2
$\begingroup$

I would really appreciate your help on a rather simple issue that I just can't solve on my own. I'd like to visualize gradient descent for a simple linear regression problem with two parameters $\theta_0$ (intercept) and $\theta_1$ (slope) using mean squared error as the cost function. What I'd like to visualize is a 3d plot of the cost function and the corresponding tangent plane for different $\theta_0$, $\theta_1$.

Here's what I have so far:

Cost function: $$J(\theta_0,\theta_1)=\frac{1}{2m}\sum_{i=1}^{m}{(\theta_0 + \theta_1 x_{1}^{(i)}-y^{(i)})^2}$$

Partial derivative wrt $\theta_0$: $$ J_{\theta_0}(\theta_0, \theta_1) = \frac{1}{m} \sum_{i=1}^{m}{(\theta_0 + \theta_1 x_{1}^{(i)} - y^{(i)})} $$

Partial derivative wrt $\theta_1$: $$ J_{\theta_1}(\theta_0, \theta_1) = \frac{1}{m} \sum_{i=1}^{m}{(\theta_0 + \theta_1 x_{1}^{(i)} - y^{(i)})(x_{1}^{(i)})} $$

Let's say my tangent plane kisses the loss function surface plot at $(\theta_0,\theta_1,z)=(1,2,3)$. So if I want to know the z-value at $(\theta_0,\theta_1)=(4,5)$ I would to the following: $$ z = J_{\theta_0}(1,2)(\theta_0-1)+J_{\theta_1}(1,2)(\theta_1-2)+3 $$ $$ z = J_{\theta_0}(1,2)(4-1)+J_{\theta_1}(1,2)(5-2)+3 $$

However, if I implement all of this in python my tangent plane won't end up where I would expect it to be. Here's my code:

import sympy as sp
import numpy as np
import plotly.graph_objects as go
import dash_html_components as html
import dash_core_components as dcc

# Generate a linear dataset
m = 100
X = 2 * np.random.rand(m, 1)
y = 4 + 3 * X + np.random.randn(m, 1)

# Append ones for bias term
X_b = np.c_[np.ones(len(X)),X]

# Loss function (j), its derivatives (djt0, djt1) and equation for tangent plane (z)
j = lambda theta0,theta1: 1/2*1/m*((X_b.dot(np.array([[theta0],[theta1]]))-y)**2).sum()
djt0 = lambda theta0,theta1: 1/m*(X_b.dot(np.array([[theta0],[theta1]]))-y).sum()
djt1 = lambda theta0,theta1: 1/m*((X_b.dot(np.array([[theta0],[theta1]]))-y)*X_b[:,1]).sum()
z = lambda theta0,theta1: djt0(10,10)*(theta0-10)+djt1(10,10)*(theta1-10)+j(10,10)

# setup sympy
theta0,theta1 = sp.symbols('theta0,theta1')
loss_surface = sp.lambdify((theta0,theta1), j(theta0,theta1))
tangent_plane = sp.lambdify((theta0,theta1), z(theta0,theta1))

# grid and plot
points = np.linspace(-20,20,90)
xgrid,ygrid=np.meshgrid(points,points)

fig = go.Figure(
    data=[
        go.Surface(z=loss_surface(xgrid,ygrid),x=xgrid,y=ygrid),
        go.Surface(z=tangent_plane(xgrid,ygrid),x=xgrid,y=ygrid)
    ]
)

fig.update_layout(
    scene = dict(
        zaxis = dict(range=[0,1600]))
)

fig.show()

Note: the z-lambda function is set for $(\theta_0,\theta_1)=(10,10)$. So I would expect to see to see a tangent plane touching the loss surface at $(10,10)$.

Here's my current result: loss function looks okay, tangent plane not really

Is there anything obviously wrong in the equations? If not, then I probably messed up the matrix computation in python.

Thank you!

$\endgroup$
1
$\begingroup$

Solved it! One of the numpy arrays had the wrong shape, which messed up the $J_{\theta_1}(\theta_0,\theta_1)$ computation.

This fixed the problem:

djt1 = lambda theta0,theta1: 1/m*((X_b.dot(np.array([[theta0],[theta1]]))-y)*X_b[:,1].reshape(-1,1).sum()

Now the results look perfect and I can plot tangent planes for different $\theta_0$, $\theta_1$ values.

enter image description here

Hope that helps.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.