def test5():
verbose = False
mat = sio.loadmat('../matlab/M.mat')
n = int(mat['n'])
r = int(mat['r'])
nt = int(mat['nt'])
rcross = int(mat['rcross'])
d = int(mat['d'])
num_eval = mat['num_eval']
init_pt = int(mat['init_pt'])
Xfull = mat['Xfull']
deg = np.squeeze(mat['deg'])
Yfull = np.squeeze(mat['Yfull'])
pi = sum(Yfull)/len(Yfull)
t1 = time.time()
Xe, b, w, deg = eigenmap(Xfull.T.dot(Xfull), d)
t2 = time.time()
f2,_,_ = AS.lreg_AS (Xe, deg, d, alpha=0.0, labels=Yfull, options={'num_eval':num_eval,'pi':pi,'n_conncomp':b,'init_pt':init_pt}, verbose=verbose)
t3 = time.time()
f1,_,_ = AS.kernel_AS (Xfull, Yfull, pi=pi, num_eval=num_eval, init_pt=init_pt, verbose=verbose)
t4 = time.time()
print "Time taken for eigenmap + computing X.T*X:", t2-t1
print "Time taken for lreg:", t3-t2
print "Time taken for kernel:", t4-t3
f1 = np.squeeze(f1)
f2 = np.squeeze(f2)
f3 = np.squeeze(mat['f'])
import IPython
IPython.embed()
def test13(): # testing sparse AS
verbose = True
#ts_data = ef.load_timestamps (tsfile)
Xfull = load_sparse_csr('Xfull1.npz')
print Xfull.shape
Xfull = Xfull[np.squeeze(np.asarray(np.nonzero(Xfull.sum(axis=1))[0])),:]
Xfull = Xfull[:,np.squeeze(np.asarray(np.nonzero(Xfull.sum(axis=0))[1]))]
r,n = Xfull.shape
nt = int(0.1*n)
num_eval = 20
# getting rid of features which are zero for all these elements
# X = np.array((Xfull[:,:n]).todense())
# X = X[np.nonzero(X.sum(axis=1))[0],:]
# X = X[:,np.nonzero(X.sum(axis=0))[1]]
# import IPython
# IPython.embed()
# X = X[:r,:]
# X = X[np.nonzero(X.sum(axis=1))[0],:]
# X = X[:,np.nonzero(X.sum(axis=0))[1]]
#X = np.load('X11.npy')
# import IPython
# IPython.embed()
num_eval = 20
# num_eval = nt*2
Y = np.array([1]*nt + [0]*(n-nt), dtype=int)
pi = sum(Y)/len(Y)
init_pt = 100
t1 = time.time()
f1,h1,s1,fs1,dtinv1 = AS.kernel_AS (Xfull, Y, pi=pi, num_eval=num_eval, init_pt=init_pt, verbose=verbose,all_fs=True,tinv=True,sparse=True)
t2 = time.time()
f2,h2,s2,fs2,dtinv2 = AS.kernel_AS (np.array(Xfull.todense()), Y, pi=pi, num_eval=num_eval, init_pt=init_pt, verbose=verbose,all_fs=True,tinv=True,sparse=False)
t3 = time.time()
import IPython
IPython.embed()
def test8():
n = 500
r = 50
nt = 100
#rcross = 50
d = 50
hubs = 1
verbose = False
num_eval = 100
#X, Y = createFakeData(n, r, nt, rcross)
#X, Y = createFakeData2(n, r, nt, hubs)
X,Y = np.load('t8data.npy')
# import IPython
# IPython.embed()
init_pt = np.nonzero(Y)[0][0]
ker = True
pi = sum(Y)/len(Y)
print "Constructing the similarity matrix:"
A = X.T.dot(X)
t1 = time.time()
if ker:
print "Performing Kernel AS"
f1,h1,s1,fs1 = AS.kernel_AS (X, Y, pi=pi, num_eval=num_eval, init_pt=init_pt, verbose=verbose,all_fs=True)
t2 = time.time()
#print "Performing Eigen decmop"
#Xe, b, w, deg = eigenmap(A, d)
#t3 = time.time()
if ker:
print "Performing Naive Shari AS"
f2,h2,s2,fs2 = AS.shari_activesearch_probs_naive(A, labels=Y, pi=pi, w0=None, eta=None, num_eval=num_eval, init_pt=init_pt, verbose=verbose, all_fs=True)
#f2,h2,s2,fs2 = AS.lreg_AS (Xe, deg, d, alpha=0.0, labels=Y, options={'num_eval':num_eval,'pi':pi,'n_conncomp':b}, verbose=verbose)
t4 = time.time()
print "Time taken for kernel:", t2-t1
#print "Time taken for eigenmap + computing X.T*X:", t3-t2
print "Time taken for Shari's method (naive):", t4-t2
if ker:
print "h_kernel: %i/%i"%(h1[-1],num_eval)
print "h_lreg: %i/%i"%(h2[-1],num_eval)
import IPython
IPython.embed()
def test2 (n, cc=2, nt=1, d=5):
# Created banded diagonal kernel matrix
X = np.eye(n)
X[range(n-1), range(1,n)] = 1
Xfull = slg.block_diag(*([X]*cc))
Yfull = [0]+[1]+[0]*(n-5)+[1]*3+[1]*((nt-1)*n) + [0]*((cc-nt)*n)
Xe, b, w, deg = eigenmap(Xfull.T.dot(Xfull), d)
num_eval = 5
init_pt = 1
pi = sum(Yfull)/len(Yfull)
f1 = AS.kernel_AS (Xfull, Yfull, pi=pi, num_eval=num_eval, init_pt=init_pt)
f2 = AS.lreg_AS (Xe, deg, d, alpha=0.0, labels=Yfull, options={'num_eval':num_eval,'pi':pi,'init_pt':init_pt,'n_conncomp':b})
import IPython
IPython.embed()
def test1 (n, cc=2, nt=1, d=4):
# Example where you chains of connections in targets
# and similar chains in non-targets
num_eval = 5
init_pt = 1
# Created banded diagonal kernel matrix
X = np.eye(n)
X[range(n-1), range(1,n)] = 1
Xfull = slg.block_diag(*([X]*cc))
Yfull = [1]*(nt*n) + [0]*((cc-nt)*n)
pi = sum(Yfull)/len(Yfull)
f1,_,_ = AS.kernel_AS (Xfull, Yfull, pi=pi, num_eval=num_eval, init_pt=init_pt)
Xe, b, w, deg = eigenmap(Xfull.T.dot(Xfull), d)
f2,_,_ = AS.lreg_AS (Xe, deg, d, alpha=0.0, labels=Yfull, options={'num_eval':num_eval,'pi':pi,'init_pt':init_pt,'n_conncomp':b})
import IPython
IPython.embed()
def test4():
n = 1000
r = 100
nt = 100
#rcross = 50
d = 10
hubs = 1
verbose = False
num_eval = 300
#init_pt = 1
#X, Y = createFakeData(n, r, nt, rcross)
X, Y = createFakeData2(n, r, nt, hubs)
pi = sum(Y)/len(Y)
print "Constructing the similarity matrix:"
A = X.T.dot(X)
t1 = time.time()
print "Performing Kernel AS"
f1,h1,s1 = AS.kernel_AS (X, Y, pi=pi, num_eval=num_eval, init_pt=None, verbose=verbose)
t2 = time.time()
print "Performing Eigen decmop"
Xe, b, w, deg = eigenmap(A, d)
t3 = time.time()
print "Performing LREG AS"
f2,h2,s2 = AS.lreg_AS (Xe, deg, d, alpha=0.0, labels=Y, options={'num_eval':num_eval,'pi':pi,'n_conncomp':b}, verbose=verbose)
t4 = time.time()
print "Time taken for kernel:", t2-t1
print "Time taken for eigenmap + computing X.T*X:", t3-t2
print "Time taken for lreg:", t4-t3
print "h_kernel: %i/%i"%(h1[-1],num_eval)
print "h_lreg: %i/%i"%(h2[-1],num_eval)
import IPython
IPython.embed()
def test_interface ():
verbose = False
#ts_data = ef.load_timestamps (tsfile)
Xfull = load_sparse_csr('Xfull1.npz')
r,n = Xfull.shape
nt = int(0.05*n)
num_eval = 1000
# num_eval = nt*2
Y = np.array([1]*nt + [0]*(n-nt), dtype=int)
pi = sum(Y)/len(Y)
init_pt = 100
t1 = time.time()
prms = ASI.Parameters(pi=pi,sparse=True, verbose=verbose)
kAS = ASI.kernelAS(prms)
kAS.initialize(Xfull)
kAS.firstMessage(init_pt)
fs2 = [kAS.f]
for i in range(num_eval):
idx = kAS.getNextMessage()
kAS.setLabelCurrent(Y[idx])
fs2.append(kAS.f)
t2 = time.time()
f1,h1,s1,fs1,dtinv1 = AS.kernel_AS (Xfull, Y, pi=pi, num_eval=num_eval, init_pt=init_pt, verbose=verbose,all_fs=True,tinv=True,sparse=True)
t3 = time.time()
checks = [np.allclose(fs1[i],fs2[i]) for i in range(len(fs1))]
import IPython
IPython.embed()
def testfake(): n = 400 r = 60 nt = 150 #rcross = 50 d = n hubs = 1 verbose = False num_eval = 390 #init_pt = 1 for i in range(100): #X, Y = createFakeData(n, r, nt, rcross) X, Y = createFakeData2(n, r, nt, hubs) pi = sum(Y)/len(Y) print i # print "Constructing the similarity matrix:" # print "Performing Kernel AS" f1,h1,s1 = AS.kernel_AS (X, Y, pi=pi, num_eval=num_eval, init_pt=None, verbose=verbose)
def test3 (hubs=3, followers=10):
# Creating as follows:
# Every alternate hub and half its followers are targets.
# Every hub shares followers with adjacent hubs
fl0 = int(followers/2)
fl1 = followers - fl0
X = np.zeros((hubs*followers, (hubs-1)*followers + hubs))
Yfull = []
t = 0
for i in range(hubs-1):
idx = i*(followers+1)
X[i*followers:(i+1)*followers, idx] = 1
X[xrange(i*followers,(i+1)*followers), xrange(idx+1, idx+followers+1)] = [1]*followers
X[xrange((i+1)*followers,(i+2)*followers), xrange(idx+1, idx+followers+1)] = [1]*followers
if t == 0:
Yfull += [0]*(fl0+1) + [1]*fl1
t = 1
else:
Yfull += [1]*(fl1+1) + [0]*fl0
t = 0
X[(hubs-1)*followers:, -1] = 1
if t == 0:
Yfull.append(0)
else:
Yfull.append(1)
Xfull = X
d = 10
num_eval = 15
init_pt = 10
pi = sum(Yfull)/len(Yfull)
# t1 = time.time()
f1,_,_ = AS.kernel_AS (Xfull, Yfull, pi=pi, num_eval=num_eval, init_pt=init_pt)
# t2 = time.time()
Xe, b, w, deg = eigenmap(Xfull.T.dot(Xfull), d)
f2,_,_ = AS.lreg_AS (Xe, deg, d, alpha=0.0, labels=Yfull, options={'num_eval':num_eval,'pi':pi,'init_pt':init_pt,'n_conncomp':b})
# t3 = time.time()
# print "Time 1:", (t2-t1)
# print "Time 2:", (t3-t2)
# print f1
# print f2
plt.plot(f1, color='r', label='kernel')
plt.plot(f2, color='b', label='lreg')
plt.plot(Yfull, color='g', label='true')
plt.legend()
plt.show()
mat = sio.loadmat('../matlab/L.mat')
Xe2 = mat['Xe']
deg2 = np.squeeze(mat['deg'])
Yfull2 = np.squeeze(mat['Yfull'])
f3,_,_ = AS.lreg_AS (Xe2, deg2, d, alpha=0.0, labels=Yfull2, options={'num_eval':num_eval,'pi':pi,'init_pt':init_pt,'n_conncomp':b})
import IPython
IPython.embed()
def test12():
verbose = True
#ts_data = ef.load_timestamps (tsfile)
Xfull = load_sparse_csr('Xfull1.npz')
r,n = Xfull.shape
nt = int(0.1*n)
num_eval = nt*2
X = np.array(Xfull.todense())
# getting rid of features which are zero for all these elements
X = X[np.nonzero(X.sum(axis=1))[0],:]
X = X[:,np.nonzero(X.sum(axis=0))[1]]
# import IPython
# IPython.embed()
X = X[:r,:]
X = X[np.nonzero(X.sum(axis=1))[0],:]
X = X[:,np.nonzero(X.sum(axis=0))[1]]
#X2 = np.load('X11.npy')
# import IPython
# IPython.embed()
nt = int(0.1*n)
num_eval = nt*2
Y = np.array([1]*nt + [0]*(n-nt), dtype=int)
#rrange = [10*i for i in range(2,11)] + [100*i for i in range(2,11,2)] + [1500, 2000, 2500, 3000]
pi = sum(Y)/len(Y)
init_pt = 100
#hits = []
#rtime = []
#for r in rrange:
t1 = time.time()
print "Performing the kernel AS"
f1,h1,s1,fs1 = AS.kernel_AS (X, Y, pi=pi, num_eval=num_eval, init_pt=init_pt, verbose=verbose,all_fs=True,sparse=False)
t2 = time.time()
# hits.append(h1[-1][0])
# rtime.append(t2-t1)
# print "r =", r
# print "hits =", h1[-1][0]
print "time =", t2-t1
# print
# print
# plt.figure()
# plt.plot(rrange, hits)
# plt.xlabel('d')
# plt.ylabel('hits')
# plt.figure()
# plt.plot(rrange, rtime)
# plt.xlabel('d')
# plt.ylabel('time in s')
# plt.show()
# Timing
# Constructing C: 71s
# Inverse of C: 202s
# Total time: 3446s
import IPython
IPython.embed()
def test11():
verbose = False
#ts_data = ef.load_timestamps (tsfile)
Xfull = load_sparse_csr('Xfull1.npz')
n = 5000
nt = int(0.1*n)
num_eval = nt*2
# getting rid of features which are zero for all these elements
# X = np.array((Xfull[:,:n]).todense())
# X = X[np.nonzero(X.sum(axis=1))[0],:]
# X = X[:,np.nonzero(X.sum(axis=0))[1]]
# import IPython
# IPython.embed()
# X = X[:r,:]
# X = X[np.nonzero(X.sum(axis=1))[0],:]
# X = X[:,np.nonzero(X.sum(axis=0))[1]]
X = np.load('X11.npy')
# import IPython
# IPython.embed()
nt = int(0.1*n)
num_eval = nt*2
Y = np.array([1]*nt + [0]*(n-nt), dtype=int)
rrange = [10*i for i in range(2,11)] + [100*i for i in range(2,11,2)] + [1500, 2000, 2500]
pi = sum(Y)/len(Y)
init_pt = 100
hits = []
rtime = []
for r in rrange:
Xr = X[:r,:]
Xr = Xr[np.nonzero(Xr.sum(axis=1))[0],:]
Xr = Xr[:,np.nonzero(Xr.sum(axis=0))[1]]
t1 = time.time()
f1,h1,s1,fs1 = AS.kernel_AS (Xr, Y, pi=pi, num_eval=num_eval, init_pt=init_pt, verbose=verbose,all_fs=True)
t2 = time.time()
hits.append(h1[-1][0])
rtime.append(t2-t1)
print "r =", r
print "hits =", h1[-1][0]
print "time =", t2-t1
print
print
# plt.figure()
# plt.plot(rrange, hits)
# plt.xlabel('d')
# plt.ylabel('hits')
# plt.figure()
# plt.plot(rrange, rtime)
# plt.xlabel('d')
# plt.ylabel('time in s')
# plt.show()
# import IPython
# IPython.embed()
return rrange, hits, rtime
def test6():
n = 10
start = 1000
end = 5000
nt_ratio = 0.6
r_ratio = 0.1
ne_ratio = 0.2
nrange = np.logspace(np.log10(start),np.log10(end), n).tolist()
nrange = [int(nv) for nv in nrange]
ntrange = [int(nt_ratio*nv) for nv in nrange]
rrange = [int(r_ratio*nv) for nv in nrange]
drange = [r*2 for r in rrange]
nerange = [int(ne_ratio*nv) for nv in nrange]
t_kernel = []
t_eigendecomp = []
t_lreg = []
max_error = []
verbose = False
for n,nt,r,d,ne in zip(nrange,ntrange,rrange,drange,nerange):
X, Y = createFakeData(n, r, nt, rcross=0)
pi = sum(Y)/len(Y)
A = X.T.dot(X) # Don't really want to include timings for this.
t1 = time.time()
f1,h1,s1 = AS.kernel_AS (X, Y, pi=pi, num_eval=ne, init_pt=None, verbose=verbose)
t2 = time.time()
Xe, b, w, deg = eigenmap(A, d)
t3 = time.time()
f2,h2,s2 = AS.lreg_AS (Xe, deg, d, alpha=0.0, labels=Y, options={'num_eval':ne,'pi':pi,'n_conncomp':b}, verbose=verbose)
t4 = time.time()
print "Parameters: n=%i, nt=%i, r=%i, d=%i, ne=%i"%(n,nt,2*r,d,ne)
print "Time taken for kernel:", t2-t1
print "Time taken for eigenmap + computing X.T*X:", t3-t2
print "Time taken for lreg:", t4-t3
print "Max error difference:", np.max(np.abs(f1-f2))
print
t_kernel.append(t2-t1)
t_eigendecomp.append(t3-t2)
t_lreg.append(t4-t3)
max_error.append(np.max(np.abs(f1-f2)))
plt.plot(nrange, t_kernel, label='kernel')
plt.plot(nrange, t_eigendecomp, label='eigendecomp')
plt.plot(nrange, t_lreg, label='lreg')
plt.legend()
plt.xlabel('Number of data points')
plt.ylabel('Time')
plt.show()
import IPython
IPython.embed()