def runupdates():
X = np.random.rand(5000, 1000)
N, d = X.shape
V = np.random.rand(N, 2) * 2 - 1
V = (V - V.mean(0)) / V.std(0)
updates = []
i = 0
for N in range(100, 5001, 100):
elm = hpelm.ELM(2, d)
elm.add_neurons(2, 'lin')
elm.add_neurons(50, 'sigm')
H = elm.project(V[:N])
A = H.dot(np.linalg.pinv(np.dot(H.T, H))).dot(H.T)
X1 = X[:N].copy()
_, info = elmvis_c(X1, A, tol=-100, maxupdate=1000, cossim=1, silent=1)
ups_c = info['ups']
X1 = X[:N].copy()
_, info = elmvis_p(X1, A, tol=-100, maxupdate=1000, cossim=1, silent=1)
ups_p = info['ups']
X1 = X[:N].copy()
_, info = elmvis_p(X1,
A,
tol=-100,
maxupdate=1000,
batch=1,
cossim=1,
silent=1)
ups_p1 = info['ups']
X1 = X[:N].copy()
_, info = elmvis_g(X1, A, tol=-100, maxupdate=1000, cossim=1, silent=1)
ups_g = info['ups']
updates.append((ups_c, ups_p, ups_p1, ups_g))
print "%d/5000" % N
updates = np.array(updates)
N = range(100, 5001, 100)
plt.plot(N, updates[:, 0], '--r', linewidth=3, alpha=0.7)
plt.plot(N, updates[:, 1], '--k')
plt.plot(N, updates[:, 2], '-k')
plt.plot(N, updates[:, 3], '-g')
plt.ylabel("updates per second")
plt.legend(("C opt, 1% batch", "Python, 1% batch", "Python", "GPU"))
plt.title("$N$ data points with 1000 features")
plt.xlabel("$N$")
plt.xlim([0, 5000])
#plt.ylim([0, 23000000])
#plt.yticks([0, 1000000, 5000000, 10000000, 15000000, 20000000, 22000000], ["0", "1 mln.", "5 mln.", "10 mln.", "15 mln.", "20 mln.", "22 mln."])
plt.savefig("plot_ups.pdf", bbox_inches="tight")
print "done"
plt.show()
def runupdates():
X = np.random.rand(5000, 1000)
N, d = X.shape
V = np.random.rand(N, 2)*2 - 1
V = (V - V.mean(0)) / V.std(0)
updates = []
i = 0
for N in range(100, 5001, 100):
elm = hpelm.ELM(2, d)
elm.add_neurons(2, 'lin')
elm.add_neurons(50, 'sigm')
H = elm.project(V[:N])
A = H.dot(np.linalg.pinv(np.dot(H.T, H))).dot(H.T)
X1 = X[:N].copy()
_, info = elmvis_c(X1, A, tol=-100, maxupdate=1000, cossim=1, silent=1)
ups_c = info['ups']
X1 = X[:N].copy()
_, info = elmvis_p(X1, A, tol=-100, maxupdate=1000, cossim=1, silent=1)
ups_p = info['ups']
X1 = X[:N].copy()
_, info = elmvis_p(X1, A, tol=-100, maxupdate=1000, batch=1, cossim=1, silent=1)
ups_p1 = info['ups']
X1 = X[:N].copy()
_, info = elmvis_g(X1, A, tol=-100, maxupdate=1000, cossim=1, silent=1)
ups_g = info['ups']
updates.append((ups_c, ups_p, ups_p1, ups_g))
print "%d/5000" % N
updates = np.array(updates)
N = range(100, 5001, 100)
plt.plot(N, updates[:,0], '--r', linewidth=3, alpha=0.7)
plt.plot(N, updates[:,1], '--k')
plt.plot(N, updates[:,2], '-k')
plt.plot(N, updates[:,3], '-g')
plt.ylabel("updates per second")
plt.legend(("C opt, 1% batch", "Python, 1% batch", "Python", "GPU"))
plt.title("$N$ data points with 1000 features")
plt.xlabel("$N$")
plt.xlim([0, 5000])
#plt.ylim([0, 23000000])
#plt.yticks([0, 1000000, 5000000, 10000000, 15000000, 20000000, 22000000], ["0", "1 mln.", "5 mln.", "10 mln.", "15 mln.", "20 mln.", "22 mln."])
plt.savefig("plot_ups.pdf", bbox_inches="tight")
print "done"
plt.show()
def runswaps():
X = np.random.rand(1000, 10000)
N, d = X.shape
V = np.random.rand(N, 2)*2 - 1
V = (V - V.mean(0)) / V.std(0)
swaps = []
i = 0
for d in range(20, 1001, 20):
elm = hpelm.ELM(2, d)
elm.add_neurons(2, 'lin')
elm.add_neurons(20, 'sigm')
H = elm.project(V)
A = H.dot(np.linalg.pinv(np.dot(H.T, H))).dot(H.T)
X1 = X[:, :d].copy()
_, info = elmvis_c(X1, A, tol=100, maxiter=10000, maxstall=10000, silent=1)
ips_c = info['ips']
X1 = X[:, :d].copy()
_, info = elmvis_p(X1, A, tol=100, maxiter=10000, maxstall=10000, silent=1)
ips_p = info['ips']
X1 = X[:, :d].copy()
_, info = elmvis_g(X1, A, tol=100, maxiter=10000, maxstall=10000, silent=1)
ips_g = info['ips']
swaps.append((ips_c, ips_p, ips_g))
print d
swaps = np.array(swaps)
d = range(20, 1001, 20)
plt.plot(d, swaps[:,0], '--r', linewidth=3, alpha=0.7)
plt.plot(d, swaps[:,1], '--k')
plt.plot(d, swaps[:,2], '--g')
plt.ylabel("swaps per second")
plt.legend(("C optimized", "Python", "GPU"))
plt.title("1000 data points with $d$ features")
plt.xlabel("$d$")
plt.xlim([0, 1000])
#plt.ylim([0, 23000000])
#plt.yticks([0, 1000000, 5000000, 10000000, 15000000, 20000000, 22000000], ["0", "1 mln.", "5 mln.", "10 mln.", "15 mln.", "20 mln.", "22 mln."])
plt.savefig("plot_ips.pdf", bbox_inches="tight")
print "done"
plt.show()
def runswaps():
X = np.random.rand(1000, 10000)
N, d = X.shape
V = np.random.rand(N, 2) * 2 - 1
V = (V - V.mean(0)) / V.std(0)
swaps = []
i = 0
for d in range(20, 1001, 20):
elm = hpelm.ELM(2, d)
elm.add_neurons(2, 'lin')
elm.add_neurons(20, 'sigm')
H = elm.project(V)
A = H.dot(np.linalg.pinv(np.dot(H.T, H))).dot(H.T)
X1 = X[:, :d].copy()
_, info = elmvis_c(X1,
A,
tol=100,
maxiter=10000,
maxstall=10000,
silent=1)
ips_c = info['ips']
X1 = X[:, :d].copy()
_, info = elmvis_p(X1,
A,
tol=100,
maxiter=10000,
maxstall=10000,
silent=1)
ips_p = info['ips']
X1 = X[:, :d].copy()
_, info = elmvis_g(X1,
A,
tol=100,
maxiter=10000,
maxstall=10000,
silent=1)
ips_g = info['ips']
swaps.append((ips_c, ips_p, ips_g))
print d
swaps = np.array(swaps)
d = range(20, 1001, 20)
plt.plot(d, swaps[:, 0], '--r', linewidth=3, alpha=0.7)
plt.plot(d, swaps[:, 1], '--k')
plt.plot(d, swaps[:, 2], '--g')
plt.ylabel("swaps per second")
plt.legend(("C optimized", "Python", "GPU"))
plt.title("1000 data points with $d$ features")
plt.xlabel("$d$")
plt.xlim([0, 1000])
#plt.ylim([0, 23000000])
#plt.yticks([0, 1000000, 5000000, 10000000, 15000000, 20000000, 22000000], ["0", "1 mln.", "5 mln.", "10 mln.", "15 mln.", "20 mln.", "22 mln."])
plt.savefig("plot_ips.pdf", bbox_inches="tight")
print "done"
plt.show()