def inner(*args):
args = garray_to_cudandarray_nested(args)
res = f(*args)
if isinstance(res, list):
res = cudandarray_to_garray_nested(res)
else:
# TODO: check for CudaNdArray instance instead
if not isinstance(res, (float, np.ndarray)):
res = gput.cudandarray_to_garray(res)
return res
def inner(*args):
args = garray_to_cudandarray_nested(args)
res = f(*args)
if isinstance(res, list):
res = cudandarray_to_garray_nested(res)
else:
# TODO: check for CudaNdArray instance instead
if not isinstance(res, (float, np.ndarray)):
res = gput.cudandarray_to_garray(res)
return res
def Run_fmin_cg(self, maxiter):
x0 = self.classifier.flat_params.get_value(borrow=True,
return_internal_type=True)
x0 = gnumpy_utils.cudandarray_to_garray(x0)
f_op, params_opt = fmin_cg(f=self.ObjFun,
x0=x0,
fprime=self.ObjFunPrime,
maxiter=maxiter)
params_opt = gnumpy_utils.garray_to_cudandarray(params_opt)
self.classifier.flat_params.set_value(params_opt)
def inner(*args):
args = garray_to_cudandarray_nested(args)
res = f(*args)
if isinstance(res, list):
res = cudandarray_to_garray_nested(res)
else:
# TODO: check for CudaNdArray instance instead
if not isinstance(res, (float, np.ndarray)):
res = gput.cudandarray_to_garray(res, copyif=True) #GWJ-True -allows copy of contiguous memory
return res
def inner(*args):
args = garray_to_cudandarray_nested(args)
res = f(*args)
#print "res:", res
#return res
#return None
if isinstance(res, list):
res = cudandarray_to_garray_nested(res)
else:
# TODO: check for CudaNdArray instance instead
if not isinstance(res, (float, np.ndarray)):
res = gput.cudandarray_to_garray(res, copyif=True)
return res
def inner(*args):
args = garray_to_cudandarray_nested(args)
res = f(*args)
# print "res:", res
# return res
# return None
if isinstance(res, list):
res = cudandarray_to_garray_nested(res)
else:
# TODO: check for CudaNdArray instance instead
if not isinstance(res, (float, np.ndarray)):
res = gput.cudandarray_to_garray(res, copyif=True)
return res
def test(shape=(3,4,5)):
"""
Make sure that the gnumpy conversion is exact.
"""
gpu = theano.sandbox.cuda.basic_ops.gpu_from_host
U = gpu(theano.tensor.ftensor3('U'))
ii = theano.function([U], gpu(U+1))
A = gnumpy.rand(*shape)
A_cnd = garray_to_cudandarray(A)
B_cnd = ii(A_cnd)
B = cudandarray_to_garray(B_cnd)
from numpy import array
B2 = array(B_cnd)
u = (A+1).asarray()
v = B.asarray()
w = B2
assert abs(u-v).max() == 0
assert abs(u-w).max() == 0
def test(shape=(3, 4, 5)):
"""
Make sure that the gnumpy conversion is exact from garray to
CudaNdarray back to garray.
"""
gpu = theano.sandbox.cuda.basic_ops.gpu_from_host
U = gpu(theano.tensor.ftensor3('U'))
ii = theano.function([U], gpu(U + 1))
A = gnumpy.rand(*shape)
A_cnd = garray_to_cudandarray(A)
assert A_cnd.shape == A.shape
# dtype always float32
# garray don't have strides
B_cnd = ii(A_cnd)
B = cudandarray_to_garray(B_cnd)
assert A_cnd.shape == A.shape
u = (A + 1).asarray()
v = B.asarray()
w = np.array(B_cnd)
assert (u == v).all()
assert (u == w).all()
def test2(shape=(3, 4, 5)):
"""
Make sure that the gnumpy conversion is exact from CudaNdarray to
garray back to CudaNdarray.
"""
gpu = theano.sandbox.cuda.basic_ops.gpu_from_host
U = gpu(theano.tensor.ftensor3('U'))
theano.function([U], gpu(U + 1))
A = np.random.rand(*shape).astype('float32')
A_cnd = theano.sandbox.cuda.CudaNdarray(A)
A_gar = cudandarray_to_garray(A_cnd)
assert A_cnd.shape == A_gar.shape
# dtype always float32
# garray don't have strides
B = garray_to_cudandarray(A_gar)
assert A_cnd.shape == B.shape
# dtype always float32
assert A_cnd._strides == B._strides
assert A_cnd.gpudata == B.gpudata
v = np.asarray(B)
assert (v == A).all()
def cudandarray_to_garray_nested(lst):
lst_flat = flatten(lst)
lst_flat = [gput.cudandarray_to_garray(i) for i in lst_flat]
lst = unflatten(lst, lst_flat)
return lst
def cudandarray_to_garray_nested(lst):
lst_flat = flatten(lst)
lst_flat = [gput.cudandarray_to_garray(i) for i in lst_flat]
lst = unflatten(lst, lst_flat)
return lst
def ObjFunPrime(self, params):
params = gnumpy_utils.garray_to_cudandarray(params)
grad = self.grad_Tfunc(params)
return gnumpy_utils.cudandarray_to_garray(grad)
