def test_performance(self):
c = random.rand(2000, 2000)
x = Variable([2000, 2000])
K = np.abs(random.rand(9, 9))
G = CompGraph(
vstack([subsample((conv_nofft(K, x) - c) * 5, [2, 4]), x * 10]))
xtest1 = random.rand(2000 * 2000).astype(np.float32)
ytest1 = np.zeros(G.output_size, dtype=np.float32)
t1_cpu = time.time()
for i in range(10):
ytest1 = G.forward(xtest1, ytest1)
t2_cpu = time.time()
xtest = gpuarray.to_gpu(xtest1.astype(np.float32))
ytest = gpuarray.to_gpu(ytest1.astype(np.float32))
t1_gpu = time.time()
for i in range(10):
ytest = G.forward_cuda(xtest, ytest)
t2_gpu = time.time()
t_cpu = t2_cpu - t1_cpu
t_gpu = t2_gpu - t1_gpu
logging.info("Forward timing: cpu=%.2f ms gpu=%.2f ms factor=%.3f" %
(t_cpu, t_gpu, t_gpu / t_cpu))
self.assertTrue(t_gpu < t_cpu)
t1_cpu = time.time()
for i in range(10):
xtest1 = G.adjoint(ytest1, xtest1)
t2_cpu = time.time()
t1_gpu = time.time()
for i in range(10):
xtest = G.adjoint_cuda(ytest, xtest)
t2_gpu = time.time()
t_cpu = t2_cpu - t1_cpu
t_gpu = t2_gpu - t1_gpu
logging.info("Adjoint timing: cpu=%.2f ms gpu=%.2f ms factor=%.3f" %
(t_cpu, t_gpu, t_gpu / t_cpu))
self.assertTrue(t_gpu < t_cpu)
def test_performance(self):
c = random.rand(2000,2000)
x = Variable([2000,2000])
K = np.abs(random.rand(9,9))
G = CompGraph(vstack([ subsample((conv_nofft(K, x) -c)*5, [2,4]), x*10 ]))
xtest1 = random.rand(2000*2000).astype(np.float32)
ytest1 = np.zeros(G.output_size, dtype=np.float32)
t1_cpu = time.time()
for i in range(10):
ytest1 = G.forward(xtest1, ytest1)
t2_cpu = time.time()
xtest = gpuarray.to_gpu(xtest1.astype(np.float32))
ytest = gpuarray.to_gpu(ytest1.astype(np.float32))
t1_gpu = time.time()
for i in range(10):
ytest = G.forward_cuda(xtest, ytest)
t2_gpu = time.time()
t_cpu = t2_cpu - t1_cpu
t_gpu = t2_gpu - t1_gpu
logging.info("Forward timing: cpu=%.2f ms gpu=%.2f ms factor=%.3f" % (t_cpu, t_gpu, t_gpu/t_cpu))
self.assertTrue(t_gpu < t_cpu)
t1_cpu = time.time()
for i in range(10):
xtest1 = G.adjoint(ytest1, xtest1)
t2_cpu = time.time()
t1_gpu = time.time()
for i in range(10):
xtest = G.adjoint_cuda(ytest, xtest)
t2_gpu = time.time()
t_cpu = t2_cpu - t1_cpu
t_gpu = t2_gpu - t1_gpu
logging.info("Adjoint timing: cpu=%.2f ms gpu=%.2f ms factor=%.3f" % (t_cpu, t_gpu, t_gpu/t_cpu))
self.assertTrue(t_gpu < t_cpu)
def _generic_check_adjoint(self, f, inshape, outshape, s,
ntests = 50, eps=1e-5, verbose=False,
in_out_sample = None):
"""
Generic tests used for all comp graph tests on a parametrizable function f
"""
x = Variable(inshape)
func = f(x)
if not type(func) is tuple:
func = (func,)
G = CompGraph(vstack(func))
nin = functools.reduce(lambda x,y: x*y, inshape, 1)
nout = functools.reduce(lambda x,y: x*y, outshape, 1)
if not in_out_sample is None:
# check against the given in/out samples
x1 = in_out_sample[0] # forward in
y1s = in_out_sample[1] # forward out
y2 = in_out_sample[2] # adjoint in
x2s = in_out_sample[3] # adjoint out
y1a = G.forward_cuda(gpuarray.to_gpu(x1.astype(np.float32)),gpuarray.to_gpu(y1s.astype(np.float32))).get()
#print(y1s)
#print(y1a)
self.assertTrue(np.amax(np.abs(y1a-y1s)) < eps)
x2a = G.adjoint_cuda(gpuarray.to_gpu(y2.astype(np.float32)),gpuarray.to_gpu(x2s.astype(np.float32))).get()
self.assertTrue(np.amax(np.abs(x2a-x2s)) < eps)
# test with random data that the forward/adjoint operators are consistent
maxerr = 0.0
random.seed(0) # make tests reproducable
for tidx in range(ntests):
x1 = random.rand(nin).astype(np.float32)
y2 = random.rand(nout).astype(np.float32)
y1 = np.zeros(nout, dtype=np.float32)
x2 = np.zeros(nin, dtype=np.float32)
if verbose: print("forward: ", end="")
y1 = G.forward_cuda(gpuarray.to_gpu(x1),gpuarray.to_gpu(y1),printt=verbose).get()
if verbose: print("adjoint: ", end="")
x2 = G.adjoint_cuda(gpuarray.to_gpu(y2),gpuarray.to_gpu(x2),printt=verbose).get()
self.assertTrue(not np.all(y1 == 0) and not np.all(x2 == 0))
y1o = G.forward(x1,y1.copy())
x2o = G.adjoint(y2,x2.copy())
erro = abs(np.dot(x1.flatten().astype(np.float64), x2o.flatten().astype(np.float64)) -
np.dot(y1o.flatten().astype(np.float64), y2.flatten().astype(np.float64)))
err = abs(np.dot(x1.flatten().astype(np.float64),x2.flatten().astype(np.float64)) -
np.dot(y1.flatten().astype(np.float64),y2.flatten().astype(np.float64)))
if err > maxerr:
maxerr = err
if verbose and err > eps:
print("forward CUDA code:")
print(G.cuda_forward_subgraphs.cuda_code)
print("backward CUDA code:")
print(G.cuda_adjoint_subgraphs.cuda_code)
print("x1\n",np.reshape(x1, inshape))
print("y1\n",np.reshape(y1, outshape))
print("y1o\n",np.reshape(y1o, outshape))
print("y2\n",np.reshape(y2, outshape))
print("x2\n",np.reshape(x2, inshape))
print("x2o\n",np.reshape(x2o, inshape))
print("(%d) Adjoint test (%s): gpu: %f, nogpu: %f" %(tidx,s,err,erro))
print("max(abs(y1-y1o)): %f" % (np.amax(np.abs(y1-y1o))))
print("max(abs(x2-x2o)): %f" % (np.amax(np.abs(x2-x2o))))
self.assertTrue(err <= eps)
if verbose: print("%s passed %d tests. Max adjoint test error: %f" % (s, ntests, maxerr))
def _test_pock_chambolle(self, impl):
"""Test pock chambolle algorithm.
"""
#print()
#print("----------------------",impl,"-------------------------")
if impl == 'pycuda':
kw = {'adapter': PyCudaAdapter()}
else:
kw = {}
X = px.Variable((10, 5))
B = np.reshape(np.arange(50), (10, 5))
prox_fns = [px.sum_squares(X, b=B)]
sltn = pc.solve(prox_fns, [],
1.0,
1.0,
1.0,
eps_rel=1e-5,
eps_abs=1e-5,
**kw)
self.assertItemsAlmostEqual(X.value, B, eps=2e-2)
self.assertAlmostEqual(sltn, 0)
prox_fns = [px.norm1(X, b=B, beta=2)]
sltn = pc.solve(prox_fns, [],
1.0,
1.0,
1.0,
eps_rel=1e-5,
eps_abs=1e-5,
**kw)
self.assertItemsAlmostEqual(X.value, B / 2., eps=2e-2)
self.assertAlmostEqual(sltn, 0, eps=2e-2)
prox_fns = [px.norm1(X), px.sum_squares(X, b=B)]
#print("----------------------------------------------------")
#print("----------------------------------------------------")
sltn = pc.solve(prox_fns, [],
0.5,
1.0,
1.0,
eps_rel=1e-5,
eps_abs=1e-5,
conv_check=1,
**kw)
cvx_X = cvx.Variable((10, 5))
cost = cvx.sum_squares(cvx_X - B) + cvx.norm(cvx_X, 1)
prob = cvx.Problem(cvx.Minimize(cost))
prob.solve()
self.assertItemsAlmostEqual(X.value, cvx_X.value, eps=2e-2)
self.assertAlmostEqual(sltn, prob.value)
psi_fns, omega_fns = pc.partition(prox_fns)
sltn = pc.solve(psi_fns,
omega_fns,
0.5,
1.0,
1.0,
eps_abs=1e-5,
eps_rel=1e-5,
**kw)
self.assertItemsAlmostEqual(X.value, cvx_X.value, eps=2e-2)
self.assertAlmostEqual(sltn, prob.value)
# With linear operators.
kernel = np.array([1, 2, 3])
kernel_mat = np.matrix("2 1 3; 3 2 1; 1 3 2")
x = px.Variable(3)
b = np.array([-41, 413, 2])
prox_fns = [px.nonneg(x), px.sum_squares(px.conv(kernel, x), b=b)]
sltn = pc.solve(prox_fns, [],
0.1,
0.1,
1.0,
max_iters=3000,
eps_abs=1e-5,
eps_rel=1e-5,
**kw)
cvx_X = cvx.Variable(3)
cost = cvx.norm(kernel_mat * cvx_X - b)
prob = cvx.Problem(cvx.Minimize(cost), [cvx_X >= 0])
prob.solve()
self.assertItemsAlmostEqual(x.value, cvx_X.value, eps=2e-2)
psi_fns, omega_fns = pc.partition(prox_fns)
sltn = pc.solve(psi_fns,
omega_fns,
0.1,
0.1,
1.0,
max_iters=3000,
eps_abs=1e-5,
eps_rel=1e-5,
**kw)
self.assertItemsAlmostEqual(x.value, cvx_X.value, eps=2e-2)
# TODO
# Multiple variables.
x = px.Variable(1)
y = px.Variable(1)
prox_fns = [
px.nonneg(x),
px.sum_squares(vstack([x, y]), b=np.arange(2))
]
sltn = pc.solve(prox_fns, [prox_fns[-1]],
0.1,
0.1,
1.0,
max_iters=3000,
eps_abs=1e-5,
eps_rel=1e-5,
try_diagonalize=False)
self.assertItemsAlmostEqual(x.value, [0])
self.assertItemsAlmostEqual(y.value, [1])
sltn = pc.solve(prox_fns, [prox_fns[-1]],
0.1,
0.1,
1.0,
max_iters=3000,
eps_abs=1e-5,
eps_rel=1e-5,
try_diagonalize=True)
self.assertItemsAlmostEqual(x.value, [0])
self.assertItemsAlmostEqual(y.value, [1])
def _generic_check_adjoint(self,
f,
inshape,
outshape,
s,
ntests=50,
eps=1e-5,
verbose=False,
in_out_sample=None):
"""
Generic tests used for all comp graph tests on a parametrizable function f
"""
x = Variable(inshape)
func = f(x)
if not type(func) is tuple:
func = (func, )
G = CompGraph(vstack(func))
nin = functools.reduce(lambda x, y: x * y, inshape, 1)
nout = functools.reduce(lambda x, y: x * y, outshape, 1)
if not in_out_sample is None:
# check against the given in/out samples
x1 = in_out_sample[0] # forward in
y1s = in_out_sample[1] # forward out
y2 = in_out_sample[2] # adjoint in
x2s = in_out_sample[3] # adjoint out
y1a = G.forward_cuda(gpuarray.to_gpu(x1.astype(np.float32)),
gpuarray.to_gpu(y1s.astype(
np.float32))).get()
#print(y1s)
#print(y1a)
self.assertTrue(np.amax(np.abs(y1a - y1s)) < eps)
x2a = G.adjoint_cuda(gpuarray.to_gpu(y2.astype(np.float32)),
gpuarray.to_gpu(x2s.astype(
np.float32))).get()
self.assertTrue(np.amax(np.abs(x2a - x2s)) < eps)
# test with random data that the forward/adjoint operators are consistent
maxerr = 0.0
random.seed(0) # make tests reproducable
for tidx in range(ntests):
x1 = random.rand(nin).astype(np.float32)
y2 = random.rand(nout).astype(np.float32)
y1 = np.zeros(nout, dtype=np.float32)
x2 = np.zeros(nin, dtype=np.float32)
if verbose:
print("forward: ", end="")
y1 = G.forward_cuda(gpuarray.to_gpu(x1),
gpuarray.to_gpu(y1),
printt=verbose).get()
if verbose:
print("adjoint: ", end="")
x2 = G.adjoint_cuda(gpuarray.to_gpu(y2),
gpuarray.to_gpu(x2),
printt=verbose).get()
self.assertTrue(not np.all(y1 == 0) and not np.all(x2 == 0))
y1o = G.forward(x1, y1.copy())
x2o = G.adjoint(y2, x2.copy())
erro = abs(
np.dot(x1.flatten().astype(np.float64),
x2o.flatten().astype(np.float64)) -
np.dot(y1o.flatten().astype(np.float64),
y2.flatten().astype(np.float64)))
err = abs(
np.dot(x1.flatten().astype(np.float64),
x2.flatten().astype(np.float64)) -
np.dot(y1.flatten().astype(np.float64),
y2.flatten().astype(np.float64)))
if err > maxerr:
maxerr = err
if verbose and err > eps:
print("forward CUDA code:")
print(G.cuda_forward_subgraphs.cuda_code)
print("backward CUDA code:")
print(G.cuda_adjoint_subgraphs.cuda_code)
print("x1\n", np.reshape(x1, inshape))
print("y1\n", np.reshape(y1, outshape))
print("y1o\n", np.reshape(y1o, outshape))
print("y2\n", np.reshape(y2, outshape))
print("x2\n", np.reshape(x2, inshape))
print("x2o\n", np.reshape(x2o, inshape))
print("(%d) Adjoint test (%s): gpu: %f, nogpu: %f" %
(tidx, s, err, erro))
print("max(abs(y1-y1o)): %f" % (np.amax(np.abs(y1 - y1o))))
print("max(abs(x2-x2o)): %f" % (np.amax(np.abs(x2 - x2o))))
self.assertTrue(err <= eps)
if verbose:
print("%s passed %d tests. Max adjoint test error: %f" %
(s, ntests, maxerr))