本文主要是介绍Introduction to Advanced Machine Learning, 第一周, week01_pa(hse-aml/intro-to-dl,简单注释,答案,附图),希望对大家解决编程问题提供一定的参考价值,需要的开发者们随着小编来一起学习吧!
这是俄罗斯高等经济学院的系列课程第一门,Introduction to Advanced Machine Learning,第一周编程作业。
这个作业一共六个任务,难易程度:容易。
1. 计算probability
2. 计算loss function
3. 计算stochastic gradient
4. 计算mini-batch gradient
5. 计算momentum gradient
6. 计算RMS prop gradient
从3到6,收敛应该越来越快,越来越稳定。
Programming assignment (Linear models, Optimization)
In this programming assignment you will implement a linear classifier and train it using stochastic gradient descent modifications and numpy.
import numpy as np
%matplotlib inline
import matplotlib.pyplot as plt
import sys
sys.path.append("..")
import grading
Two-dimensional classification
To make things more intuitive, let’s solve a 2D classification problem with synthetic data.
with open('train.npy', 'rb') as fin:X = np.load(fin)with open('target.npy', 'rb') as fin:y = np.load(fin)plt.scatter(X[:, 0], X[:, 1], c=y, cmap=plt.cm.Paired, s=20)
plt.show()
!
Task
Features
As you can notice the data above isn’t linearly separable. Since that we should add features (or use non-linear model). Note that decision line between two classes have form of circle, since that we can add quadratic features to make the problem linearly separable. The idea under this displayed on image below:
def expand(X):"""Adds quadratic features. This expansion allows your linear model to make non-linear separation.For each sample (row in matrix), compute an expanded row:[feature0, feature1, feature0^2, feature1^2, feature0*feature1, 1]:param X: matrix of features, shape [n_samples,2]:returns: expanded features of shape [n_samples,6]"""X_expanded = np.ones((X.shape[0], 6))X_expanded[:,0] = X[:,0]X_expanded[:,1] = X[:,1]X_expanded[:,2] = X[:,0] * X[:,0]X_expanded[:,3] = X[:,1] * X[:,1]X_expanded[:,4] = X[:,0] * X[:,1]return X_expanded
X_expanded = expand(X)
print(X_expanded)
[[ 1.20798057 0.0844994 1.45921706 0.00714015 0.10207364 1. ][ 0.76121787 0.72510869 0.57945265 0.52578261 0.5519657 1. ][ 0.55256189 0.51937292 0.30532464 0.26974823 0.28698568 1. ]..., [-1.22224754 0.45743421 1.49388906 0.20924606 -0.55909785 1. ][ 0.43973452 -1.47275142 0.19336645 2.16899674 -0.64761963 1. ][ 1.4928118 1.15683375 2.22848708 1.33826433 1.72693508 1. ]]
Here are some tests for your implementation of expand
function.
# simple test on random numbersdummy_X = np.array([[0,0],[1,0],[2.61,-1.28],[-0.59,2.1]])# call your expand function
dummy_expanded = expand(dummy_X)# what it should have returned: x0 x1 x0^2 x1^2 x0*x1 1
dummy_expanded_ans = np.array([[ 0. , 0. , 0. , 0. , 0. , 1. ],[ 1. , 0. , 1. , 0. , 0. , 1. ],[ 2.61 , -1.28 , 6.8121, 1.6384, -3.3408, 1. ],[-0.59 , 2.1 , 0.3481, 4.41 , -1.239 , 1. ]])#tests
assert isinstance(dummy_expanded,np.ndarray), "please make sure you return numpy array"
assert dummy_expanded.shape == dummy_expanded_ans.shape, "please make sure your shape is correct"
assert np.allclose(dummy_expanded,dummy_expanded_ans,1e-3), "Something's out of order with features"print("Seems legit!")
Seems legit!
Logistic regression
To classify objects we will obtain probability of object belongs to class ‘1’. To predict probability we will use output of linear model and logistic function:
def probability(X, w):"""Given input features and weightsreturn predicted probabilities of y==1 given x, P(y=1|x), see description aboveDon't forget to use expand(X) function (where necessary) in this and subsequent functions.:param X: feature matrix X of shape [n_samples,6] (expanded):param w: weight vector w of shape [6] for each of the expanded features:returns: an array of predicted probabilities in [0,1] interval."""# TODO:<your code here>prob = 1/(1+np.exp(-np.dot(X,w)))
dummy_weights = np.linspace(-1, 1, 6)
ans_part1 = probability(X_expanded[:1, :], dummy_weights)[0]
## GRADED PART, DO NOT CHANGE!
grader.set_answer("xU7U4", ans_part1)
# you can make submission with answers so far to check yourself at this stage
grader.submit(COURSERA_EMAIL, COURSERA_TOKEN)
In logistic regression the optimal parameters w w are found by cross-entropy minimization:
def compute_loss(X, y, w):"""Given feature matrix X [n_samples,6], target vector [n_samples] of 1/0,and weight vector w [6], compute scalar loss function using formula above."""# TODO:<your code here>prob = probability(X,w)n_sample = X.shape[0]loss = -sum(y * np.log(prob) + (1-y) * np.log(1-prob))/n_sample
# use output of this cell to fill answer field
ans_part2 = compute_loss(X_expanded, y, dummy_weights)
## GRADED PART, DO NOT CHANGE!
grader.set_answer("HyTF6", ans_part2)
# you can make submission with answers so far to check yourself at this stage
grader.submit(COURSERA_EMAIL, COURSERA_TOKEN)
Since we train our model with gradient descent, we should compute gradients.
To be specific, we need a derivative of loss function over each weight [6 of them].
We won’t be giving you the exact formula this time — instead, try figuring out a derivative with pen and paper.
As usual, we’ve made a small test for you, but if you need more, feel free to check your math against finite differences (estimate how L L changes if you shift by 10−5 10 − 5 or so).
def compute_grad(X, y, w):"""Given feature matrix X [n_samples,6], target vector [n_samples] of 1/0,and weight vector w [6], compute vector [6] of derivatives of L over each weights."""# X [n,d] n examples, d features# y [n,] n examples, outputs# w [d,] d features# grad[d]# np.dot(X.T, dz) [d,n][n,] = [d,]# TODO<your code here>a = probability(X,w)dz = a - y #[n,]grad = -1.0 / X.shape[0] * np.dot(X.T, dz)# because the minus here, the following update is positive, instead of negative.
# use output of this cell to fill answer field
ans_part3 = np.linalg.norm(compute_grad(X_expanded, y, dummy_weights))
## GRADED PART, DO NOT CHANGE!
grader.set_answer("uNidL", ans_part3)
# you can make submission with answers so far to check yourself at this stage
grader.submit(COURSERA_EMAIL, COURSERA_TOKEN)
Here’s an auxiliary function that visualizes the predictions:
from IPython import displayh = 0.01
x_min, x_max = X[:, 0].min() - 1, X[:, 0].max() + 1
y_min, y_max = X[:, 1].min() - 1, X[:, 1].max() + 1
xx, yy = np.meshgrid(np.arange(x_min, x_max, h), np.arange(y_min, y_max, h))def visualize(X, y, w, history):"""draws classifier prediction with matplotlib magic"""Z = probability(expand(np.c_[xx.ravel(), yy.ravel()]), w)Z = Z.reshape(xx.shape)plt.subplot(1, 2, 1)plt.contourf(xx, yy, Z, alpha=0.8)plt.scatter(X[:, 0], X[:, 1], c=y, cmap=plt.cm.Paired)plt.xlim(xx.min(), xx.max())plt.ylim(yy.min(), yy.max())plt.subplot(1, 2, 2)plt.plot(history)plt.grid()ymin, ymax = plt.ylim()plt.ylim(0, ymax)display.clear_output(wait=True)plt.show()
visualize(X, y, dummy_weights, [0.5, 0.5, 0.25])
Training
In this section we’ll use the functions you wrote to train our classifier using stochastic gradient descent.
You can try change hyperparameters like batch size, learning rate and so on to find the best one, but use our hyperparameters when fill answers.
Mini-batch SGD
Stochastic gradient descent just takes a random example on each iteration, calculates a gradient of the loss on it and makes a step:
# please use np.random.seed(42), eta=0.1, n_iter=100 and batch_size=4 for deterministic resultsnp.random.seed(42)
w = np.array([0, 0, 0, 0, 0, 1])eta= 0.1 # learning raten_iter = 100
batch_size = 4
loss = np.zeros(n_iter)
plt.figure(figsize=(12, 5))for i in range(n_iter):ind = np.random.choice(X_expanded.shape[0], batch_size)loss[i] = compute_loss(X_expanded, y, w)if i % 10 == 0:visualize(X_expanded[ind, :], y[ind], w, loss)# TODO:<your code here>grad = compute_grad(X_expanded[ind, :],y[ind],w)w = w + eta * grad
visualize(X, y, w, loss)
plt.clf()
# use output of this cell to fill answer field ans_part4 = compute_loss(X_expanded, y, w)
## GRADED PART, DO NOT CHANGE!
grader.set_answer("ToK7N", ans_part4)
# you can make submission with answers so far to check yourself at this stage
grader.submit(COURSERA_EMAIL, COURSERA_TOKEN)
SGD with momentum
Momentum is a method that helps accelerate SGD in the relevant direction and dampens oscillations as can be seen in image below. It does this by adding a fraction α α of the update vector of the past time step to the current update vector.
# please use np.random.seed(42), eta=0.05, alpha=0.9, n_iter=100 and batch_size=4 for deterministic results
np.random.seed(42)
w = np.array([0, 0, 0, 0, 0, 1])eta = 0.05 # learning rate
alpha = 0.9 # momentum
nu = np.zeros_like(w)n_iter = 100
batch_size = 4
loss = np.zeros(n_iter)
plt.figure(figsize=(12, 5))for i in range(n_iter):ind = np.random.choice(X_expanded.shape[0], batch_size)loss[i] = compute_loss(X_expanded, y, w)if i % 10 == 0:visualize(X_expanded[ind, :], y[ind], w, loss)# TODO:<your code here>nu = alpha * nu + eta * compute_grad(X_expanded[ind, :],y[ind],w)w = w + nu
visualize(X, y, w, loss)
plt.clf()
# use output of this cell to fill answer field ans_part5 = compute_loss(X_expanded, y, w)
## GRADED PART, DO NOT CHANGE!
grader.set_answer("GBdgZ", ans_part5)
# you can make submission with answers so far to check yourself at this stage
grader.submit(COURSERA_EMAIL, COURSERA_TOKEN)
RMSprop
Implement RMSPROP algorithm, which use squared gradients to adjust learning rate:
# please use np.random.seed(42), eta=0.1, alpha=0.9, n_iter=100 and batch_size=4 for deterministic results
np.random.seed(42)w = np.array([0, 0, 0, 0, 0, 1.])eta = 0.1 # learning rate
alpha = 0.9 # moving average of gradient norm squared
g2 = np.zeros_like(w)
eps = 1e-8n_iter = 100
batch_size = 4
loss = np.zeros(n_iter)
plt.figure(figsize=(12,5))
for i in range(n_iter):ind = np.random.choice(X_expanded.shape[0], batch_size)loss[i] = compute_loss(X_expanded, y, w)if i % 10 == 0:visualize(X_expanded[ind, :], y[ind], w, loss)# TODO:<your code here>grad = compute_grad(X_expanded[ind, :],y[ind],w)g2 = alpha * g2 + (1-alpha) * grad ** 2w = w + eta/np.sqrt(g2 + eps) * grad
visualize(X, y, w, loss)
plt.clf()
# use output of this cell to fill answer field
ans_part6 = compute_loss(X_expanded, y, w)
## GRADED PART, DO NOT CHANGE!
grader.set_answer("dLdHG", ans_part6)
grader.submit(COURSERA_EMAIL, COURSERA_TOKEN)
这篇关于Introduction to Advanced Machine Learning, 第一周, week01_pa(hse-aml/intro-to-dl,简单注释,答案,附图)的文章就介绍到这儿,希望我们推荐的文章对编程师们有所帮助!