-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathbmmSlow.py
180 lines (147 loc) · 7.86 KB
/
bmmSlow.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
import argparse
import math
import os
import pickle
import numpy as np
import torch
import torchvision
from torchvision.datasets import MNIST
from torchvision import transforms
from torch.utils.data import Dataset, DataLoader
from tqdm import tqdm
import matplotlib
import matplotlib.pyplot as plt
def mnist_loader(data_type,batch_size):
"""Setup MNIST data loader."""
# Load datasets.
if data_type == 'float':
train_dataset = MNIST('data/pytorch_data/MNIST', train=True, download=True, transform=transforms.Compose([transforms.ToTensor()]))
test_dataset = MNIST('data/pytorch_data/MNIST', train=False, download=True, transform=transforms.Compose([transforms.ToTensor()]))
elif data_type == 'binary':
train_dataset = MNIST('data/pytorch_data/MNIST', train=True, download=True, transform=transforms.Compose([transforms.ToTensor(), lambda x:x>0, lambda x: x.float()]))
test_dataset = MNIST('data/pytorch_data/MNIST', train=False, download=True, transform=transforms.Compose([transforms.ToTensor(), lambda x:x>0, lambda x: x.float()]))
kwargs = {'num_workers': 8, 'pin_memory': True, 'batch_size': batch_size, 'shuffle': False}
train_loader = DataLoader(train_dataset, **kwargs)
test_loader = DataLoader(test_dataset, **kwargs)
return train_loader, test_loader, len(train_dataset)
def Simplex(K):
X = []
X.append(0)
for i in range(K-1):
X.append(np.random.uniform(0,1))
X.append(1)
X.sort()
Z = []
for i in range(K):
Z.append(X[i+1] - X[i])
return Z
def ComputePosterior(data_i, component_k, pi, theta, K_mixture, J_parameter_dimension, device):
current_posterior_gamma_ik = 0
# Computing numerator
numerator = 1
for d in range(J_parameter_dimension):
numerator = numerator * ( (theta[component_k][d]**data_i[d]) * ( (1 - theta[component_k][d])**(1 - data_i[d]) ) )
# Computing denominator
denominator = 0
for k in range(K_mixture):
temp = 1
for d in range(J_parameter_dimension):
temp = temp * ( (theta[k][d]**data_i[d]) * ( (1 - theta[k][d])**(1 - data_i[d]) ) )
denominator = denominator + (pi[k] * temp)
current_posterior_gamma_ik = (pi[component_k] * numerator) / denominator
return current_posterior_gamma_ik
def ComputeMarginal(K_mixture, J_parameter_dimension, train_loader, pi, theta, device, batch_size):
marginal = 0
for i,data in enumerate(train_loader):
#data_i = data[0].to(device)
data_i = data[0]
data_i = torch.squeeze(data_i)
data_i = data_i.view(-1)
sum_k = 0
for k in range(K_mixture):
temp = 1
for d in range(J_parameter_dimension):
temp = temp * ( (theta[k][d]**data_i[d]) * ( (1 - theta[k][d])**(1 - data_i[d]) ) )
sum_k = sum_k + (pi[k] * temp)
marginal = marginal + math.log(sum_k)
return marginal
def train(data_type, epoch_num, batch_size, K_mixture, J_parameter_dimension, device, results):
"""pi is a vector of length K, theta is of shape K*J, and in my case K is 10 and J=D=784"""
train_loader, test_loader, N_number_data = mnist_loader(data_type, batch_size)
# initializing pi in simplex(K-1), and theta of shape K*J
pi = Simplex(K_mixture)
theta = np.random.uniform(0,1,(K_mixture, J_parameter_dimension))
alpha = np.random.uniform(0,1,(K_mixture))
marginal_log_like = []
epoch_list = []
for epoch in tqdm(range(epoch_num)):
theta_numerator = np.zeros((K_mixture, J_parameter_dimension))
theta_denominator = np.zeros((K_mixture, J_parameter_dimension))
pi_numerator = np.zeros((K_mixture))
for i,data in enumerate(train_loader):
# data[0] is the batch_size*1*28*28 matrix and data[1] is the label
# removing dimensions of size 1
#data_i = data[0].to(device)
data_i = data[0]
data_i = torch.squeeze(data_i)
# convert the shape of tensor from 28*28 to 784
data_i = data_i.view(-1)
# data_i[d] represents x_{i,d}
for k in range(K_mixture):
# we pass theta which is parameters to compute current posterior
posterior_gamma_ik = ComputePosterior(data_i, k, pi, theta, K_mixture, J_parameter_dimension, device)
pi_numerator[k] = pi_numerator[k] + posterior_gamma_ik
for d in range(J_parameter_dimension):
# is it okay if I use posterior_gamma_ik that is computed outside this for?
#posterior_gamma_ik = ComputePosterior(data_i, k, pi, theta)
theta_numerator[k][d] = theta_numerator[k][d] + (posterior_gamma_ik * data_i[d])
theta_denominator[k][d] = theta_denominator[k][d] + posterior_gamma_ik
print('epoch:', epoch, 'data processed so far:', (i+1)*batch_size)
# Now that we have gone through all the data, we can update parameters:
theta = theta_numerator / theta_denominator # Shall I have a loop or it works in python
pi = (pi_numerator + alpha - 1) / ( sum(alpha) + N_number_data - K_mixture) # Shall I have a loop or it works in python
epoch_list.append(epoch)
marginal = ComputeMarginal(K_mixture, J_parameter_dimension, train_loader, pi, theta, device, batch_size)
print('epoch:', epoch, 'marginal log likelihood:', marginal )
marginal_log_like.append(marginal)
pickle.dump( ([epoch_list, marginal_log_like]), open( results+'epoch_marginal.pkl', "wb" ) )
# plot marginal log likelihood for each epoch
plt.plot(epoch_list,marginal_log_like)
plt.title('Marginal Log Likelihood')
plt.xlabel('epoch')
plt.ylabel('marginal log likelihood')
plt.savefig('Marginal.pdf')
#saving the marginal after training is done to compare to other experiments
pickle.dump( (marginal_log_like[-1]), open( results+'finalmarginal.pkl', "wb" ) )
#saving the parameters of the model to able to replicate the results
pickle.dump( (pi), open( results+'pi.pkl', "wb" ) )
pickle.dump( (theta), open( results+'theta.pkl', "wb" ) )
pickle.dump( (alpha), open( results+'alpha.pkl', "wb" ) )
# Note that we have K components each of which are J dimensional where J=D and D is the number of pixels. So we can show images (instances) based on each K component
# saving some instances per component.
all_sampled_data = []
for k in range(K_mixture):
for instance in range(3):
sampled_data = np.zeros(J_parameter_dimension)
for d in range(J_parameter_dimension):
sampled_data[d] = np.random.binomial(size=1, n=1, p= theta[k][d])
sampled_data = sampled_data.reshape(28,28)
matplotlib.image.imsave(results+'component'+str(k)+'instance'+str(instance)+'.png', sampled_data)
all_sampled_data.append(sampled_data)
pickle.dump( (all_sampled_data), open( results+'sampledImages.pkl', "wb" ) )
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--experiment_name', default='Random', type=str)
parser.add_argument('--data_type', default='binary', type=str)
parser.add_argument('--epoch_num', default=10, type=int)
parser.add_argument('--batch_size', default=1, type=int)
parser.add_argument('--K_mixture', default=10, type=int)
parser.add_argument('--J_parameter_dimension', default=784, type=int)
args = parser.parse_args()
device_name = 'cuda:'+str(0) if torch.cuda.is_available() else 'cpu'
device = torch.device(device_name)
results = './results/' + args.experiment_name + '/'
if not os.path.exists(results):
os.makedirs(results)
print('device in use:', device)
train(args.data_type,args.epoch_num,args.batch_size,args.K_mixture,args.J_parameter_dimension, device, results)