def recover(data):
mu = np.mean(data)
sigma = np.var(data)
init_params = [(mu + 0.1, sigma), (mu - 0.1, sigma)]
weight, distributions, ll = mixem.em(
data, [NormalDistribution(mu, sigma) for mu_sigma in init_params])
print(weight, distributions, ll)
def train_lambda(data):
weights, distributions, ll = mixem.em(np.array(data), [
mixem.distribution.NormalDistribution(0, 1),
mixem.distribution.NormalDistribution(3, 4),
mixem.distribution.NormalDistribution(7, 8)
])
return weights
def recover(data):
init_params = np.random.choice(data, size=2)
weight, distributions, ll = mixem.em(data, [ExponentialDistribution(l) for l in init_params])
print(weight, distributions, ll)
def fitDataBMM(data, depth, lowerLimit, upperLimit, init_proportions, components=2):
""" Fit data and return Binomial Mixture Model"""
if components > 9:
logging.error('Too many components specified. Max components 9.')
exit()
distros = []
# Init distros
for i in range(components):
distros.append(BinomialDistribution(init_proportions[i], depth))
# Format data as pairs of success and trials
data_points = []
for x,y in zip(data[:,:-2].flatten(), np.repeat(data[:,-1],4)):
# do filtering on each proportion
if x > lowerLimit and x < upperLimit:
charCount = x*0.01*y # convert proportion to count
data_points.append([charCount,y])
else:
continue
data = np.array(data_points)
weights, distros, log_like = mixem.em(data, distros, initial_weights=None, progress_callback=None, max_iterations=500, tol_iters=200, tol=0.1)
return BinomialMixture(weights, distros, log_like)
def recover(data):
weight, distributions, ll = mixem.em(data, [
GeometricDistribution(0.8),
GeometricDistribution(0.1),
])
print(weight, distributions, ll)
def recover(data):
init_params = np.random.choice(data, size=2)
weight, distributions, ll = mixem.em(
data, [ExponentialDistribution(l) for l in init_params])
print(weight, distributions, ll)
def train_lambda(data):
# plt.scatter(np.array(range(data.shape[0])), data)
# plt.show();
weights, distributions, ll = mixem.em(np.sort(np.array(data)), [
mixem.distribution.NormalDistribution(0, 1),
mixem.distribution.NormalDistribution(0.3, 5),
mixem.distribution.NormalDistribution(1, 9)
])
return weights
def recover(data):
mu = np.mean(data)
sigma = np.var(data)
init_params = [(np.array([mu + 0.1]), np.diag([sigma])), (np.array([mu - 0.1]), np.diag([sigma]))]
weight, distributions, ll = mixem.em(data, [MultivariateNormalDistribution(mu, sigma) for mu, sigma in init_params])
print(weight, distributions, ll)
def train_lambda(data):
'''
train_lambda takes training data and returns the lambdas which will used
in the EM algorithm.
We use a implementation of the Expectation-Maximization (EM) algorithm
called "mixem" to tune the parameters lambda in the interpolation.
https://pypi.python.org/pypi/mixem
'''
weights, distributions, ll = mixem.em(np.sort(np.array(data)), [mixem.distribution.NormalDistribution(0,1),mixem.distribution.NormalDistribution(0.3,5),mixem.distribution.NormalDistribution(1,9)])
return weights
def recover(data):
mu = np.mean(data)
sigma = np.var(data)
init_params = [
(np.array([mu + 0.1]), np.diag([sigma])),
(np.array([mu - 0.1]), np.diag([sigma]))
]
# start = time.time()
weight, distributions, ll = mixem.em(data, [MultivariateNormalDistribution(mu, sigma) for mu, sigma in init_params])
print(weight, distributions, ll)
def recover(data):
mu = np.mean(data)
sigma = np.var(data)
init_params = [
(mu + 0.1, sigma),
(mu - 0.1, sigma)
]
weight, distributions, ll = mixem.em(data, [NormalDistribution(mu, sigma) for mu_sigma in init_params])
print(weight, distributions, ll)
def main():
data = pd.read_csv(os.path.join(os.path.dirname(__file__), "faithful.csv"))
data = np.array(data)
init_params = [
(np.array((2, 50)), np.identity(2)),
(np.array((4, 80)), np.identity(2)),
]
weight, distributions, ll = mixem.em(data, [MultivariateNormalDistribution(mu, sigma) for mu, sigma in init_params], initial_weights=[0.3, 0.7])
print(weight, distributions, ll)
def main():
startTime = time.time()
folderName = "assignment-2-data-files/"
fileName = "P1M1L1.txt"
data, lengthData = readTrainTxt(folderName + fileName)
data = takeBatches(data, 1000)
init_params = [(0, 2), (4, 2), (8, 2), (12, 2)]
weight, distributions, ll = mixem.em(
data, [NormalDistribution(mu, sigma) for (mu, sigma) in init_params])
print(weight, distributions, ll)
print(time.time() - startTime)
def recover(data):
mu = [np.mean(data[0, :]), np.mean(data[1, :])]
sigma = [np.var(data[0, :]), np.var(data[1, :])]
#print mu,sigma
init_params = [
(np.array((mu[0] - 1, mu[0] + 1)), np.identity(2)),
(np.array((mu[1] - 1, mu[1] + 1)), np.identity(2)),
]
start = time.time()
weight, distributions, ll, iteration = mixem.em(data, [
MultivariateNormalDistribution(mu, sigma) for mu, sigma in init_params
])
#print(weight, distributions, ll)
#print 'iterate time: ' + str(time.time() - start) + ' seconds'
return weight, distributions, iteration, (time.time() - start)
def get_normals_mixture(x, data):
dist_list_norm = [
mixem.distribution.NormalDistribution(mu=0.5, sigma=1),
mixem.distribution.NormalDistribution(mu=2, sigma=1)
]
norm_mixture = Data_Rep(data, dist_list_norm)
#removed ref to train set
norm_weights, norm_dists, norm_log_l = mixem.em(data,
dist_list_norm,
max_iterations=200,
progress_callback=None)
post_scipy_dist_1_norm, post_scipy_dist_2_norm = norm_mixture.scipy_dists
norm_pdf1 = [post_scipy_dist_1_norm.pdf(i) for i in x]
norm_pdf2 = [post_scipy_dist_2_norm.pdf(i) for i in x]
norm_joint_pdf = [
norm_weights[0] * norm_pdf1[i] + norm_weights[1] * norm_pdf2[i]
for i in range(len(norm_pdf1))
]
return norm_mixture, norm_joint_pdf
def organize_data(x, data1, data2, dist1, dist2):
data = np.concatenate((data1, data2))
np.random.shuffle(data)
# train_set = data[1000:]
# val_set = data[:1000]
# sorted_val = np.array(sorted(val_set))
dist_list = [dist1, dist2]
mixture = Data_Rep(data, dist_list)
post_scipy_dist_1, post_scipy_dist_2 = mixture.scipy_dists
# took out train/val split here
weights, distributions, log_l = mixem.em(data,
dist_list,
max_iterations=200,
progress_callback=None)
pdf1 = [post_scipy_dist_1.pdf(i) for i in x]
pdf2 = [post_scipy_dist_2.pdf(i) for i in x]
joint_pdf = [
weights[0] * pdf1[i] + weights[1] * pdf2[i] for i in range(len(pdf1))
]
return mixture, joint_pdf, data
def _mixture_cdf(self, data, dist_list): self.weights, self.distributions, self.log_l = mixem.em(data, dist_list, max_iterations=200, progress_callback=None) self.scipy_dists = self.get_scipy_dists(self.distributions) return lambda query: sum([w * dist.cdf(query) for w, dist in zip(self.weights, self.scipy_dists)])
# prepare some data
data = pd.read_csv("faithful.csv")
#print(data.head())
# output to static HTML file
output_file("Fitting_Data_Contour.html", title="Old Faithful Data")
# create a new plot with a title and axis labels
'''
fig = figure(title="Old Faithful Data", x_axis_label="Eruption duration (minutes)", y_axis_label="Waiting time (minutes)")
fig.scatter(x=data.eruptions, y=data.waiting)
show(fig);
'''
weights, distributions, ll, iteration = mixem.em(np.array(data), [
mixem.distribution.MultivariateNormalDistribution(np.array(
(2, 50)), np.identity(2)),
mixem.distribution.MultivariateNormalDistribution(np.array(
(4, 80)), np.identity(2)),
])
N = 100
x = np.linspace(np.min(data.eruptions), np.max(data.eruptions), num=N)
y = np.linspace(np.min(data.waiting), np.max(data.waiting), num=N)
xx, yy = np.meshgrid(x, y, indexing="ij")
#print x,y
# Convert meshgrid into a ((N*N), 2) array of coordinates
xxyy = np.array([xx.flatten(), yy.flatten()]).T
# Compute model probabilities
p = mixem.probability(xxyy, weights, distributions).reshape((N, N))
#print p
fig2 = figure(title="Fitted Old Faithful Data",