def GpuConvOp_from_ConvOp(op):
logical_img_hw = None
if op.kshp_logical is not None and op.kshp_logical != op.kshp:
return None
#print op.kshp, op.imshp[1:3]
#print op.kshp_logical, logical_img_hw
ret = GpuConv(
border_mode=op.out_mode,
subsample=(op.dx, op.dy),
logical_img_hw=logical_img_hw,
logical_kern_hw=op.kshp_logical,
logical_kern_align_top=op.kshp_logical_top_aligned,
kshp=op.kshp,
version=op.version,
verbose=op.verbose,
imshp=op.imshp,
)
if op.imshp_logical is not None:
logical_img_hw = op.imshp_logical[1:3]
if logical_img_hw != op.imshp[1:3]:
# this case is not implemented
#return None
rstride = int(
numpy.ceil(op.imshp_logical[1] / float(op.imshp[1])))
cstride = int(
numpy.ceil(op.imshp_logical[2] / float(op.imshp[2])))
def make_graph(img, kern):
buf = tensor.alloc(numpy.asarray(0, dtype=img.dtype),
img.shape[0], *op.imshp_logical)
img = tensor.set_subtensor(buf[:, :, ::rstride, ::cstride],
img)
img = gpu_from_host(img)
return ret(img, kern)
return make_graph
return ret
def _params_allgood(ishape, kshape, mode, subsample=(1, 1), img_stride=(1, 1),
kern_stride=(1, 1), version=-1, verbose=0, random=True,
print_=None, id=None, rtol=1e-5, atol=1e-8,
nb_iter=0, ones=False, compile_kshp=None):
#
# This function is the core of several of the big unit-test drivers,
# but it can also be used very directly on its own to test a specific
# kind of convolution.
#
# See `test_example` (above) for an example of how to use this directly.
#
# :param kshape: (4d)The shape of the kernel at run time.
# :param compile_kshp: (2d) hardcode the shape of the kernel in
# the generated code This is supposed to be
# faster, but we need to check That we raise
# an error if the input have the wrong shape.
#
if ones:
assert not random
npy_img = theano._asarray(numpy.ones(ishape), dtype='float32')
npy_kern = -theano._asarray(numpy.ones(kshape), dtype='float32')
elif random:
npy_img = theano._asarray(numpy.random.rand(*ishape) + 1,
dtype='float32')
npy_kern = theano._asarray(numpy.random.rand(*kshape) - 2,
dtype='float32')
else:
npy_img = theano._asarray(numpy.arange(
numpy.prod(ishape)).reshape(ishape), dtype='float32') + 1
npy_kern = -(theano._asarray(numpy.arange(
numpy.prod(kshape)).reshape(kshape), dtype='float32') + 1)
img = pygpu.array(npy_img)
kern = pygpu.array(npy_kern)
#we take the stride after the transfert as we make c_contiguous
#data on the GPU.
if img_stride != (1, 1):
img = img[:, :, ::img_stride[0], ::img_stride[1]]
npy_img = npy_img[:, :, ::img_stride[0], ::img_stride[1]]
if kern_stride != (1, 1):
kern = kern[:, :, ::kern_stride[0], ::kern_stride[1]]
npy_kern = npy_kern[:, :, ::kern_stride[0], ::kern_stride[1]]
t2 = None
rval = True
try:
t0 = time.time()
cpuval = py_conv(npy_img, npy_kern, mode, subsample)
t1 = time.time()
i = gftensor4()
k = gftensor4()
op = GpuConv(border_mode=mode,
subsample=subsample,
version=version,
verbose=verbose,
kshp=compile_kshp)(i, k)
f = theano.function([i, k], op, mode=mode_with_gpu)
gpuval = f(img, kern)
t2 = time.time()
for i in range(nb_iter):
gpuval2 = f(img, kern)
assert numpy.allclose(numpy.asarray(gpuval),
numpy.asarray(gpuval2))
assert (numpy.asarray(gpuval) == numpy.asarray(gpuval2)).all()
gpuval = numpy.asarray(gpuval)
if gpuval.shape != cpuval.shape:
print >> sys.stdout, "ERROR: shape mismatch",
print >> sys.stdout, gpuval.shape, cpuval.shape
rval = False
if rval:
rval = numpy.allclose(cpuval, gpuval, rtol=rtol)
assert numpy.all(numpy.isfinite(gpuval))
except NotImplementedError, e:
print >> sys.stdout, '_params_allgood Failed allclose', e
rval = False