def fit(self, X, K, eps=pow(10, -2)):
# fits the parameters of the HMM using EM algorithm
# X is the sequence of observations (array of size (T,D)),
# K is the number of hidden states
# eps : tolerance on log likelihood difference between two iterations for convergence of EM algorithm
self.K = K
T, D = X.shape
# initialization of means and covariances with GMM
print(
"Initialization of Gaussians parameters (means and covariances) with GMM : "
)
gmm_model = GMM(isotropic=False)
gmm_model.fit(X, K, eps=eps)
self.mus = gmm_model.mus
self.Sigmas2 = gmm_model.Sigmas2
print("\nFit of HMM : ")
# initialization of pis and A at random
self.pis = np.random.rand(self.K)
self.pis /= np.sum(self.pis)
self.A = np.random.rand(self.K, self.K)
self.A /= np.sum(self.A, axis=1)[:, None]
lik = self.compute_log_likelihood(X)
print("Initial log-likelihood : ", lik)
delta_lik = 1
cpt_iter = 1
while (delta_lik > eps):
# Expectation step
pi = self.compute_proba_Zt_cond_X(
X) # array (T,K) (t,i) -> p(z_t = i|X; θ)
pij = self.compute_proba_Zt_and_Znext_cond_X(
X) # tensor (T-1,K,K) (t,i,j) -> p(z_(t+1) = j, z t = i|X; θ)
# Maximization step
self.pis = pi[0, :]
pi_repeated = pi[:, :, np.newaxis] # (T,K,D)
self.mus = np.sum(pi_repeated * X[:, np.newaxis, :],
axis=0) / np.sum(pi_repeated, axis=0)
self.Sigmas2 = []
for k in range(self.K):
Xc = X - self.mus[k]
Sigmas2k = 0
for t in range(T):
xt = Xc[t, :][:, None] # size (d,1)
Sigmas2k += np.dot(xt, xt.T) * pi[t, k]
Sigmas2k /= np.sum(pi[:, k])
self.Sigmas2.append(Sigmas2k)
self.Sigmas2 = np.array(self.Sigmas2)
self.A = np.sum(pij, axis=0) / np.sum(pi[:-1], axis=0)[:, None]
# Computing new likelihood, and deciding if we should stop
old_lik = lik # storing old_likelihood to compute delta_lik
lik = self.compute_log_likelihood(X) # storing new likelihood
delta_lik = lik - old_lik # measure to decide if we should stop or iterate again
print("Iter " + str(cpt_iter) + " ; log_likelihood : " + str(lik))
cpt_iter += 1
print("EM algorithm converged.")
print("initial distribution found (rounded, 2 decimals) : ",
np.round(self.pis, 2))
print("transition matrix found (rounded, 2 decimals) : ",
np.round(self.A, 2))
from GMM import GMM
import numpy as np
from sklearn.datasets import make_moons
from sklearn.model_selection import train_test_split
import matplotlib.pyplot as plt
from util import *
# 构造聚类数据,X是特征数据，Y是相应的label，此时生成的是半环形图
X, Y = make_moons(n_samples=1000, noise=0.04, random_state=0)
# 划分数据，一部分用于训练聚类，一部分用于分类
X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.2)
model = GMM(X_train, K=10)
# 获取各个类别的概率
result = model.fit()
print('每条数据属于各个类别的概率如下: ', result)
# 获取每条数据所在的类别
label_train = np.argmax(result, axis=1)
print(label_train)
# 获取测试数据所在的类别的概率
result_test = model.predict(X_test)
# 获取测试数据的类别
label_test = np.argmax(result_test, axis=1)
# 展示原始数据分布及其label
ax1 = plt.subplot(211)
ax1.scatter(X[:, 0],
X[:, 1],
#!/usr/bin/env python3 # -*- coding: utf-8 -*- """ Created on Sun Apr 21 02:43:24 2019 @author: maachou """ from sklearn.datasets.samples_generator import make_blobs import matplotlib.pyplot as plt from GMM import GMM mix=GMM(K=6) X,Y = make_blobs(cluster_std=0.5,random_state=20,n_samples=100,centers=6) plt.scatter(X[:,0],X[:,1]) print(X.shape) mix.fit(X) mix.Means() Y=mix.predict(X) plt.scatter(X[:,0],X[:,1],c=Y)
import numpy as np
import matplotlib.pyplot as plt
from GMM import GMM
if __name__ == '__main__':
group_a = np.random.normal(loc=(20.00, 14.00), scale=(4.0, 4.0), size=(1000, 2))
group_b = np.random.normal(loc=(15.00, 8.00), scale=(2.0, 2.0), size=(1000, 2))
group_c = np.random.normal(loc=(30.00, 40.00), scale=(2.0, 2.0), size=(1000, 2))
group_d = np.random.normal(loc=(25.00, 32.00), scale=(7.0, 7.0), size=(1000, 2))
group_e = np.random.normal(loc=(10.00, 32.00), scale=(7.0, 7.0), size=(1000, 2))
DATA = np.concatenate((group_a, group_b, group_c, group_d, group_e))
S = GMM(5, DATA, 1e-3)
S.fit()
S.print_status()
testdata = np.random.rand(10000, 2)*50
labels = S.Classify(testdata)
plt.scatter(testdata[:, 0], testdata[:, 1], c=list(map(lambda i : {0:'b',1:'g',2:'r',3:'y',4:'k'}[i], labels)))
plt.show()
# generate the dataset
X, Y = make_classification(n_samples=1000,
n_features=2,
n_redundant=0,
n_informative=2,
n_clusters_per_class=2)
X = preprocessing.scale(X)
num_clusters = 3
num_epochs = 50
gmm_model = GMM()
phi, pi_dist, mean, covariance = gmm_model.fit(X,
num_clusters=num_clusters,
num_epochs=num_epochs)
gmm_sklearn = mixture.GaussianMixture(n_components=2)
gmm_sklearn.fit(X)
plt.figure(figsize=(8, 8))
plt.subplots_adjust(left=0.05, bottom=0.05, right=0.95, top=0.9)
plt.subplot(211)
plt.title('Plot for the unclustered data', fontsize='small')
plt.scatter(X[:, 0], X[:, 1], s=25, c=None)
plt.subplot(212)
plt.title('Plot for the clustered data', fontsize='small')
plt.scatter(X[:, 0], X[:, 1], s=25, c=phi)
class QuickBrush(Brush):
lWorksize = (16, 16)
def __init__(self, context, devices, d_img, d_labels):
Brush.__init__(self, context, devices, d_labels)
self.context = context
self.queue = cl.CommandQueue(context,
properties=cl.command_queue_properties.PROFILING_ENABLE)
nComponentsFg = 4
nComponentsBg = 4
self.nDim = 3
self.dim = d_img.dim
filename = os.path.join(os.path.dirname(__file__), 'quick.cl')
program = createProgram(context, context.devices, [], filename)
# self.kernSampleBg = cl.Kernel(program, 'sampleBg')
self.kern_get_samples = cl.Kernel(program, 'get_samples')
self.lWorksize = (16, 16)
self.gWorksize = roundUp(self.dim, self.lWorksize)
nSamples = 4 * (self.gWorksize[0] / self.lWorksize[0]) * (
self.gWorksize[1] / self.lWorksize[1])
# self.gmmFg_cpu = mixture.GMM(4)
self.gmmFg = GMM(context, 65, nComponentsFg, 10240)
self.gmmBg = GMM(context, 65, nComponentsBg, nSamples)
self.hScore = np.empty(self.dim, np.float32)
self.hSampleFg = np.empty((10240, ), np.uint32)
self.hSampleBg = np.empty((12000, ), np.uint32)
self.hA = np.empty((max(nComponentsFg, nComponentsBg), 8), np.float32)
self.d_img = d_img
cm = cl.mem_flags
self.dSampleFg = cl.Buffer(context, cm.READ_WRITE, size=4 * 10240)
self.dSampleBg = cl.Buffer(context, cm.READ_WRITE, size=4 * 12000)
self.dA = cl.Buffer(context, cm.READ_ONLY | cm.COPY_HOST_PTR, hostbuf=self.hA)
self.dScoreFg = Buffer2D(context, cm.READ_WRITE, self.dim, np.float32)
self.dScoreBg = Buffer2D(context, cm.READ_WRITE, self.dim, np.float32)
#self.points = Set()
self.capPoints = 200 * 200 * 300 #brush radius 200, stroke length 300
self.points = np.empty((self.capPoints), np.uint32)
# self.colorize = Colorize.Colorize(clContext, clContext.devices)
# self.hTriFlat = self.hTri.reshape(-1)
# self.probBg(1200)
self.h_img = np.empty(self.dim, np.uint32)
self.h_img = self.h_img.ravel()
cl.enqueue_copy(self.queue, self.h_img, self.d_img, origin=(0, 0), region=self.dim).wait()
self.samples_bg_idx = np.random.randint(0, self.dim[0] * self.dim[1], 12000)
self.hSampleBg = self.h_img[self.samples_bg_idx]
cl.enqueue_copy(self.queue, self.dSampleBg, self.hSampleBg).wait()
w,m,c = self.gmmBg.fit(self.dSampleBg, 300, retParams=True)
print w
print m
print c
self.gmmBg.score(self.d_img, self.dScoreBg)
pass
def draw(self, p0, p1):
Brush.draw(self, p0, p1)
#self.probFg(x1-20, x1+20, y1-20, y1+20)
#return
"""color = self.colorTri[self.type]
#self.argsScore[5] = np.int32(self.nComponentsFg)
#seed = []
hasSeeds = False
redoBg = False
minX = sys.maxint
maxX = -sys.maxint
minY = sys.maxint
maxY = -sys.maxint
for point in self.points[0:nPoints]:
#if self.hTriFlat[point] != color:
self.hTriFlat[point] = color
#seed += point
hasSeeds = True
minX = min(minX, point%self.width)
maxX = max(maxX, point%self.width)
minY = min(minY, point/self.width)
maxY = max(maxY, point/self.width)
#if (point[1]*self.width + point[0]) in self.randIdx:
# redoBg = True
#if redoBg:
# self.probBg(0)
#if len(seed) == 0:
if not hasSeeds:
return
minX = max(0, minX-DILATE)
maxX = min(self.width-1, maxX + DILATE)
minY = max(0, minY-DILATE)
maxY = min(self.height-1, maxY + DILATE)
"""
args = [
np.int32(self.n_points),
self.d_points,
cl.Sampler(self.context, False, cl.addressing_mode.NONE,
cl.filter_mode.NEAREST),
self.d_img,
self.dSampleFg
]
gWorksize = roundUp((self.n_points, ), (256, ))
self.kern_get_samples(self.queue, gWorksize, (256,), *args).wait()
cl.enqueue_copy(self.queue, self.hSampleFg, self.dSampleFg)
# print self.hSampleFg.view(np.uint8).reshape(10240, 4)[0:self.n_points, :]
# print self.n_points
self.gmmFg.fit(self.dSampleFg, self.n_points)
# print w
# print m
# print c
self.gmmFg.score(self.d_img, self.dScoreFg)
# self.argsSampleBg = [
# self.d_labels,
# np.int32(self.label),
# cl.Sampler(self.context, False, cl.addressing_mode.NONE,
# cl.filter_mode.NEAREST),
# self.d_img,
# self.dSampleFg
# ]
#
# gWorksize = roundUp(self.dim, (16, 16))
#
# self.kernSampleBg(self.queue, gWorksize, (16, 16),
# *(self.argsSampleBg)).wait()
# cl.enqueue_copy(self.queue, self.hSampleBg, self.dSampleBg).wait()
pass
def probFg(self, d_samples, n_points):
# if True:
# tri = self.hTri[minY:maxY, minX:maxX]
# b = (tri == self.colorTri[self.type])
#
# samplesFg = self.hSrc[minY:maxY, minX:maxX]
# samplesFg = samplesFg[b]
# else:
# DILATE = 5
# samplesFg = self.hSrc[minY:maxY, minX:maxX].ravel()
#gpu = False
#self.prob(self.gmmFG, samplesFg, self.dScoreFg, gpu)
#self.gmmFg_cpu.fit(samplesFg)
#print 'cpu', self.gmmFg_cpu.weights_
#a = calcA_cpu(self.gmmFg_cpu.weights_.astype(np.float32), self.gmmFg_cpu.means_.astype(np.float32), self.gmmFg_cpu.covars_.astype(np.float32))
#cl.enqueue_copy(self.queue, self.gmmFg.dA, a).wait()
#weights, means, covars = self.gmmFg.fit(samplesFg, retParams=True)
#a = calcA_cpu(weights, means[:, 0:3], covars[:, 0:3])
#cl.enqueue_copy(self.queue, self.gmmFg.dA, a).wait()
w,m,c = self.gmmFg.fit(d_samples, n_points, retParams=True)
print w
print m
print c
#print 'gpu', weights
self.gmmFg.score(self.d_img, self.dScoreFg)
#score returns float64, not float32 -> convert with astype
#self.hScore = -self.gmmFG.score(self.rgb.reshape(-1, 3)).astype(np.float32)
"""
def drawCircle(self, xc, yc, points=None):
r = self.radius
for y in xrange(-r, r):
for x in xrange(-r, r):
if points != None:
points.add((xc+x, yc+y))
"""
def probBg(self, nSamples):
#self.kernSampleBg(self.queue, self.gWorksize, self.lWorksize, *(self.argsSampleBg)).wait()
#cl.enqueue_copy(self.queue, self.hSampleBg, self.dSampleBg).wait()
self.bgIdx = np.where(self.hTri.ravel() != self.colorTri[self.type])[0]
self.randIdx = self.bgIdx[np.random.randint(0, len(self.bgIdx), 2000)]
self.bgIdx = np.setdiff1d(self.bgIdx, self.randIdx)
self.hSampleBg[0:len(self.randIdx)] = self.hSrc.view(np.uint32).ravel()[
self.randIdx]
cl.enqueue_copy(self.queue, self.dSampleBg, self.hSampleBg).wait()
#print self.gmmBg.fit(self.hSrc.view(np.uint32).ravel()[self.randIdx], retParams=True)
self.gmmBg.fit(self.hSrc.view(np.uint32).ravel()[self.randIdx])
#self.gmmBg.fit(self.dSampleBg, nSamples=len(self.randIdx))
self.gmmBg.score(self.dSrc, self.dScoreBg)
gmm_cpu = mixture.GMM(nComp)
gmm_cpu.dtype = np.float32
gmm_cpu.init_params = ''
gmm_cpu.means_ = means
gmm_cpu.weights_ = weights
gmm_cpu.covars_ = covars
gmm_cpu.fit(samples)
gmm = GMM(context, nIter, nComp, nSamples)
a = calcA_cpu(weights, means, covars)
cl.enqueue_copy(queue, gmm.dA, a).wait()
gmm.has_preset_wmc = True
w,m,c = gmm.fit(dSamples, nSamples, retParams=True)
print 'converged: {0}'.format(gmm.has_converged)
print gmm_cpu.weights_
print w
print
print gmm_cpu.means_
print m
print
print gmm_cpu.covars_
print c
gmm_cpu.init_params = 'wmc'
iter = 10
#to estimate wmc on cpu
from GMM import GMM
from sklearn.datasets import make_blobs
import matplotlib.pyplot as plt
X, y = make_blobs(n_samples=1000, centers=4, n_features=2)
gmm_cls = GMM(initializer='uniform', cov_type='diag')
gmm_cls.fit(X, 4)
colors = []
for l in gmm_cls.kmeans_cls_.predict(X):
if l == 0:
colors.append('red')
if l == 1:
colors.append('green')
if l == 2:
colors.append('orange')
if l == 3:
colors.append('blue')
plt.scatter(X[:, 0], X[:, 1], c=colors, alpha=0.1)
plt.scatter(gmm_cls.means_[:, 0], gmm_cls.means_[:, 1], c='k')
plt.show()
plt.scatter(X[:, 0], X[:, 1], c=colors, alpha=0.1)
plt.scatter(gmm_cls.kmeans_cls_.means_[:, 0],
gmm_cls.kmeans_cls_.means_[:, 1],
c='k')
plt.show()
kmeans_obj = KMeans(3, x) kmeans_obj.fit(3, 0.002) means = kmeans_obj.mean_vec cov_mat_list = kmeans_obj.CovMatrix() mixture_coeff = kmeans_obj.MixtureCoeff() print(cov_mat_list) """from sklearn.cluster import KMeans obj = KMeans(n_clusters = 3, init = 'k-means++', max_iter = 100, n_init = 10, random_state = 0) y_Kmeans = obj.fit_predict(x) print(obj.cluster_centers_[:])""" GMM_obj = GMM(3, x, means, cov_mat_list, mixture_coeff) GMM_obj.fit(0.0002) print(GMM_obj.mean_vec) print(GMM_obj.cov_mat) print(GMM_obj.mixture_coeff) y_pred = GMM_obj.ClusterPredict(x) plt.scatter(GMM_obj.x_train[y_pred == 0, 0], GMM_obj.x_train[y_pred == 0, 1], s = 20, c = 'red', label = 'Cluster 1') plt.scatter(GMM_obj.x_train[y_pred == 1, 0], GMM_obj.x_train[y_pred == 1, 1], s = 20, c = 'green', label = 'Cluster 2') plt.scatter(GMM_obj.x_train[y_pred == 2, 0], GMM_obj.x_train[y_pred == 2, 1], s = 20, c = 'blue', label = 'Cluster 3') plt.scatter(GMM_obj.mean_vec[:, 0], GMM_obj.mean_vec[:, 1], s = 50, c = 'yellow', label = 'Centroids') plt.show() plt.scatter(GMM_obj.x_train[:, 0], GMM_obj.x_train[:, 1]) plt.show()
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns; sns.set()
from sklearn.datasets.samples_generator import make_blobs
from sklearn.model_selection import train_test_split
from util import *
# 构造聚类数据,X是特征数据，Y是相应的label，此时生成的是半环形图
X, Y = make_blobs(n_samples=700, centers=4,cluster_std=0.5, random_state=2019)
# 划分数据，一部分用于训练聚类，一部分用于分类
X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.2)
model = GMM(X_train,K=4)
# 获取训练数据各个类别的概率
result_train = model.fit()
print('每条数据属于各个类别的概率如下: ',result_train)
# 获取训练数据所在的类别
label_train = np.argmax(result_train,axis=1)
print(label_train)
# 获取测试数据所在的类别的概率
result_test = model.predict(X_test)
# 获取测试数据的类别
label_test = np.argmax(result_test,axis=1)
# 展示原始数据分布及其label
ax1 = plt.subplot(211)
ax1.scatter(X[:,0],X[:,1],s=50,c=Y,marker='x',cmap='viridis',label="Original")
ax1.set_title('Original Data and label Distribution')
