def run_bench(num_iter, s, k):
wrld = ctf.comm()
M = ctf.random.random((s,s))
X = ctf.random.random((k,s))
[U,S,VT] = ctf.svd(M)
S = np.arange(0,s)+1
M = ctf.dot(U*S,U.T())
te = ctf.timer_epoch("BENCHMARK: SPD SOLVE")
te.begin()
times = []
for i in range(num_iter):
t0 = time.time()
X = ctf.solve_spd(M,X)
times.append(time.time()-t0)
te.end()
if ctf.comm().rank() == 0:
print("ctf.solve_spd average time:",np.sum(times)/num_iter,"sec")
print("ctf.solve_spd iteration timings:",times)
te = ctf.timer_epoch("BENCHMARK: Manual Cholesky+TRSM SPD SOLVE")
te.begin()
times = []
for i in range(num_iter):
t0 = time.time()
L = ctf.cholesky(M)
X = ctf.solve_tri(M,X,from_left=False)
times.append(time.time()-t0)
te.end()
if ctf.comm().rank() == 0:
print("ctf.cholesky+solve_tri average time:",np.sum(times)/num_iter,"sec")
print("ctf.cholesky+solve_tri iteration timings:",times)
def getALS_SVD(T, U, V, W, regParam, omega, I, J, K, r):
it = 0
E = ctf.tensor((I, J, K))
E.i("ijk"
) << T.i("ijk") - omega.i("ijk") * U.i("iu") * V.i("ju") * W.i("ku")
curr_err_norm = ctf.vecnorm(E) + (ctf.vecnorm(U) + ctf.vecnorm(V) +
ctf.vecnorm(W)) * regParam
while True:
U = updateFactor_SVD(T, U, V, W, regParam, omega, I, J, K, r, "U")
V = updateFactor_SVD(T, U, V, W, regParam, omega, I, J, K, r, "V")
W = updateFactor_SVD(T, U, V, W, regParam, omega, I, J, K, r, "W")
E.set_zero()
E.i("ijk") << T.i(
"ijk") - omega.i("ijk") * U.i("iu") * V.i("ju") * W.i("ku")
next_err_norm = ctf.vecnorm(E) + (ctf.vecnorm(U) + ctf.vecnorm(V) +
ctf.vecnorm(W)) * regParam
if ctf.comm().rank() == 0:
print(curr_err_norm, next_err_norm)
if abs(curr_err_norm - next_err_norm) < .001 or it > 20:
break
curr_err_norm = next_err_norm
it += 1
if ctf.comm().rank() == 0:
print("Number of iterations: ", it)
return U, V, W
def run_tests():
numpy.random.seed(5330);
wrld = ctf.comm()
if ctf.comm().rank() != 0:
result = unittest.TextTestRunner(stream = open(os.devnull, 'w')).run(unittest.TestSuite(unittest.TestLoader().loadTestsFromTestCase(KnowValues)))
else:
print("Tests for basic numpy ndarray functionality")
result = unittest.TextTestRunner().run(unittest.TestSuite(unittest.TestLoader().loadTestsFromTestCase(KnowValues)))
return result
def getALS_CG(T, U, V, W, regParam, omega, I, J, K, r, block):
t0 = time.time()
it = 0
E = ctf.tensor((I, J, K), sp=True)
E.i("ijk"
) << T.i("ijk") - omega.i("ijk") * U.i("iu") * V.i("ju") * W.i("ku")
if ctf.comm().rank() == 0:
print("contraction to form the error tensor cost %f seconds" %
(np.round_(time.time() - t0, 4)))
assert (E.sp == 1)
curr_err_norm = ctf.vecnorm(E) + (ctf.vecnorm(U) + ctf.vecnorm(V) +
ctf.vecnorm(W)) * regParam
while True:
t = time.time()
U = updateFactor_CG(T, U, V, W, regParam, omega, I, J, K, r, block,
"U")
assert (U.sp == 1)
V = updateFactor_CG(T, U, V, W, regParam, omega, I, J, K, r, block,
"V")
assert (V.sp == 1)
W = updateFactor_CG(T, U, V, W, regParam, omega, I, J, K, r, block,
"W")
assert (W.sp == 1)
if ctf.comm().rank() == 0:
print(
"CG update factor matrices in the following iteration cost %f seconds"
% np.round_(time.time() - t, 4))
E.set_zero()
E.i("ijk") << T.i(
"ijk") - omega.i("ijk") * U.i("iu") * V.i("ju") * W.i("ku")
assert (E.sp == 1)
next_err_norm = ctf.vecnorm(E) + (ctf.vecnorm(U) + ctf.vecnorm(V) +
ctf.vecnorm(W)) * regParam
if ctf.comm().rank() == 0:
print(curr_err_norm, next_err_norm)
it += 1
if abs(curr_err_norm - next_err_norm) < .001 or it > 100:
break
curr_err_norm = next_err_norm
nt = np.round_(time.time() - t0, 4)
return it, nt
def run_tests():
numpy.random.seed(5330)
wrld = ctf.comm()
print("I am rank")
if ctf.comm().rank() != 0:
result = unittest.TextTestRunner(stream=open(os.devnull, 'w')).run(
unittest.TestSuite(
unittest.TestLoader().loadTestsFromTestCase(KnowValues)))
else:
print("Tests for partition")
result = unittest.TextTestRunner().run(
unittest.TestSuite(
unittest.TestLoader().loadTestsFromTestCase(KnowValues)))
return result
def main():
regParam = 0.00001
file_name = sys.argv[1]
I = int(sys.argv[2])
J = int(sys.argv[3])
K = int(sys.argv[4])
stepSize = int(sys.argv[5])
r = int(sys.argv[6])
sample_rate = float(sys.argv[7])
T = read_from_frostt(file_name, I, J, K)
# T.read_from_file("T.txt")
Omega = getOmega(T)
s = T.sum()
os = Omega.sum()
if ctf.comm().rank() == 0:
print(s, os)
print(sys.argv)
#T = function_tensor(I, J, K, sparsity)
U = ctf.random.random((I, r))
V = ctf.random.random((J, r))
W = ctf.random.random((K, r))
#T.write_to_file("T.txt")
sparse_SGD(T, U, V, W, regParam, Omega, I, J, K, r, stepSize, sample_rate,
100, 1.e-5, 1.E3, 30, 10)
def sparse_SGD(T, U, V, W, Lambda, omega, I, J, K, r, stepSize, sample_rate, num_iter, errThresh, time_limit, work_cycle, use_MTTKRP):
times = [0 for i in range(7)]
iteration_count = 0
total_count = 0
R = ctf.tensor((I, J, K), sp=T.sp)
if T.sp == True:
nnz_tot = T.nnz_tot
else:
nnz_tot = ctf.sum(omega)
start_time = time.time()
starting_time = time.time()
dtime = 0
R.i("ijk") << T.i("ijk") - ctf.TTTP(omega, [U, V, W]).i("ijk")
curr_err_norm = ctf.vecnorm(R) + (ctf.vecnorm(U) + ctf.vecnorm(V) + ctf.vecnorm(W)) * Lambda
times[0] += time.time() - starting_time
norm = [curr_err_norm]
step = stepSize * 0.5
t_obj_calc = 0.
while iteration_count < num_iter and time.time() - start_time - t_obj_calc < time_limit:
iteration_count += 1
starting_time = time.time()
sampled_T = T.copy()
sampled_T.sample(sample_rate)
times[1] += time.time() - starting_time
sparse_update(sampled_T, [U, V, W], Lambda, [I, J, K], r, stepSize * 0.5 + step, sample_rate, times, use_MTTKRP)
#step *= 0.99
sampled_T.set_zero()
if iteration_count % work_cycle == 0:
duration = time.time() - start_time - t_obj_calc
t_b_obj = time.time()
total_count += 1
R.set_zero()
R.i("ijk") << T.i("ijk") - ctf.TTTP(omega, [U, V, W]).i("ijk")
diff_norm = ctf.vecnorm(R)
RMSE = diff_norm/(nnz_tot**.5)
next_err_norm = diff_norm + (ctf.vecnorm(U) + ctf.vecnorm(V) + ctf.vecnorm(W)) * Lambda
if glob_comm.rank() == 0:
print('Objective after',duration,'seconds (',iteration_count,'iterations) is: {}'.format(next_err_norm))
print('RMSE after',duration,'seconds (',iteration_count,'iterations) is: {}'.format(RMSE))
t_obj_calc += time.time() - t_b_obj
if abs(curr_err_norm - next_err_norm) < errThresh:
break
curr_err_norm = next_err_norm
norm.append(curr_err_norm)
duration = time.time() - start_time - t_obj_calc
if ctf.comm().rank() == 0:
print('SGD amortized seconds per sweep: {}'.format(duration/(iteration_count*sample_rate)))
print("Time/SGD iteration: {}".format(duration/iteration_count))
return norm
def run_tests():
np.random.seed(5330)
wrld = ctf.comm()
if wrld.rank() != 0:
result = unittest.TextTestRunner(stream=open(os.devnull, 'w')).run(
unittest.TestSuite(
unittest.TestLoader().loadTestsFromTestCase(KnowValues)))
else:
print("Tests for linear algebra functionality")
result = unittest.TextTestRunner().run(
unittest.TestSuite(
unittest.TestLoader().loadTestsFromTestCase(KnowValues)))
return result
def getALS_CG(T, U, V, W, regParam, omega, I, J, K, r, block):
it = 0
E = ctf.tensor((I, J, K), sp=True)
E.i("ijk"
) << T.i("ijk") - omega.i("ijk") * U.i("iu") * V.i("ju") * W.i("ku")
NNZ = T.read_local_nnz()[1].shape[
0] # number of nonzero entries, i.e. sample size
curr_err_norm = (
ctf.vecnorm(E) +
(ctf.vecnorm(U) + ctf.vecnorm(V) + ctf.vecnorm(W)) * regParam) / NNZ
norm = [curr_err_norm]
timeList = [0]
t = time.time()
while True:
U = updateFactor_CG(T, U, V, W, regParam, omega, I, J, K, r, block,
"U")
V = updateFactor_CG(T, U, V, W, regParam, omega, I, J, K, r, block,
"V")
W = updateFactor_CG(T, U, V, W, regParam, omega, I, J, K, r, block,
"W")
E.set_zero()
E.i("ijk") << T.i(
"ijk") - omega.i("ijk") * U.i("iu") * V.i("ju") * W.i("ku")
next_err_norm = (ctf.vecnorm(E) +
(ctf.vecnorm(U) + ctf.vecnorm(V) + ctf.vecnorm(W)) *
regParam) / NNZ
if ctf.comm().rank() == 0:
print(curr_err_norm, next_err_norm)
if abs(curr_err_norm - next_err_norm) < .0001 or it > 100:
break
curr_err_norm = next_err_norm
norm.append(curr_err_norm)
timeList.append(np.round_(time.time() - t, 4))
it += 1
return norm, it, timeList
def getALS_CG(T, U, V, W, regParam, omega, I, J, K, r, block):
it = 0
E = ctf.tensor((I, J, K), sp=True)
#E.i("ijk") << T.i("ijk") - omega.i("ijk")*U.i("iu")*V.i("ju")*W.i("ku")
E.i("ijk") << T.i("ijk") - ctf.TTTP(omega, [U, V, W]).i("ijk")
assert (E.sp == 1)
curr_err_norm = ctf.vecnorm(E) + (ctf.vecnorm(U) + ctf.vecnorm(V) +
ctf.vecnorm(W)) * regParam
t = time.time()
while True:
U = updateFactor_CG(T, U, V, W, regParam, omega, I, J, K, r, block,
"U")
assert (U.sp == 1)
V = updateFactor_CG(T, U, V, W, regParam, omega, I, J, K, r, block,
"V")
assert (V.sp == 1)
W = updateFactor_CG(T, U, V, W, regParam, omega, I, J, K, r, block,
"W")
assert (W.sp == 1)
E.set_zero()
#E.i("ijk") << T.i("ijk") - omega.i("ijk")*U.i("iu")*V.i("ju")*W.i("ku")
E.i("ijk") << T.i("ijk") - ctf.TTTP(omega, [U, V, W]).i("ijk")
assert (E.sp == 1)
next_err_norm = ctf.vecnorm(E) + (ctf.vecnorm(U) + ctf.vecnorm(V) +
ctf.vecnorm(W)) * regParam
if ctf.comm().rank() == 0:
print(curr_err_norm, next_err_norm)
it += 1
if abs(curr_err_norm - next_err_norm) < .001 or it > 100:
break
curr_err_norm = next_err_norm
nt = np.round_(time.time() - t, 4)
return it, nt
def nproc(self):
return ctf.comm().np()
def is_master_proc():
if ctf.comm().rank() == 0:
return True
else:
return False
def printf(*string):
if ctf.comm().rank() == 0:
print(string)
def print_CG_time_summary():
if ctf.comm().rank() == 0:
print("total block CG contraction time %d" % blockCG_contraction_time)
import cProfile
# Get inputs
Nx = int(argv[1])
Ny = int(argv[2])
D = int(argv[3])
chi = int(argv[4])
Zn = int(argv[5])
if Zn == 0:
Zn = None
backend = argv[6]
d = 2
if backend == 'ctf':
import ctf
wrld = ctf.comm()
# TEBD Parameters
step_sizes = [0.1]
n_step = [5]
# Get Hamiltonian
if Zn is None:
ham = return_op(Nx, Ny, sym=None, backend=backend)
else:
ham = return_op(Nx, Ny, sym='Z2', backend=backend)
# Create PEPS
peps = PEPS(Nx,
Ny,
d,
a2 = ctf.astensor(numpy.ones((2,5)))
self.assertTrue(ctf.hstack((a1, a2)).shape == (2,9))
a1 = ctf.astensor(numpy.ones((2,4)))
a2 = ctf.astensor(numpy.ones((2,5))+0j)
self.assertTrue(ctf.hstack((a1, a2)).shape == (2,9))
self.assertTrue(ctf.hstack((a1, a2)).dtype == numpy.complex128)
a1 = numpy.ones((2,4))
a2 = ctf.astensor(numpy.ones((2,5))+0j)
self.assertTrue(ctf.hstack((a1, a2)).shape == (2,9))
na2 = numpy.ones((2,5))+0j
self.assertTrue(ctf.all(ctf.hstack((a1, a2)) == numpy.hstack((a1,na2))))
a1 = ctf.astensor(numpy.ones(4))
self.assertTrue(ctf.hstack((a1, 1.5)).shape == (5,))
a1 = ctf.astensor(numpy.ones((2,4,2)))
a2 = ctf.astensor(numpy.ones((2,5,2)))
self.assertTrue(ctf.hstack((a1, a2)).shape == (2,9,2))
if __name__ == "__main__":
if ctf.comm().rank() != 0:
result = unittest.TextTestRunner(stream = open(os.devnull, 'w')).run(unittest.TestSuite(unittest.TestLoader().loadTestsFromTestCase(KnowValues)))
else:
print("Tests for basic numpy ndarray functionality")
result = unittest.TextTestRunner().run(unittest.TestSuite(unittest.TestLoader().loadTestsFromTestCase(KnowValues)))
ctf.MPI_Stop()
sys.exit(not result)
def run_bench(num_iter, s_start, s_end, mult, R, sp, sp_init):
wrld = ctf.comm()
s = s_start
nnz = float(s_start * s_start * s_start) * sp_init
agg_s = []
agg_avg_times = []
agg_min_times = []
agg_max_times = []
agg_min_95 = []
agg_max_95 = []
if num_iter > 1:
if ctf.comm().rank() == 0:
print("Performing MTTKRP WARMUP with s =", s, "nnz =", nnz, "sp =",
sp, "sp_init =", sp_init)
T = ctf.tensor((s, s, s), sp=sp)
T.fill_sp_random(-1., 1., float(nnz) / float(s * s * s))
U = ctf.random.random((s, R))
V = ctf.random.random((s, R))
W = ctf.random.random((s, R))
U = ctf.einsum("ijk,jr,kr->ir", T, V, W)
V = ctf.einsum("ijk,ir,kr->jr", T, U, W)
W = ctf.einsum("ijk,ir,jr->kr", T, U, V)
if ctf.comm().rank() == 0:
print("Completed MTTKRP WARMUP with s =", s, "nnz =", nnz, "sp =",
sp, "sp_init =", sp_init)
while s <= s_end:
agg_s.append(s)
T = ctf.tensor((s, s, s), sp=sp)
T.fill_sp_random(-1., 1., float(nnz) / float(s * s * s))
if ctf.comm().rank() == 0:
print("Performing MTTKRP with s =", s, "nnz =", nnz, "sp =", sp,
"sp_init =", sp_init)
U = ctf.random.random((s, R))
V = ctf.random.random((s, R))
W = ctf.random.random((s, R))
te1 = 0.
te2 = 0.
te3 = 0.
avg_times = []
for i in range(num_iter):
t0 = time.time()
U = ctf.einsum("ijk,jr,kr->ir", T, V, W)
t1 = time.time()
ite1 = t1 - t0
te1 += ite1
t0 = time.time()
V = ctf.einsum("ijk,ir,kr->jr", T, U, W)
t1 = time.time()
ite2 = t1 - t0
te2 += ite2
t0 = time.time()
W = ctf.einsum("ijk,ir,jr->kr", T, U, V)
t1 = time.time()
ite3 = t1 - t0
te3 += ite3
if ctf.comm().rank() == 0:
print(ite1, ite2, ite3, "avg:", (ite1 + ite2 + ite3) / 3.)
avg_times.append((ite1 + ite2 + ite3) / 3.)
if ctf.comm().rank() == 0:
print("Completed", num_iter, "iterations, took", te1 / num_iter,
te2 / num_iter, te3 / num_iter,
"seconds on average for 3 variants.")
avg_time = (te1 + te2 + te3) / (3 * num_iter)
agg_avg_times.append(avg_time)
print("MTTKRP took", avg_times,
"seconds on average across variants with s =", s, "nnz =",
nnz, "sp =", sp, "sp_init =", sp_init)
min_time = np.min(avg_times)
max_time = np.max(avg_times)
agg_min_times.append(min_time)
agg_max_times.append(max_time)
print("min/max interval is [", min_time, ",", max_time, "]")
stddev = np.std(avg_times)
min_95 = (te1 + te2 + te3) / (3 * num_iter) - 2 * stddev
max_95 = (te1 + te2 + te3) / (3 * num_iter) + 2 * stddev
agg_min_95.append(min_95)
agg_max_95.append(max_95)
print("95% confidence interval is [", min_95, ",", max_95, "]")
s = int(s * mult)
if ctf.comm().rank() == 0:
print("s min_time min_95 avg_time max_95 max_time")
for i in range(len(agg_s)):
print(agg_s[i], agg_min_times[i], agg_min_95[i], agg_avg_times[i],
agg_max_95[i], agg_max_times[i])
def main():
#ut = UnitTests()
#ut.runAllTests()
#I = random.randint(6,6)
#J = random.randint(6,6)
#K = random.randint(6,6)
I = 1000
J = 1000
K = 1000
r = 2
sparsity = .000001
regParam = .1
block = 100
# 3rd-order tensor
T_SVD = ctf.tensor((I, J, K), sp=True)
T_SVD.fill_sp_random(0, 1, sparsity)
#T_SVD = function_tensor(I,J,K,sparsity)
assert (T_SVD.sp == 1)
#omega = updateOmega(T_SVD,I,J,K)
omega = getOmega(T_SVD)
assert (omega.sp == 1)
ctf.random.seed(42)
U_SVD = ctf.random.random((I, r), sp=True)
V_SVD = ctf.random.random((J, r), sp=True)
W_SVD = ctf.random.random((K, r), sp=True)
U_CG = ctf.copy(U_SVD)
V_CG = ctf.copy(V_SVD)
W_CG = ctf.copy(W_SVD)
T_CG = ctf.copy(T_SVD)
U_CG2 = ctf.copy(U_SVD)
V_CG2 = ctf.copy(V_SVD)
W_CG2 = ctf.copy(W_SVD)
T_CG2 = ctf.copy(T_SVD)
#t = time.time()
#getALS_SVD(T_SVD,U_SVD,V_SVD,W_SVD,regParam,omega,I,J,K,r)
#print("ALS SVD costs time = ",np.round_(time.time()- t,4))
if ctf.comm().rank() == 0:
print(
"--------------------------------ALS iterative CG------------------------"
)
blockCGit, blockCGtime = getALS_CG(T_CG, U_CG, V_CG, W_CG, regParam, omega,
I, J, K, r, block)
if ctf.comm().rank() == 0:
print("Number of iterations: %d" % (blockCGit))
print("CG block size: %d " % (block))
print("ALS iterative CG costs time: %f" % (blockCGtime))
if ctf.comm().rank() == 0:
print(
"--------------------------------ALS direct SVD------------------------"
)
kressnerit, kressnertime = getALS_Kressner(T_CG2, U_CG2, V_CG2, W_CG2,
regParam, omega, I, J, K, r)
if ctf.comm().rank() == 0:
print("Number of iterations: %d" % (kressnerit))
print("ALS direct CG costs time: %f" % (kressnertime))
def getALS_CG(T,
U,
V,
W,
regParam,
omega,
I,
J,
K,
r,
block_size,
num_iter=100,
err_thresh=.001,
time_limit=600,
use_implicit=True):
if use_implicit == True:
t_ALS_CG = ctf.timer_epoch("als_CG_implicit")
if ctf.comm().rank() == 0:
print(
"--------------------------------ALS with implicit CG------------------------"
)
else:
t_ALS_CG = ctf.timer_epoch("als_CG_explicit")
if ctf.comm().rank() == 0:
print(
"--------------------------------ALS with explicit CG------------------------"
)
if T.sp == True:
nnz_tot = T.nnz_tot
else:
nnz_tot = ctf.sum(omega)
t_ALS_CG.begin()
it = 0
if block_size <= 0:
block_size = max(I, J, K)
t_init_error_norm = ctf.timer("ALS_init_error_tensor_norm")
t_init_error_norm.start()
t0 = time.time()
E = ctf.tensor((I, J, K), sp=T.sp)
#E.i("ijk") << T.i("ijk") - omega.i("ijk")*U.i("iu")*V.i("ju")*W.i("ku")
E.i("ijk") << T.i("ijk") - ctf.TTTP(omega, [U, V, W]).i("ijk")
t1 = time.time()
curr_err_norm = ctf.vecnorm(E) + (ctf.vecnorm(U) + ctf.vecnorm(V) +
ctf.vecnorm(W)) * regParam
t2 = time.time()
t_init_error_norm.stop()
if ctf.comm().rank() == 0 and status_prints == True:
print('ctf.TTTP() takes {}'.format(t1 - t0))
print('ctf.vecnorm {}'.format(t2 - t1))
t_before_loop = time.time()
t_obj_calc = 0.
ctf.random.seed(42)
while True:
t_upd_cg = ctf.timer("ALS_upd_cg")
t_upd_cg.start()
U = updateFactor(T, U, V, W, regParam, omega, I, J, K, r, block_size,
"U", use_implicit)
V = updateFactor(T, U, V, W, regParam, omega, I, J, K, r, block_size,
"V", use_implicit)
W = updateFactor(T, U, V, W, regParam, omega, I, J, K, r, block_size,
"W", use_implicit)
duration = time.time() - t_before_loop - t_obj_calc
t_b_obj = time.time()
E.set_zero()
#E.i("ijk") << T.i("ijk") - omega.i("ijk")*U.i("iu")*V.i("ju")*W.i("ku")
E.i("ijk") << T.i("ijk") - ctf.TTTP(omega, [U, V, W]).i("ijk")
diff_norm = ctf.vecnorm(E)
RMSE = diff_norm / (nnz_tot**.5)
next_err_norm = diff_norm + (ctf.vecnorm(U) + ctf.vecnorm(V) +
ctf.vecnorm(W)) * regParam
t_obj_calc += time.time() - t_b_obj
t_upd_cg.stop()
it += 1
if ctf.comm().rank() == 0:
#print("Last residual:",curr_err_norm,"New residual",next_err_norm)
print('Objective after', duration, 'seconds (', it,
'iterations) is: {}'.format(next_err_norm))
print('RMSE after', duration, 'seconds (', it,
'iterations) is: {}'.format(RMSE))
if abs(curr_err_norm - next_err_norm
) < err_thresh or it >= num_iter or duration > time_limit:
break
curr_err_norm = next_err_norm
t_ALS_CG.end()
duration = time.time() - t_before_loop - t_obj_calc
if glob_comm.rank() == 0:
print('ALS (implicit =', use_implicit,
') time per sweep: {}'.format(duration / it))
def CG(A, b, x0, r, regParam, I, is_implicit=False):
t_batch_cg = ctf.timer("ALS_exp_cg")
t_batch_cg.start()
Ax0 = ctf.tensor((I, r))
if is_implicit:
Ax0.i("ir") << A.mul("ir", x0)
else:
Ax0.i("ir") << A.i("irl") * x0.i("il")
Ax0 += regParam * x0
rk = b - Ax0
sk = rk
xk = x0
for i in range(sk.shape[-1]): # how many iterations?
Ask = ctf.tensor((I, r))
t_cg_bmvec = ctf.timer("ALS_exp_cg_mvec")
t_cg_bmvec.start()
t0 = time.time()
if is_implicit:
Ask.i("ir") << A.mul("ir", sk)
else:
Ask.i("ir") << A.i("irl") * sk.i("il")
t1 = time.time()
if ctf.comm().rank == 0 and status_prints == True:
print('form Ask takes {}'.format(t1 - t0))
t_cg_bmvec.stop()
Ask += regParam * sk
rnorm = ctf.tensor(I)
rnorm.i("i") << rk.i("ir") * rk.i("ir")
skAsk = ctf.tensor(I)
skAsk.i("i") << sk.i("ir") * Ask.i("ir")
alpha = rnorm / (skAsk + 1.e-30)
alphask = ctf.tensor((I, r))
alphask.i("ir") << alpha.i("i") * sk.i("ir")
xk1 = xk + alphask
alphaask = ctf.tensor((I, r))
alphaask.i("ir") << alpha.i("i") * Ask.i("ir")
rk1 = rk - alphaask
rk1norm = ctf.tensor(I)
rk1norm.i("i") << rk1.i("ir") * rk1.i("ir")
beta = rk1norm / (rnorm + 1.e-30)
betask = ctf.tensor((I, r))
betask.i("ir") << beta.i("i") * sk.i("ir")
sk1 = rk1 + betask
rk = rk1
xk = xk1
sk = sk1
if ctf.vecnorm(rk) < CG_thresh:
break
#print("explicit CG residual after",sk.shape[-1],"iterations is",ctf.vecnorm(rk))
t_batch_cg.stop()
return xk
def train(model,
features,
adj,
y_train,
y_val,
train_mask,
val_mask,
lr=0.1,
epochs=200,
patience=0,
save_best=False):
begin = time.time()
val_loss_history = []
min_epoch = -1
computation_time = 0
for epoch in range(epochs):
# the backpropogation time
start = time.time()
model.backward(features, y_train, adj, train_mask, lr)
end = time.time()
train_loss = model.loss(adj, features, y_train, train_mask)
val_loss = model.loss(adj, features, y_val, val_mask)
train_accuracy = model.accuracy(adj, features, y_train, train_mask)
val_accuracy = model.accuracy(adj, features, y_val, val_mask)
if model.package == "ctf" and ctf.comm().rank() == 0:
print("Epoch:", '%04d' % (epoch + 1), "train_loss=",
"{:.5f}".format(train_loss), "train_acc=",
"{:.5f}".format(train_accuracy), "val_loss=",
"{:.5f}".format(val_loss), "val_acc=",
"{:.5f}".format(val_accuracy), "time=",
"{:.5f}".format(end - start))
elif model.package != "ctf":
print("Epoch:", '%04d' % (epoch + 1), "train_loss=",
"{:.5f}".format(train_loss), "train_acc=",
"{:.5f}".format(train_accuracy), "val_loss=",
"{:.5f}".format(val_loss), "val_acc=",
"{:.5f}".format(val_accuracy), "time=",
"{:.5f}".format(end - start))
computation_time += (end - start)
# if the validation loss is not decreasing in patience times,
if patience != 0:
if len(val_loss_history) == 0:
val_loss_history.append(val_loss)
continue
else:
min_epoch = np.argmin(np.array(val_loss_history))
min_loss = val_loss_history[min_epoch]
val_loss_history.append(val_loss)
if val_loss < min_loss:
min_loss = val_loss
min_epoch = epoch
if save_best == True:
model.save('params/best.pkl')
else:
if epoch - min_epoch >= patience:
print("Validation loss has not been improved for",
'%04d' % patience,
"epochs, reaching the specified patience.")
break
# total time
stop = time.time()
if model.package == "ctf" and ctf.comm().rank() == 0:
print("Total time: {:.4f}s, ".format(stop - begin),
"Computation time: {:.4f}s".format(computation_time))
elif model.package != "ctf":
print("Total time: {:.4f}s, ".format(stop - begin),
"Computation time: {:.4f}s".format(computation_time))
def updateFactor_CG(T, U, V, W, regParam, omega, I, J, K, r, block, string):
if (string == "U"):
#M1 = ctf.tensor((J,K,r))
#M1.i("jku") << V.i("ju")*W.i("ku")
#num_nonzero, dense_omega = getDenseOmega(T,U,V,W,regParam,omega,I,J,K,r,"i")
#Z = ctf.tensor((I,J*K,r))
#Z.i("itr") << dense_omega.i("jkti")*M1.i("jkr")
#Tbar = ctf.tensor((I,num_nonzero))
#Tbar.i("it") << dense_omega.i("ijkt") *T.i("ijk")
size = int(I / block)
for n in range(block):
t = time.time()
nomega = omega[n * size:(n + 1) * size, :, :]
if print_flag and ctf.comm().rank() == 0:
print("slicing omega cost %f seconds" %
(np.round_(time.time() - t, 4)))
#assert(nomega.sp == 1)
# ------------------ SPARSITY NOT PRESERVED IN THE ABOVE LINE ----------------#
x0 = ctf.random.random((size, r))
Ax0 = ctf.tensor((size, r), sp=True)
#Ax0.i("ir") << M.i("jkr")*dense_omega.i("jkti")*dense_omega.i("jktI")*M.i("jkR")*x0.i("IR")
Ax0.i("ir") << V.i("Jr") * W.i("Kr") * nomega.i("iJK") * V.i(
"JR") * W.i("KR") * x0.i(
"iR") # LHS; ATA using matrix-vector multiplication
Ax0 += regParam * x0
if print_flag and ctf.comm().rank() == 0:
#blockCG_contraction_time += time.time()- t
print("contraction to form LHS cost %f seconds" %
(np.round_(time.time() - t, 4)))
assert (Ax0.sp == 1)
b = ctf.tensor((size, r), sp=True)
#b.i("ir") << M.i("JKr") * dense_omega.i("JKti") * dense_omega.i("JKtI") * T.i("IJK")
b.i("ir") << V.i("Jr") * W.i("Kr") * T[n * size:
(n + 1) * size, :, :].i(
"iJK") # RHS; ATb
if print_flag and ctf.comm().rank() == 0:
#blockCG_contraction_time += time.time()- t
print("contraction to form RHS cost %f seconds" %
(np.round_(time.time() - t, 4)))
assert (b.sp == 1)
U[n * size:(n + 1) * size, :].set_zero()
U[n * size:(n + 1) * size, :] = CG(Ax0, b, x0, V, W, r, regParam,
nomega, size, "U")
assert (U.sp == 1)
return U
if (string == "V"):
#M2 = ctf.tensor((I,K,r))
#M2.i("iku") << U.i("iu")*W.i("ku")
#num_nonzero, dense_omega = getDenseOmega(T,U,V,W,regParam,omega,I,J,K,r)
#Z = ctf.tensor((J,num_nonzero,r))
#Z.i("jtr") << dense_omega.i("ijkt")*M2.i("ikr")
#Tbar = ctf.tensor((J,num_nonzero))
#Tbar.i("jt") << dense_omega.i("ijkt") *T.i("ijk")
size = int(J / block)
for n in range(block):
nomega = omega[:, n * size:(n + 1) * size, :]
x0 = ctf.random.random((size, r))
Ax0 = ctf.tensor((size, r), sp=True)
Ax0.i("jr") << U.i("Ir") * W.i("Kr") * nomega.i("IjK") * U.i(
"IR") * W.i("KR") * x0.i(
"jR") # LHS; ATA using matrix-vector multiplication
Ax0 += regParam * x0
assert (Ax0.sp == 1)
b = ctf.tensor((size, r), sp=True)
b.i("jr") << U.i("Ir") * W.i(
"Kr") * T[:, n * size:(n + 1) * size, :].i("IjK") # RHS; ATb
assert (b.sp == 1)
V[n * size:(n + 1) * size, :].set_zero()
V[n * size:(n + 1) * size, :] = CG(Ax0, b, x0, U, W, r, regParam,
nomega, size, "V")
assert (V.sp == 1)
return V
if (string == "W"):
#M3 = ctf.tensor((I,J,r))
#M3.i("iju") << U.i("iu")*V.i("ju")
#num_nonzero, dense_omega = getDenseOmega(T,U,V,W,regParam,omega,I,J,K,r)
#Z = ctf.tensor((K,num_nonzero,r))
#Z.i("ktr") << dense_omega.i("ijkt")*M3.i("ijr")
#Tbar = ctf.tensor((K,num_nonzero))
#Tbar.i("kt") << dense_omega.i("ijkt") *T.i("ijk")
size = int(K / block)
for n in range(block):
nomega = omega[:, :, n * size:(n + 1) * size]
x0 = ctf.random.random((size, r))
Ax0 = ctf.tensor((size, r), sp=True)
Ax0.i("kr") << U.i("Ir") * V.i("Jr") * nomega.i("IJk") * U.i(
"IR") * V.i("JR") * x0.i(
"kR") # LHS; ATA using matrix-vector multiplication
Ax0 += regParam * x0
assert (Ax0.sp == 1)
b = ctf.tensor((size, r), sp=True)
b.i("kr") << U.i("Ir") * V.i("Jr") * T[:, :, n * size:(n + 1) *
size].i("IJk") # RHS; ATb
assert (b.sp == 1)
W[n * size:(n + 1) * size, :].set_zero()
W[n * size:(n + 1) * size, :] = CG(Ax0, b, x0, U, V, r, regParam,
nomega, size, "W")
assert (W.sp == 1)
return W
self.ovov = None
self.foo = None
self.fvv = None
self.fov = None
if __name__ == '__main__':
assert (len(sys.argv) <= 4)
assert (len(sys.argv) >= 3)
nocc = int(sys.argv[1])
nvir = int(sys.argv[2])
cutoff = None
if (len(sys.argv) > 3):
cutoff = float(sys.argv[3])
cm = ctf.comm()
eris = integrals()
NS = ctf.SYM.NS
SY = ctf.SYM.SY
eris.oovv = ctf.tensor([nocc, nocc, nvir, nvir], sym=[NS, NS, SY, NS])
eris.ovvv = ctf.tensor([nocc, nvir, nvir, nvir], sym=[NS, NS, SY, NS])
eris.vvvv = ctf.tensor([nvir, nvir, nvir, nvir], sym=[SY, NS, SY, NS])
eris.ovov = ctf.tensor([nocc, nvir, nocc, nvir], sym=[NS, NS, NS, NS])
eris.ooov = ctf.tensor([nocc, nocc, nocc, nvir], sym=[NS, NS, NS, NS])
eris.oooo = ctf.tensor([nocc, nocc, nocc, nocc], sym=[NS, NS, NS, NS])
eris.fvv = ctf.tensor([nvir, nvir])
eris.fov = ctf.tensor([nocc, nvir])
eris.foo = ctf.tensor([nocc, nocc])
for e in [
eris.ovvv, eris.oovv, eris.oooo, eris.ooov, eris.vvvv, eris.ovov,
#!/usr/bin/env python
import ctf
from ctf import random
A = ctf.random.random((32,32))
[U,S,VT]=ctf.svd(A)
err = A-ctf.dot(U,ctf.dot(ctf.diag(S),VT))
success=True
err_nrm = err.norm2()
if err_nrm > 1.E-6:
success=False
if ctf.comm().rank() == 0:
if success:
print("success, norm is ", err_nrm)
else:
print("failure, norm is ", err_nrm)
ctf.MPI_Stop()
def rank(self):
return ctf.comm().rank()
print("95% confidence interval for fast is [", avg_fast - 2 * stddev_fast,
",", avg_fast + 2 * stddev_fast, "]")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--n',
type=int,
default=4,
metavar='int',
help='Dimension of symmetric modes (default: 4)')
parser.add_argument('--b',
type=int,
default=10,
metavar='int',
help='Dimension of nonsymmetric modes (default: 10)')
parser.add_argument(
'--niter',
type=int,
default=10,
metavar='int',
help='Number of iterations for benchmarking (default: 10)')
args, _ = parser.parse_known_args()
n = args.n
b = args.b
niter = args.niter
w = ctf.comm()
test(n, b)
bench(n, b, niter)
from ctf import random as crandom
import gzip
import shutil
import os
import argparse
import arg_defs as arg_defs
sys.path.insert(0, '../SGD')
from gradient1 import sparse_SGD
sys.path.insert(0, '../ALS')
from als_sp import getALS_CG
sys.path.insert(0, '../CCD')
from ccd_sp import run_CCD
from ccd_sp import get_objective
glob_comm = ctf.comm()
def getOmega(T):
[inds, data] = T.read_local_nnz()
data[:] = 1.
Omega = ctf.tensor(T.shape, sp=T.sp)
Omega.write(inds, data)
return Omega
def read_frostt_tensor(file_name, I, J, K, use_sp_rep):
unzipped_file_name = file_name + '.tns'
exists = os.path.isfile(unzipped_file_name)
if not exists:
def main():
#ut = UnitTests()
#ut.runAllTests()
#I = random.randint(6,6)
#J = random.randint(6,6)
#K = random.randint(6,6)
I = 8
J = 8
K = 8
r = 2
sparsity = .1
regParam = .1
block = 4
ntrails = 5
# 3rd-order tensor
#T_SVD = ctf.tensor((I,J,K),sp=True)
#T_SVD.fill_sp_random(0,1,sparsity)
T_SVD = function_tensor(I, J, K, sparsity)
assert (T_SVD.sp == 1)
omega = updateOmega(T_SVD, I, J, K)
assert (omega.sp == 1)
blockCGerrList = []
blockCGtimeList = []
KressnererrList = []
KressnertimeList = []
for i in range(ntrails):
ctf.random.seed(42 + i)
U_SVD = ctf.random.random((I, r), sp=True)
V_SVD = ctf.random.random((J, r), sp=True)
W_SVD = ctf.random.random((K, r), sp=True)
U_CG = ctf.copy(U_SVD)
V_CG = ctf.copy(V_SVD)
W_CG = ctf.copy(W_SVD)
T_CG = ctf.copy(T_SVD)
U_CG2 = ctf.copy(U_SVD)
V_CG2 = ctf.copy(V_SVD)
W_CG2 = ctf.copy(W_SVD)
T_CG2 = ctf.copy(T_SVD)
#t = time.time()
#getALS_SVD(T_SVD,U_SVD,V_SVD,W_SVD,regParam,omega,I,J,K,r)
#print("ALS SVD costs time = ",np.round_(time.time()- t,4))
t = time.time()
blockCGnorm, blockCGit, blockCGtime = getALS_CG(
T_CG, U_CG, V_CG, W_CG, regParam, omega, I, J, K, r, block)
blockCGerrList.append(blockCGnorm)
blockCGtimeList.append(blockCGtime)
if ctf.comm().rank() == 0:
print("Number of iterations: %d" % (blockCGit))
print("CG block size: %d " % (block))
print("ALS iterative CG costs time: %f" %
(np.round_(time.time() - t, 4)))
t = time.time()
kressnernorm, kressnerit, kressnertime = getALS_Kressner(
T_CG2, U_CG2, V_CG2, W_CG2, regParam, omega, I, J, K, r)
KressnererrList.append(kressnernorm)
KressnertimeList.append(kressnertime)
if ctf.comm().rank() == 0:
print("Number of iterations: %d" % (kressnerit))
print("ALS direct CG costs time: %f" %
(np.round_(time.time() - t, 4)))
#----------------------------------------- plot -------------------------------------------------#
plt.figure()
for i in range(ntrails):
if ctf.comm().rank() == 0:
plt.plot(blockCGtimeList[i],
blockCGerrList[i],
label="trail %d" % (i + 1))
plt.legend()
plt.title(
"Function tensor(%d*%d*%d), iterative block CG, block size %d, rank %d, sparsity %f"
% (I, J, K, block, r, sparsity))
plt.xlabel("Time[s]")
plt.ylabel("Training Error Norm")
plt.savefig('iterative_block_CG.png')
plt.figure()
for i in range(ntrails):
if ctf.comm().rank() == 0:
plt.plot(KressnertimeList[i],
KressnererrList[i],
label="trail %d" % (i + 1))
plt.legend()
plt.title(
"Funtion tensor(%d*%d*%d),direct CG (Kressner), rank %d, sparsity %f "
% (I, J, K, r, sparsity))
plt.xlabel("Time[s]")
plt.ylabel("Training Error Norm")
plt.savefig('direct_CG.png')
max_95 = (te1 + te2 + te3) / (3 * num_iter) + 2 * stddev
agg_min_95.append(min_95)
agg_max_95.append(max_95)
print("95% confidence interval is [", min_95, ",", max_95, "]")
s = int(s * mult)
if ctf.comm().rank() == 0:
print("s min_time min_95 avg_time max_95 max_time")
for i in range(len(agg_s)):
print(agg_s[i], agg_min_times[i], agg_min_95[i], agg_avg_times[i],
agg_max_95[i], agg_max_times[i])
if __name__ == "__main__":
parser = argparse.ArgumentParser()
sargs.add_arguments(parser)
args, _ = parser.parse_known_args()
num_iter = args.num_iter
s_start = args.s_start
s_end = args.s_end
mult = args.mult
R = args.R
sp = args.sp
sp_init = args.sp_init
if ctf.comm().rank() == 0:
print("num_iter is", num_iter, "s_start is", s_start, "s_end is",
s_end, "mult is", mult, "R is", R, "sp is", sp, "sp_init is",
sp_init)
run_bench(num_iter, s_start, s_end, mult, R, sp, sp_init)
import ctf,time,random
import numpy as np
import numpy.linalg as la
from ctf import random as crandom
glob_comm = ctf.comm()
from scipy.sparse.linalg import lsqr as lsqr
# In[6]:
class UnitTests:
def test_3d_purturb1(self):
I = random.randint(3,5)
J = random.randint(3,5)
K = random.randint(3,5)
r = 2
sparsity = .2
regParam = 10
ctf.random.seed(42)
U = ctf.random.random((I,r))
V= ctf.random.random((J,r))
W= ctf.random.random((K,r))
# 3rd-order tensor
T = ctf.tensor((I,J,K))
T.fill_random(0,1)
self.assertTrue(ctf.hstack((a1, a2)).shape == (2, 9))
self.assertTrue(ctf.hstack((a1, a2)).dtype == numpy.complex128)
a1 = numpy.ones((2, 4))
a2 = ctf.astensor(numpy.ones((2, 5)) + 0j)
self.assertTrue(ctf.hstack((a1, a2)).shape == (2, 9))
na2 = numpy.ones((2, 5)) + 0j
self.assertTrue(
ctf.all(ctf.hstack((a1, a2)) == numpy.hstack((a1, na2))))
a1 = ctf.astensor(numpy.ones(4))
self.assertTrue(ctf.hstack((a1, 1.5)).shape == (5, ))
a1 = ctf.astensor(numpy.ones((2, 4, 2)))
a2 = ctf.astensor(numpy.ones((2, 5, 2)))
self.assertTrue(ctf.hstack((a1, a2)).shape == (2, 9, 2))
if __name__ == "__main__":
if ctf.comm().rank() != 0:
result = unittest.TextTestRunner(stream=open(os.devnull, 'w')).run(
unittest.TestSuite(
unittest.TestLoader().loadTestsFromTestCase(KnowValues)))
else:
print("Tests for basic numpy ndarray functionality")
result = unittest.TextTestRunner().run(
unittest.TestSuite(
unittest.TestLoader().loadTestsFromTestCase(KnowValues)))
ctf.MPI_Stop()
sys.exit(not result)
def updateFactor_Kressner(T, U, V, W, regParam, omega, I, J, K, r, string):
if (string == "U"):
#M1 = ctf.tensor((J,K,r))
#M1.i("jku") << V.i("ju")*W.i("ku")
for i in range(I):
t = time.time()
A = ctf.tensor((r, r), sp=True)
sliced_omega = omega[i, :, :]
if print_flag and ctf.comm().rank() == 0:
print("slicing omega cost %f seconds" %
(np.round_(time.time() - t, 4)))
A.i("uv") << V.i("Ju") * W.i("Ku") * sliced_omega.i("JK") * V.i(
"Jv") * W.i("Kv")
if print_flag and ctf.comm().rank() == 0:
print("contraction to form LHS cost %f seconds" %
(np.round_(time.time() - t, 4)))
assert (A.sp == 1)
#assert(omega[i,:,:].sp==1) TODO!
b = ctf.tensor(r, sp=True)
sliced_T = T[i, :, :]
if print_flag and ctf.comm().rank() == 0:
print("slicing original tensor cost %f seconds" %
(np.round_(time.time() - t, 4)))
b.i("r") << V.i("Jr") * W.i("Kr") * sliced_T.i("JK") # RHS; ATb
if print_flag and ctf.comm().rank() == 0:
print("contraction to form RHS cost %f seconds" %
(np.round_(time.time() - t, 4)))
assert (b.sp == 1)
U[i, :].set_zero()
U[i, :] = Kressner(A, b, U[i, :], r, regParam)
assert (U.sp == 1)
return U
if (string == "V"):
#M2 = ctf.tensor((I,K,r))
#M2.i("iku") << U.i("iu")*W.i("ku")
for j in range(J):
A = ctf.tensor((r, r), sp=True)
A.i("uv") << U.i("Iu") * W.i("Ku") * omega[:, j, :].i("IK") * U.i(
"Iv") * W.i("Kv")
assert (A.sp == 1)
b = ctf.tensor(r, sp=True)
b.i("r") << U.i("Ir") * W.i("Kr") * T[:, j, :].i("IK") # RHS; ATb
assert (b.sp == 1)
V[j, :].set_zero()
V[j, :] = Kressner(A, b, V[j, :], r, regParam)
assert (V.sp == 1)
return V
if (string == "W"):
#M3 = ctf.tensor((I,J,r))
#M3.i("iju") << U.i("iu")*V.i("ju")
for k in range(K):
A = ctf.tensor((r, r), sp=True)
A.i("uv") << U.i(
"Iu") * V.i("Ju") * omega[:, :, k].i("IJ") * U.i("Iv") * V.i(
"Jv") # LHS; ATA using matrix-vector multiplication
assert (A.sp == 1)
b = ctf.tensor(r, sp=True)
b.i("r") << U.i("Ir") * V.i("Jr") * T[:, :, k].i("IJ") # RHS; ATb
assert (b.sp == 1)
W[k, :].set_zero()
W[k, :] = Kressner(A, b, W[k, :], r, regParam)
assert (W.sp == 1)
return W