.ion()
# Create persistent figure ONCE
fig, ax = plt.subplots(figsize=(8, 6))
ax.plot(xx, yy, linewidth=2, label=r"$f(x)=x^2$")
point, = ax.plot([], [], "ro", markersize=10, label="Current $x_t$")
path, = ax.plot([], [], "b--", alpha=0.4) # trail (history)
ax.set_xlim(-100, 100)
ax.set_ylim(0, 10000)
ax.set_xlabel("x")
ax.set_ylabel("f(x)")
ax.legend()
# To store trajectory for smoother visual effect
xs, ys = [], []
# Main animation loop
for t in range(steps):
# Sample mini-batch Z_1,...,Z_m ~ N(0,1)
z_batch = rng.normal(0.0, 1.0, size=m)
# Gradient: g = 2(a_i - z_j) mean → g = 2(x - mean(z_j))
g = 2.0 * (x - z_batch.mean())
# Update
x = x - eta * g
# Store for trail
xs.append(x)
ys.append(f(x))
# Update moving point + trail
point.set_data([x], [f(x)])
path.set_data(xs, ys)
ax.set_title(
f"Mini-batch GD on $x^2$ | m={m} | iter={t} | x={x:.2f} | g=2(x - mean(z))"
)
# Smooth animation
fig.canvas.draw()
fig.canvas.flush_events()
plt.pause(0.01)
plt.ioff()
plt.show()#Stochastic
# -*- coding: utf-8 -*-
"""
Stochastic Gradient Descent on f(x)=x^2 — Smooth Video Animation (NO SAVING)
"""
import numpy as np
import matplotlib
matplotlib.use("Qt5Agg") # <<< IMPORTANT for REAL animation
import matplotlib.pyplot as plt
# Target function
def f(x):
return x**2
# Visualization grid
xx = np.linspace(-100, 100, 800)
yy = f(xx)
# SGD hyperparameters
x = 80.0 # initial point
eta = 0.7 # learning rate
steps = 800 # number of updates
rng = np.random.default_rng(0) # reproducible randomness
# Activate interactive mode
plt.ion()
# Create persistent figure
fig, ax = plt.subplots(figsize=(8, 6))
ax.plot(xx, yy, linewidth=2, label=r"$f(x)=x^2$")
point, = ax.plot([], [], "ro", markersize=10, label="Current $x_t$")
path, = ax.plot([], [], "b--", alpha=0.4, label="Trajectory")
ax.set_xlim(-100, 100)
ax.set_ylim(0, 10000)
ax.set_xlabel("x")
ax.set_ylabel("f(x)")
ax.legend()
# Store trajectory
xs, ys = [], []
# Animation loop
for t in range(steps):
# Draw one sample Z ~ N(0,1)
z = rng.normal(0.0, 1.0)
# Stochastic gradient: g = 2(x - z)
g = 2.0 * (x - z)
# Update
x = x - eta * g
# Record trajectory
xs.append(x)
ys.append(f(x))
# Update point + trail
point.set_data([x], [f(x)])
path.set_data(xs, ys)
ax.set_title(
f"Stochastic GD on $x^2$ | iter={t} | x={x:.2f} | g = 2(x - z)"
)
# Smooth animation
fig.canvas.draw()
fig.canvas.flush_events()
plt.pause(0.05)
plt.ioff()
plt.show()