dl_tutorial/examples/pytorch06_grad_descent.py at master · mint-lab/dl_tutorial · GitHub

Repository navigation

History

import torch

import matplotlib.pyplot as plt

f = lambda x: 0.1*x**3 - 0.8*x**2 - 1.5*x + 5.4

viz_range = torch.FloatTensor([-6, 12])

learn_rate = 0.1

max_iter = 100

min_tol = 1e-6

x_init = 12.

# Prepare visualization

xs = torch.linspace(*viz_range, 100)

plt.plot(xs, f(xs), 'r-', label='f(x)', linewidth=2)

plt.plot(x_init, f(x_init), 'b.', label='Each step', markersize=12)

plt.axis((*viz_range, *f(viz_range)))

plt.legend()

x = torch.tensor(x_init, requires_grad=True)

for i in range(max_iter):

# Derive gradient with Autograd

if x.grad != None:

x.grad.zero_() # Reset gradient tracking

y = f(x) # Calculate the function (forward)

y.backward() # Calculate the gradient (backward)

# Run the gradient descent

xp = x.clone().detach() # Note) xp = x

with torch.no_grad(): # Disable gradient tracking

x -= learn_rate*x.grad # Note) x = x - learn_rate*fd(x) is an original code.

# x = x - learn_rate*x.grad() does not work!

# Update visualization for each iteration

print(f'Iter: {i}, x = {xp:.3f} to {x:.3f}, f(x) = {f(xp):.3f} to {f(x):.3f} (f\'(x) = {x.grad:.3f})')

lcolor = torch.rand(3).tolist()

approx = x.grad*(xs-xp) + f(xp)

plt.plot(xs, approx, '-', linewidth=1, color=lcolor, alpha=0.5)

xc = x.clone().detach() # Copy 'x' for plotting

plt.plot(xc, f(xc), '.', color=lcolor, markersize=12)

# Check the terminal condition

if abs(x - xp) < min_tol:

break;

plt.show()

dl_tutorial/examples/pytorch06_grad_descent.py at master · mint-lab/dl_tutorial · GitHub

Repository navigation

Expand file tree

Breadcrumbs

pytorch06_grad_descent.py

Latest commit

History

Breadcrumbs

pytorch06_grad_descent.py

File metadata and controls