In [1]:
import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
from mpl_toolkits.mplot3d import Axes3D # required even though not ref'd!
import random
from matplotlib import rcParams
EXACT_COEFF = [-2977.3484892930505, 113.13288354356122]
rcParams["font.size"] = 12
def line_fit(B, x):
return B[2-1]*x + B[1-1]
In [2]:
df = pd.read_csv('data/cheese_deaths.csv')
df
Out[2]:
In [4]:
fig, ax1 = plt.subplots(figsize=(8,2.5))
...
plt.show()
In [3]:
...
In [5]:
def MSE(B):
"""
Compute and return the mean squared error with:
(x_i, y_i) = (df.cheese, df.deaths).
The line coefficients are B = [y-intercept, slope].
"""
In [18]:
def MSE_gradient(B):
"""
Compute and return the gradient vector containing the partial derivatives
of the MSE cost function with respect to b_1 and b_2. Don't scale by 2
and divide by the number of observations as it's a waste of computation.
We can adjust the learning rate to take this into consideration.
"""
In [19]:
def minimize(B0, eta, precision):
"""
Minimize the MSE cost function starting at model parameters B0
and using learning rate vector eta. The model has converged when
the change in MSE from one iteration to the next is less than
precision. Return both the B at which your algorithm converged
and also the list of B points between B0 and converged B, inclusively.
"""
trace = []
In [8]:
LEARNING_RATES = [0.02,0.0001]
PRECISION = 0.000000001 # can't be too small as x-xprev prec is low
B0 = np.array([random.randrange(-6000, 4000), random.randrange(-200, 300)])
In [20]:
%time B,trace = minimize(B0, LEARNING_RATES, PRECISION)
In [9]:
print(f"descent gives {B} in {len(trace)} steps")
print(f"exact is {EXACT_COEFF}")
print(f"cost of approx is {MSE(B):1.10f}")
print(f"cost of exact is {MSE(EXACT_COEFF):1.10f}")
In [10]:
def get_surface(Cost, b1, b2):
"""
Given an arbitrary function of two scalar parameters, f,
return a list of lists, C, containing f evaluated at f(b1[i],b2[j]). i is the
b_1 coordinate and j is the b_2 coordinate. b_1 is the np.arange of
y-intercept values and b_2 is the np.arange of slope values.
"""
In [13]:
def heatmap(b0, b1, C, trace): # trace is a list of [b1, b2] pairs
"""
Given b0 and b1, y-intercept and slope values, and cost 2D matrix, plot
a heatmap using plt.imshow(). Make the heat map look like those in
the project description.
Once the heat map is displayed, plot the trace coordinates in color
#FEE08F. Plot the first coordinate in color #FB060B and the last
coordinate in black.
Returns nothing.
"""
fig = plt.figure(figsize=(6,6))
...
plt.imshow(C,
origin='lower',
cmap='coolwarm',
alpha=0.65,
...
)
...
# Plot all trace cordinates at once instead of using a Python loop
# (MUCH faster this way)
ax.plot(trace[:,0], trace[:,1], color='#FEE08F', marker='.', markersize=.1)
...
In [14]:
b1 = np.arange(-6000, 4000, 30) # y intercept
b2 = np.arange(-200, 300, 1) # slope
C = get_surface(MSE, b1, b2)
heatmap(b1, b2, C, trace)