Python/machine_learning/sequential_minimum_optimization.py

623 lines
20 KiB
Python

"""
Sequential minimal optimization (SMO) for support vector machines (SVM)
Sequential minimal optimization (SMO) is an algorithm for solving the quadratic
programming (QP) problem that arises during the training of SVMs. It was invented by
John Platt in 1998.
Input:
0: type: numpy.ndarray.
1: first column of ndarray must be tags of samples, must be 1 or -1.
2: rows of ndarray represent samples.
Usage:
Command:
python3 sequential_minimum_optimization.py
Code:
from sequential_minimum_optimization import SmoSVM, Kernel
kernel = Kernel(kernel='poly', degree=3., coef0=1., gamma=0.5)
init_alphas = np.zeros(train.shape[0])
SVM = SmoSVM(train=train, alpha_list=init_alphas, kernel_func=kernel, cost=0.4,
b=0.0, tolerance=0.001)
SVM.fit()
predict = SVM.predict(test_samples)
Reference:
https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/smo-book.pdf
https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/tr-98-14.pdf
"""
import os
import sys
import urllib.request
import numpy as np
import pandas as pd
from matplotlib import pyplot as plt
from sklearn.datasets import make_blobs, make_circles
from sklearn.preprocessing import StandardScaler
CANCER_DATASET_URL = (
"https://archive.ics.uci.edu/ml/machine-learning-databases/"
"breast-cancer-wisconsin/wdbc.data"
)
class SmoSVM:
def __init__(
self,
train,
kernel_func,
alpha_list=None,
cost=0.4,
b=0.0,
tolerance=0.001,
auto_norm=True,
):
self._init = True
self._auto_norm = auto_norm
self._c = np.float64(cost)
self._b = np.float64(b)
self._tol = np.float64(tolerance) if tolerance > 0.0001 else np.float64(0.001)
self.tags = train[:, 0]
self.samples = self._norm(train[:, 1:]) if self._auto_norm else train[:, 1:]
self.alphas = alpha_list if alpha_list is not None else np.zeros(train.shape[0])
self.Kernel = kernel_func
self._eps = 0.001
self._all_samples = list(range(self.length))
self._K_matrix = self._calculate_k_matrix()
self._error = np.zeros(self.length)
self._unbound = []
self.choose_alpha = self._choose_alphas()
# Calculate alphas using SMO algorithm
def fit(self):
k = self._k
state = None
while True:
# 1: Find alpha1, alpha2
try:
i1, i2 = self.choose_alpha.send(state)
state = None
except StopIteration:
print("Optimization done!\nEvery sample satisfy the KKT condition!")
break
# 2: calculate new alpha2 and new alpha1
y1, y2 = self.tags[i1], self.tags[i2]
a1, a2 = self.alphas[i1].copy(), self.alphas[i2].copy()
e1, e2 = self._e(i1), self._e(i2)
args = (i1, i2, a1, a2, e1, e2, y1, y2)
a1_new, a2_new = self._get_new_alpha(*args)
if not a1_new and not a2_new:
state = False
continue
self.alphas[i1], self.alphas[i2] = a1_new, a2_new
# 3: update threshold(b)
b1_new = np.float64(
-e1
- y1 * k(i1, i1) * (a1_new - a1)
- y2 * k(i2, i1) * (a2_new - a2)
+ self._b
)
b2_new = np.float64(
-e2
- y2 * k(i2, i2) * (a2_new - a2)
- y1 * k(i1, i2) * (a1_new - a1)
+ self._b
)
if 0.0 < a1_new < self._c:
b = b1_new
if 0.0 < a2_new < self._c:
b = b2_new
if not (np.float64(0) < a2_new < self._c) and not (
np.float64(0) < a1_new < self._c
):
b = (b1_new + b2_new) / 2.0
b_old = self._b
self._b = b
# 4: update error, here we only calculate the error for non-bound samples
self._unbound = [i for i in self._all_samples if self._is_unbound(i)]
for s in self.unbound:
if s in (i1, i2):
continue
self._error[s] += (
y1 * (a1_new - a1) * k(i1, s)
+ y2 * (a2_new - a2) * k(i2, s)
+ (self._b - b_old)
)
# if i1 or i2 is non-bound, update their error value to zero
if self._is_unbound(i1):
self._error[i1] = 0
if self._is_unbound(i2):
self._error[i2] = 0
# Predict test samples
def predict(self, test_samples, classify=True):
if test_samples.shape[1] > self.samples.shape[1]:
raise ValueError(
"Test samples' feature length does not equal to that of train samples"
)
if self._auto_norm:
test_samples = self._norm(test_samples)
results = []
for test_sample in test_samples:
result = self._predict(test_sample)
if classify:
results.append(1 if result > 0 else -1)
else:
results.append(result)
return np.array(results)
# Check if alpha violates the KKT condition
def _check_obey_kkt(self, index):
alphas = self.alphas
tol = self._tol
r = self._e(index) * self.tags[index]
c = self._c
return (r < -tol and alphas[index] < c) or (r > tol and alphas[index] > 0.0)
# Get value calculated from kernel function
def _k(self, i1, i2):
# for test samples, use kernel function
if isinstance(i2, np.ndarray):
return self.Kernel(self.samples[i1], i2)
# for training samples, kernel values have been saved in matrix
else:
return self._K_matrix[i1, i2]
# Get error for sample
def _e(self, index):
"""
Two cases:
1: Sample[index] is non-bound, fetch error from list: _error
2: sample[index] is bound, use predicted value minus true value: g(xi) - yi
"""
# get from error data
if self._is_unbound(index):
return self._error[index]
# get by g(xi) - yi
else:
gx = np.dot(self.alphas * self.tags, self._K_matrix[:, index]) + self._b
yi = self.tags[index]
return gx - yi
# Calculate kernel matrix of all possible i1, i2, saving time
def _calculate_k_matrix(self):
k_matrix = np.zeros([self.length, self.length])
for i in self._all_samples:
for j in self._all_samples:
k_matrix[i, j] = np.float64(
self.Kernel(self.samples[i, :], self.samples[j, :])
)
return k_matrix
# Predict tag for test sample
def _predict(self, sample):
k = self._k
predicted_value = (
np.sum(
[
self.alphas[i1] * self.tags[i1] * k(i1, sample)
for i1 in self._all_samples
]
)
+ self._b
)
return predicted_value
# Choose alpha1 and alpha2
def _choose_alphas(self):
loci = yield from self._choose_a1()
if not loci:
return None
return loci
def _choose_a1(self):
"""
Choose first alpha
Steps:
1: First loop over all samples
2: Second loop over all non-bound samples until no non-bound samples violate
the KKT condition.
3: Repeat these two processes until no samples violate the KKT condition
after the first loop.
"""
while True:
all_not_obey = True
# all sample
print("Scanning all samples!")
for i1 in [i for i in self._all_samples if self._check_obey_kkt(i)]:
all_not_obey = False
yield from self._choose_a2(i1)
# non-bound sample
print("Scanning non-bound samples!")
while True:
not_obey = True
for i1 in [
i
for i in self._all_samples
if self._check_obey_kkt(i) and self._is_unbound(i)
]:
not_obey = False
yield from self._choose_a2(i1)
if not_obey:
print("All non-bound samples satisfy the KKT condition!")
break
if all_not_obey:
print("All samples satisfy the KKT condition!")
break
return False
def _choose_a2(self, i1):
"""
Choose the second alpha using a heuristic algorithm
Steps:
1: Choose alpha2 that maximizes the step size (|E1 - E2|).
2: Start in a random point, loop over all non-bound samples till alpha1 and
alpha2 are optimized.
3: Start in a random point, loop over all samples till alpha1 and alpha2 are
optimized.
"""
self._unbound = [i for i in self._all_samples if self._is_unbound(i)]
if len(self.unbound) > 0:
tmp_error = self._error.copy().tolist()
tmp_error_dict = {
index: value
for index, value in enumerate(tmp_error)
if self._is_unbound(index)
}
if self._e(i1) >= 0:
i2 = min(tmp_error_dict, key=lambda index: tmp_error_dict[index])
else:
i2 = max(tmp_error_dict, key=lambda index: tmp_error_dict[index])
cmd = yield i1, i2
if cmd is None:
return
rng = np.random.default_rng()
for i2 in np.roll(self.unbound, rng.choice(self.length)):
cmd = yield i1, i2
if cmd is None:
return
for i2 in np.roll(self._all_samples, rng.choice(self.length)):
cmd = yield i1, i2
if cmd is None:
return
# Get the new alpha2 and new alpha1
def _get_new_alpha(self, i1, i2, a1, a2, e1, e2, y1, y2):
k = self._k
if i1 == i2:
return None, None
# calculate L and H which bound the new alpha2
s = y1 * y2
if s == -1:
l, h = max(0.0, a2 - a1), min(self._c, self._c + a2 - a1) # noqa: E741
else:
l, h = max(0.0, a2 + a1 - self._c), min(self._c, a2 + a1) # noqa: E741
if l == h:
return None, None
# calculate eta
k11 = k(i1, i1)
k22 = k(i2, i2)
k12 = k(i1, i2)
# select the new alpha2 which could achieve the minimal objectives
if (eta := k11 + k22 - 2.0 * k12) > 0.0:
a2_new_unc = a2 + (y2 * (e1 - e2)) / eta
# a2_new has a boundary
if a2_new_unc >= h:
a2_new = h
elif a2_new_unc <= l:
a2_new = l
else:
a2_new = a2_new_unc
else:
b = self._b
l1 = a1 + s * (a2 - l)
h1 = a1 + s * (a2 - h)
# Method 1
f1 = y1 * (e1 + b) - a1 * k(i1, i1) - s * a2 * k(i1, i2)
f2 = y2 * (e2 + b) - a2 * k(i2, i2) - s * a1 * k(i1, i2)
ol = (
l1 * f1
+ l * f2
+ 1 / 2 * l1**2 * k(i1, i1)
+ 1 / 2 * l**2 * k(i2, i2)
+ s * l * l1 * k(i1, i2)
)
oh = (
h1 * f1
+ h * f2
+ 1 / 2 * h1**2 * k(i1, i1)
+ 1 / 2 * h**2 * k(i2, i2)
+ s * h * h1 * k(i1, i2)
)
"""
Method 2: Use objective function to check which alpha2_new could achieve the
minimal objectives
"""
if ol < (oh - self._eps):
a2_new = l
elif ol > oh + self._eps:
a2_new = h
else:
a2_new = a2
# a1_new has a boundary too
a1_new = a1 + s * (a2 - a2_new)
if a1_new < 0:
a2_new += s * a1_new
a1_new = 0
if a1_new > self._c:
a2_new += s * (a1_new - self._c)
a1_new = self._c
return a1_new, a2_new
# Normalize data using min-max method
def _norm(self, data):
if self._init:
self._min = np.min(data, axis=0)
self._max = np.max(data, axis=0)
self._init = False
return (data - self._min) / (self._max - self._min)
else:
return (data - self._min) / (self._max - self._min)
def _is_unbound(self, index):
return bool(0.0 < self.alphas[index] < self._c)
def _is_support(self, index):
return bool(self.alphas[index] > 0)
@property
def unbound(self):
return self._unbound
@property
def support(self):
return [i for i in range(self.length) if self._is_support(i)]
@property
def length(self):
return self.samples.shape[0]
class Kernel:
def __init__(self, kernel, degree=1.0, coef0=0.0, gamma=1.0):
self.degree = np.float64(degree)
self.coef0 = np.float64(coef0)
self.gamma = np.float64(gamma)
self._kernel_name = kernel
self._kernel = self._get_kernel(kernel_name=kernel)
self._check()
def _polynomial(self, v1, v2):
return (self.gamma * np.inner(v1, v2) + self.coef0) ** self.degree
def _linear(self, v1, v2):
return np.inner(v1, v2) + self.coef0
def _rbf(self, v1, v2):
return np.exp(-1 * (self.gamma * np.linalg.norm(v1 - v2) ** 2))
def _check(self):
if self._kernel == self._rbf and self.gamma < 0:
raise ValueError("gamma value must be non-negative")
def _get_kernel(self, kernel_name):
maps = {"linear": self._linear, "poly": self._polynomial, "rbf": self._rbf}
return maps[kernel_name]
def __call__(self, v1, v2):
return self._kernel(v1, v2)
def __repr__(self):
return self._kernel_name
def count_time(func):
def call_func(*args, **kwargs):
import time
start_time = time.time()
func(*args, **kwargs)
end_time = time.time()
print(f"SMO algorithm cost {end_time - start_time} seconds")
return call_func
@count_time
def test_cancer_data():
print("Hello!\nStart test SVM using the SMO algorithm!")
# 0: download dataset and load into pandas' dataframe
if not os.path.exists(r"cancer_data.csv"):
request = urllib.request.Request( # noqa: S310
CANCER_DATASET_URL,
headers={"User-Agent": "Mozilla/4.0 (compatible; MSIE 5.5; Windows NT)"},
)
response = urllib.request.urlopen(request) # noqa: S310
content = response.read().decode("utf-8")
with open(r"cancer_data.csv", "w") as f:
f.write(content)
data = pd.read_csv(
"cancer_data.csv",
header=None,
dtype={0: str}, # Assuming the first column contains string data
)
# 1: pre-processing data
del data[data.columns.tolist()[0]]
data = data.dropna(axis=0)
data = data.replace({"M": np.float64(1), "B": np.float64(-1)})
samples = np.array(data)[:, :]
# 2: dividing data into train_data data and test_data data
train_data, test_data = samples[:328, :], samples[328:, :]
test_tags, test_samples = test_data[:, 0], test_data[:, 1:]
# 3: choose kernel function, and set initial alphas to zero (optional)
my_kernel = Kernel(kernel="rbf", degree=5, coef0=1, gamma=0.5)
al = np.zeros(train_data.shape[0])
# 4: calculating best alphas using SMO algorithm and predict test_data samples
mysvm = SmoSVM(
train=train_data,
kernel_func=my_kernel,
alpha_list=al,
cost=0.4,
b=0.0,
tolerance=0.001,
)
mysvm.fit()
predict = mysvm.predict(test_samples)
# 5: check accuracy
score = 0
test_num = test_tags.shape[0]
for i in range(test_tags.shape[0]):
if test_tags[i] == predict[i]:
score += 1
print(f"\nAll: {test_num}\nCorrect: {score}\nIncorrect: {test_num - score}")
print(f"Rough Accuracy: {score / test_tags.shape[0]}")
def test_demonstration():
# change stdout
print("\nStarting plot, please wait!")
sys.stdout = open(os.devnull, "w")
ax1 = plt.subplot2grid((2, 2), (0, 0))
ax2 = plt.subplot2grid((2, 2), (0, 1))
ax3 = plt.subplot2grid((2, 2), (1, 0))
ax4 = plt.subplot2grid((2, 2), (1, 1))
ax1.set_title("Linear SVM, cost = 0.1")
test_linear_kernel(ax1, cost=0.1)
ax2.set_title("Linear SVM, cost = 500")
test_linear_kernel(ax2, cost=500)
ax3.set_title("RBF kernel SVM, cost = 0.1")
test_rbf_kernel(ax3, cost=0.1)
ax4.set_title("RBF kernel SVM, cost = 500")
test_rbf_kernel(ax4, cost=500)
sys.stdout = sys.__stdout__
print("Plot done!")
def test_linear_kernel(ax, cost):
train_x, train_y = make_blobs(
n_samples=500, centers=2, n_features=2, random_state=1
)
train_y[train_y == 0] = -1
scaler = StandardScaler()
train_x_scaled = scaler.fit_transform(train_x, train_y)
train_data = np.hstack((train_y.reshape(500, 1), train_x_scaled))
my_kernel = Kernel(kernel="linear", degree=5, coef0=1, gamma=0.5)
mysvm = SmoSVM(
train=train_data,
kernel_func=my_kernel,
cost=cost,
tolerance=0.001,
auto_norm=False,
)
mysvm.fit()
plot_partition_boundary(mysvm, train_data, ax=ax)
def test_rbf_kernel(ax, cost):
train_x, train_y = make_circles(
n_samples=500, noise=0.1, factor=0.1, random_state=1
)
train_y[train_y == 0] = -1
scaler = StandardScaler()
train_x_scaled = scaler.fit_transform(train_x, train_y)
train_data = np.hstack((train_y.reshape(500, 1), train_x_scaled))
my_kernel = Kernel(kernel="rbf", degree=5, coef0=1, gamma=0.5)
mysvm = SmoSVM(
train=train_data,
kernel_func=my_kernel,
cost=cost,
tolerance=0.001,
auto_norm=False,
)
mysvm.fit()
plot_partition_boundary(mysvm, train_data, ax=ax)
def plot_partition_boundary(
model, train_data, ax, resolution=100, colors=("b", "k", "r")
):
"""
We cannot get the optimal w of our kernel SVM model, which is different from a
linear SVM. For this reason, we generate randomly distributed points with high
density, and predicted values of these points are calculated using our trained
model. Then we could use this predicted values to draw contour map, and this contour
map represents the SVM's partition boundary.
"""
train_data_x = train_data[:, 1]
train_data_y = train_data[:, 2]
train_data_tags = train_data[:, 0]
xrange = np.linspace(train_data_x.min(), train_data_x.max(), resolution)
yrange = np.linspace(train_data_y.min(), train_data_y.max(), resolution)
test_samples = np.array([(x, y) for x in xrange for y in yrange]).reshape(
resolution * resolution, 2
)
test_tags = model.predict(test_samples, classify=False)
grid = test_tags.reshape((len(xrange), len(yrange)))
# Plot contour map which represents the partition boundary
ax.contour(
xrange,
yrange,
np.asmatrix(grid).T,
levels=(-1, 0, 1),
linestyles=("--", "-", "--"),
linewidths=(1, 1, 1),
colors=colors,
)
# Plot all train samples
ax.scatter(
train_data_x,
train_data_y,
c=train_data_tags,
cmap=plt.cm.Dark2,
lw=0,
alpha=0.5,
)
# Plot support vectors
support = model.support
ax.scatter(
train_data_x[support],
train_data_y[support],
c=train_data_tags[support],
cmap=plt.cm.Dark2,
)
if __name__ == "__main__":
test_cancer_data()
test_demonstration()
plt.show()