"gnumpy = chain_rules(a, x, y, grad_np_kernel(x, y, sigma, kernel=k), b)",
globals=globals(),
repeat=3,
number=1,
)
print("Time for NumPy: {:.4f}s".format(
np.median(speed_numpy[k])))
else:
gnumpy = torch.zeros_like(xc).data.cpu().numpy()
# Vanilla pytorch (with cuda if available, and cpu otherwise)
if use_vanilla:
try:
aKxy_b = torch.dot(
ac.view(-1),
(torch_kernel(xc, yc, sigmac, kernel=k) @ bc).view(-1))
g3 = torch.autograd.grad(aKxy_b, xc, create_graph=False)[0].cpu()
torch.cuda.synchronize()
speed_pytorch[k] = np.array(
timeit.repeat(
setup=
"cost = torch.dot(ac.view(-1), (torch_kernel(xc, yc, sigmac, kernel=k) @ bc).view(-1))",
stmt=
"g3 = torch.autograd.grad(cost, xc, create_graph=False)[0] ; torch.cuda.synchronize()",
globals=globals(),
repeat=REPEAT,
number=1,
))
print(
"Time for PyTorch: {:.4f}s".format(
np.median(speed_pytorch[k])),
for k in kernel_to_test:
print('kernel: ' + k)
# Pure numpy
g_numpy = np.matmul(np_kernel(x, y, sigma, kernel=k), b)
speed_numpy[k] = timeit.repeat(
'gnumpy = np.matmul( np_kernel(x, y, sigma, kernel=k), b)',
globals=globals(),
repeat=5,
number=1)
print('Time for NumPy: {:.4f}s'.format(
np.median(speed_numpy[k])))
# Vanilla pytorch (with cuda if available, and cpu otherwise)
try:
g_pytorch = torch_kernel(xc, yc, sigmac, kernel=k) @ bc
torch.cuda.synchronize()
speed_pytorch[k] = np.array(
timeit.repeat(
"torch_kernel(xc, yc, sigmac, kernel=k) @ bc; torch.cuda.synchronize()",
globals=globals(),
repeat=REPEAT,
number=4)) / 4
print('Time for PyTorch: {:.4f}s'.format(
np.median(speed_pytorch[k])),
end='')
print(' (absolute error: ',
np.max(np.abs(g_pytorch.cpu().numpy() - g_numpy)), ')')
except:
print('Time for PyTorch: Not Done')
speed_pykeops_gen = timeit.Timer(
'g1 = kernel_product( params,xc,yc,bc, mode=mode).cpu()',
GC,
globals=globals(),
timer=time.time).timeit(LOOPS)
print(
"Time for keops generic: {:.4f}s".format(speed_pykeops_gen),
end="")
print(" (absolute error: ",
np.max(np.abs(g1.data.numpy() - gnumpy)), ")")
except:
pass
# vanilla pytorch (with cuda if available else uses cpu)
try:
g0 = torch.mm(torch_kernel(xc, yc, sigmac, kernel=k), bc).cpu().numpy()
speed_pytorch = timeit.Timer(
'g0 = torch.mm(torch_kernel(xc,yc,sigmac,kernel=k),bc)#.cpu().numpy()',
GC,
globals=globals(),
timer=time.time).timeit(LOOPS)
print("Time for Pytorch: {:.4f}s".format(speed_pytorch),
end="")
print(" (absolute error: ", np.max(np.abs(g0 - gnumpy)), ")")
except:
pass
# specific cuda tiled implementation (if cuda is available)
if gpu_available:
try:
from pykeops.numpy.convolutions.radial_kernels import radial_kernels_conv