nikcleju@55
|
1 """
|
nikcleju@55
|
2 #=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#
|
nikcleju@55
|
3 # Bob L. Sturm <bst@create.aau.dk> 20111018
|
nikcleju@55
|
4 # Department of Architecture, Design and Media Technology
|
nikcleju@55
|
5 # Aalborg University Copenhagen
|
nikcleju@55
|
6 # Lautrupvang 15, 2750 Ballerup, Denmark
|
nikcleju@55
|
7 #=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#
|
nikcleju@55
|
8 """
|
nikcleju@55
|
9
|
nikcleju@55
|
10 import numpy as np
|
nikcleju@55
|
11 from sklearn.utils import check_random_state
|
nikcleju@55
|
12 import time
|
nikcleju@55
|
13
|
nikcleju@55
|
14 from omp_sk_bugfix import orthogonal_mp
|
nikcleju@55
|
15 from omp_QR import greed_omp_qr
|
nikcleju@55
|
16 from omp_QR import omp_qr
|
nikcleju@55
|
17
|
nikcleju@55
|
18 """
|
nikcleju@55
|
19 Run a problem suite involving sparse vectors in
|
nikcleju@55
|
20 ambientDimension dimensional space, with a resolution
|
nikcleju@55
|
21 in the phase plane of numGradations x numGradations,
|
nikcleju@55
|
22 and at each indeterminacy and sparsity pair run
|
nikcleju@55
|
23 numTrials independent trials.
|
nikcleju@55
|
24
|
nikcleju@55
|
25 Outputs a text file denoting successes at each phase point.
|
nikcleju@55
|
26 For more on phase transitions, see:
|
nikcleju@55
|
27 D. L. Donoho and J. Tanner, "Precise undersampling theorems,"
|
nikcleju@55
|
28 Proc. IEEE, vol. 98, no. 6, pp. 913-924, June 2010.
|
nikcleju@55
|
29 """
|
nikcleju@55
|
30
|
nikcleju@55
|
31 def runProblemSuite(ambientDimension,numGradations,numTrials):
|
nikcleju@55
|
32
|
nikcleju@55
|
33 idx = np.arange(ambientDimension)
|
nikcleju@55
|
34 phaseDelta = np.linspace(0.05,1,numGradations)
|
nikcleju@55
|
35 phaseRho = np.linspace(0.05,1,numGradations)
|
nikcleju@55
|
36 success = np.zeros((numGradations, numGradations))
|
nikcleju@55
|
37
|
nikcleju@55
|
38 #Nic: init timers
|
nikcleju@55
|
39 t1all = 0
|
nikcleju@55
|
40 t2all = 0
|
nikcleju@55
|
41 t3all = 0
|
nikcleju@55
|
42
|
nikcleju@55
|
43 deltaCounter = 0
|
nikcleju@55
|
44 # delta is number of measurements/
|
nikcleju@55
|
45 for delta in phaseDelta[:17]:
|
nikcleju@55
|
46 rhoCounter = 0
|
nikcleju@55
|
47 for rho in phaseRho:
|
nikcleju@55
|
48 print(deltaCounter,rhoCounter)
|
nikcleju@55
|
49 numMeasurements = int(delta*ambientDimension)
|
nikcleju@55
|
50 sparsity = int(rho*numMeasurements)
|
nikcleju@55
|
51 # how do I set the following to be random each time?
|
nikcleju@55
|
52 generator = check_random_state(100)
|
nikcleju@55
|
53 # create unit norm dictionary
|
nikcleju@55
|
54 D = generator.randn(numMeasurements, ambientDimension)
|
nikcleju@55
|
55 D /= np.sqrt(np.sum((D ** 2), axis=0))
|
nikcleju@55
|
56 # compute Gramian (for efficiency)
|
nikcleju@55
|
57 DTD = np.dot(D.T,D)
|
nikcleju@55
|
58
|
nikcleju@55
|
59 successCounter = 0
|
nikcleju@55
|
60 trial = numTrials
|
nikcleju@55
|
61 while trial > 0:
|
nikcleju@55
|
62 # generate sparse signal with a minimum non-zero value
|
nikcleju@55
|
63 x = np.zeros((ambientDimension, 1))
|
nikcleju@55
|
64 idx2 = idx
|
nikcleju@55
|
65 generator.shuffle(idx2)
|
nikcleju@55
|
66 idx3 = idx2[:sparsity]
|
nikcleju@55
|
67 while np.min(np.abs(x[idx3,0])) < 1e-10 :
|
nikcleju@55
|
68 x[idx3,0] = generator.randn(sparsity)
|
nikcleju@55
|
69 # sense sparse signal
|
nikcleju@55
|
70 y = np.dot(D, x)
|
nikcleju@55
|
71
|
nikcleju@55
|
72 # Nic: Use sparsify OMP function (translated from Matlab)
|
nikcleju@55
|
73 ompopts = dict({'stopCrit':'M', 'stopTol':2*sparsity})
|
nikcleju@55
|
74 starttime = time.time() # start timer
|
nikcleju@55
|
75 x_r2, errs, times = greed_omp_qr(y.squeeze().copy(), D.copy(), D.shape[1], ompopts)
|
nikcleju@55
|
76 t2all = t2all + time.time() - starttime # stop timer
|
nikcleju@55
|
77 idx_r2 = np.nonzero(x_r2)[0]
|
nikcleju@55
|
78
|
nikcleju@55
|
79 # run to two times expected sparsity, or tolerance
|
nikcleju@55
|
80 # why? Often times, OMP can retrieve the correct solution
|
nikcleju@55
|
81 # when it is run for more than the expected sparsity
|
nikcleju@55
|
82 #x_r, idx_r = omp_qr(y,D,DTD,2*sparsity,1e-5)
|
nikcleju@55
|
83 # Nic: adjust tolerance to match with other function
|
nikcleju@55
|
84 starttime = time.time() # start timer
|
nikcleju@55
|
85 x_r, idx_r = omp_qr(y.copy(),D.copy(),DTD.copy(),2*sparsity,numMeasurements*1e-14/np.vdot(y,y))
|
nikcleju@55
|
86 t1all = t1all + time.time() - starttime # stop timer
|
nikcleju@55
|
87
|
nikcleju@55
|
88 # Nic: test sklearn omp
|
nikcleju@55
|
89 starttime = time.time() # start timer
|
nikcleju@55
|
90 x_r3 = orthogonal_mp(D.copy(), y.copy(), n_nonzero_coefs=2*sparsity, tol=numMeasurements*1e-14, precompute_gram=False, copy_X=True)
|
nikcleju@55
|
91 idx_r3 = np.nonzero(x_r3)[0]
|
nikcleju@55
|
92 t3all = t3all + time.time() - starttime # stop timer
|
nikcleju@55
|
93
|
nikcleju@55
|
94 # Nic: compare results
|
nikcleju@55
|
95 print 'diff1 = ',np.linalg.norm(x_r.squeeze() - x_r2.squeeze())
|
nikcleju@55
|
96 print 'diff2 = ',np.linalg.norm(x_r.squeeze() - x_r3.squeeze())
|
nikcleju@55
|
97 print 'diff3 = ',np.linalg.norm(x_r2.squeeze() - x_r3.squeeze())
|
nikcleju@55
|
98 print "Bob's total time = ", t1all
|
nikcleju@55
|
99 print "Nic's total time = ", t2all
|
nikcleju@55
|
100 print "Skl's total time = ", t3all
|
nikcleju@55
|
101 if np.linalg.norm(x_r.squeeze() - x_r2.squeeze()) > 1e-10 or \
|
nikcleju@55
|
102 np.linalg.norm(x_r.squeeze() - x_r3.squeeze()) > 1e-10 or \
|
nikcleju@55
|
103 np.linalg.norm(x_r2.squeeze() - x_r3.squeeze()) > 1e-10:
|
nikcleju@55
|
104 print "STOP: Different results"
|
nikcleju@55
|
105 print "Bob's residual: ||y - D x_r ||_2 = ",np.linalg.norm(y.squeeze() - np.dot(D,x_r).squeeze())
|
nikcleju@55
|
106 print "Nic's residual: ||y - D x_r ||_2 = ",np.linalg.norm(y.squeeze() - np.dot(D,x_r2).squeeze())
|
nikcleju@55
|
107 print "Skl's residual: ||y - D x_r ||_2 = ",np.linalg.norm(y.squeeze() - np.dot(D,x_r3).squeeze())
|
nikcleju@55
|
108 raise ValueError("Different results")
|
nikcleju@55
|
109
|
nikcleju@55
|
110 # debais to remove small entries
|
nikcleju@55
|
111 for nn in idx_r:
|
nikcleju@55
|
112 if abs(x_r[nn]) < 1e-10:
|
nikcleju@55
|
113 x_r[nn] = 0
|
nikcleju@55
|
114
|
nikcleju@55
|
115 # exact recovery condition using support
|
nikcleju@55
|
116 #if sorted(np.flatnonzero(x_r)) == sorted(np.flatnonzero(x)):
|
nikcleju@55
|
117 # successCounter += 1
|
nikcleju@55
|
118 # exact recovery condition using error in solution
|
nikcleju@55
|
119 error = x - x_r
|
nikcleju@55
|
120 """ the following is the exact recovery condition in: A. Maleki
|
nikcleju@55
|
121 and D. L. Donoho, "Optimally tuned iterative reconstruction
|
nikcleju@55
|
122 algorithms for compressed sensing," IEEE J. Selected Topics
|
nikcleju@55
|
123 in Signal Process., vol. 4, pp. 330-341, Apr. 2010. """
|
nikcleju@55
|
124 if np.vdot(error,error) < np.vdot(x,x)*1e-4:
|
nikcleju@55
|
125 successCounter += 1
|
nikcleju@55
|
126 trial -= 1
|
nikcleju@55
|
127
|
nikcleju@55
|
128 success[rhoCounter,deltaCounter] = successCounter
|
nikcleju@55
|
129 if successCounter == 0:
|
nikcleju@55
|
130 break
|
nikcleju@55
|
131
|
nikcleju@55
|
132 rhoCounter += 1
|
nikcleju@55
|
133 #np.savetxt('test.txt',success,fmt='#2.1d',delimiter=',')
|
nikcleju@55
|
134 deltaCounter += 1
|
nikcleju@55
|
135
|
nikcleju@55
|
136 if __name__ == '__main__':
|
nikcleju@55
|
137 print ('Running problem suite')
|
nikcleju@55
|
138 ambientDimension = 400
|
nikcleju@55
|
139 numGradations = 30
|
nikcleju@55
|
140 numTrials = 1
|
nikcleju@55
|
141
|
nikcleju@55
|
142 #import cProfile
|
nikcleju@55
|
143 #cProfile.run('runProblemSuite(ambientDimension,numGradations,numTrials)','profres')
|
nikcleju@55
|
144 runProblemSuite(ambientDimension,numGradations,numTrials)
|
nikcleju@55
|
145 print "Done"
|
nikcleju@55
|
146
|
nikcleju@55
|
147 #import pstats
|
nikcleju@55
|
148 #p = pstats.Stats('D:\Nic\Dev2\profres')
|
nikcleju@55
|
149 #p.sort_stats('cumulative').print_stats(10)
|