-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathsemi_gmm.py
141 lines (110 loc) · 4.95 KB
/
semi_gmm.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
import numpy as np
from scipy.stats import multivariate_normal as mvn
import matplotlib.pyplot as plt
import data
import util
import pdb
inputList = "data/data.txt"
gtList = "data/gt.txt"
clusterInName = "data/kmeans_cluster.npy"
np_cluster_means = np.load(clusterInName)
#y by x
patchSize = (5, 5)
numIterations = 50
(trainData, trainGt, testData, testGt) = data.getImages(inputList, gtList)
(numTrain, ny, nx, drop) = trainGt.shape
trainData = np.concatenate((trainData, testData), axis=0)
(numTest, ny, nx, drop) = testGt.shape
numTotal = numTrain + numTest
#Set test data as unsupervised
unsupGtData = np.zeros((numTest, ny, nx, 5))
unsupGtData[:, :, :, 4] = 1
trainGt = np.concatenate((trainGt, unsupGtData), axis=0)
#Get data patches
xData = util.tfExtractPatches(trainData, patchSize)
gtData = np.reshape(trainGt, [-1, 5])
#Initialize clusters
#clusterMeans is [numClusters, numFeatures]
clusterMeans = np_cluster_means.astype(np.float64)
#clusterStds is [numClusters, numFeatures, numFeatures]
#clusterStds = np.tile(np.eye(25), [4, 1, 1]).astype(np.float64)
clusterStds = np.zeros((4, 25, 25))
for k in range(4):
clusterStds[k] = np.diag(np.abs(-(clusterMeans[k]*clusterMeans[k].transpose())))
#clusterPrior is [numClusters]
clusterPrior = np.array([.25, .25, .25, .25]).astype(np.float64)
#Grab supervised data
posIdxs = np.nonzero(gtData[:, 4] == 0)
sup_gtData = gtData[posIdxs][:, :-1] #Drop distractor class
sup_xData = xData[posIdxs]
(numSup, drop) = sup_xData.shape
#Grab unsupervised data
negIdxs = np.nonzero(gtData[:, 4] == 1)
unsup_xData = xData[negIdxs]
(numUnsup, drop) = unsup_xData.shape
#Run EM
for iteration in range(numIterations):
#E step for all data
#mvn_data is [numData, numClusters]
unsup_mvn_data = np.transpose(np.array([mvn.pdf(unsup_xData, clusterMeans[k], clusterStds[k]) for k in range(4)]))
unsup_respNum = clusterPrior[np.newaxis, :] * unsup_mvn_data
unsup_respDen = np.sum(unsup_respNum, axis=1)
#resp is [numData, numClusters]
unsup_resp = unsup_respNum/unsup_respDen[:, np.newaxis]
#respNorm is [numClusters]
respNorm = np.sum(unsup_resp, axis=0)
gtNorm = np.sum(sup_gtData, axis=0)
norm = gtNorm + respNorm
#M step on prior
new_clusterPrior = norm/(numSup + numUnsup)
sup_clusterMeans = np.sum(sup_gtData[:, :, np.newaxis] * sup_xData[:, np.newaxis, :], axis=0)
unsup_clusterMeans = np.sum(unsup_resp[:, :, np.newaxis] * unsup_xData[:, np.newaxis, :], axis=0)
new_clusterMeans = (sup_clusterMeans + unsup_clusterMeans)/norm[:, np.newaxis]
sup_cData = sup_xData[:, np.newaxis, :] - new_clusterMeans[np.newaxis, :, :]
sup_cData_cDataT = sup_cData[:, :, :, np.newaxis] * sup_cData[:, :, np.newaxis, :]
sup_clusterStds = np.sum(sup_gtData[:, :, np.newaxis, np.newaxis] * sup_cData_cDataT, axis=0)
unsup_cData = unsup_xData[:, np.newaxis, :] - new_clusterMeans[np.newaxis, :, :]
unsup_cData_cDataT = unsup_cData[:, :, :, np.newaxis] * unsup_cData[:, :, np.newaxis, :]
unsup_clusterStds = np.sum(unsup_resp[:, :, np.newaxis, np.newaxis] * unsup_cData_cDataT, axis=0)
new_clusterStds = (sup_clusterStds + unsup_clusterStds)/norm[:, np.newaxis, np.newaxis]
#Calculate sup LL
sup_prior_ll = np.sum(sup_gtData * np.log(new_clusterPrior[np.newaxis, :]))
sup_mvn_logdata = np.transpose(np.array([mvn.logpdf(sup_xData, new_clusterMeans[k], new_clusterStds[k]) for k in range(4)]))
sup_param_ll = np.sum(sup_gtData * sup_mvn_logdata)
sup_ll = sup_prior_ll + sup_param_ll
#Calculate unsup LL
unsup_prior_ll = np.sum(unsup_resp * np.log(new_clusterPrior[np.newaxis, :]))
unsup_mvn_logdata = np.transpose(np.array([mvn.logpdf(unsup_xData, new_clusterMeans[k], new_clusterStds[k]) for k in range(4)]))
unsup_param_ll = np.sum(unsup_resp * unsup_mvn_logdata)
unsup_ll = unsup_prior_ll + unsup_param_ll
ll = sup_ll + unsup_ll
print("Iteration", iteration, " ll:", ll)
#Set new params
clusterMeans = new_clusterMeans
clusterStds = new_clusterStds
clusterPrior = new_clusterPrior
#E step for all data
#mvn_data is [numData, numClusters]
all_mvn_data = np.transpose(np.array([mvn.pdf(xData, clusterMeans[k], clusterStds[k]) for k in range(4)]))
all_respNum = clusterPrior[np.newaxis, :] * all_mvn_data
all_respDen = np.sum(all_respNum, axis=1)
#resp is [numData, numClusters]
all_resp = all_respNum/all_respDen[:, np.newaxis]
#Visualize
resp_argmax = np.argmax(all_resp, axis=1)
resp_onehot = np.zeros(all_resp.shape)
resp_onehot[np.arange(len(resp_argmax)), resp_argmax] = 1
estImage = np.reshape(resp_onehot, [numTotal, ny, nx, 4])
#Calculate accuracy
estTestImg = estImage[7:]
testGtPosIdx = np.nonzero(testGt[:, :, :, 4] == 0)
posEst = estTestImg[testGtPosIdx]
posGt = testGt[testGtPosIdx][:, :-1]
accuracy = np.mean(np.equal(posEst, posGt).astype(np.float32))
print("Final accurcy:", accuracy)
plt.figure()
util.visualizeGt(estImage[-1, :, :, :], "estImage")
plt.figure()
util.visualizeGt(testGt[-1, :, :, :4], "gtImage")
plt.show()
pdb.set_trace()