def convolution2DParam(out_dims, dy, dx, sy, sx, ph, pw, pd, pr):
cp = conv2DParam()
cp.out_dims = intVector()
for d in out_dims:
cp.out_dims.push_back(d)
cp.dilate_y, cp.dilate_x = (dy - 1), (dx - 1)
cp.sy, cp.sx = sy, sx
cp.pad_lh, cp.pad_lw = ph, pw
cp.pad_rh, cp.pad_rw = pd, pr
return cp
def pooling2DParam(out_dims, kh, kw, sy, sx, ph, pw, pd, pr, algo):
pp = pol2DParam()
pp.out_dims = intVector()
for d in out_dims:
pp.out_dims.push_back(d)
pp.kh, pp.kw = kh, kw
pp.sy, pp.sx = sy, sx
pp.pad_lh, pp.pad_lw = ph, pw
pp.pad_rh, pp.pad_rw = pd, pr
pp.algo_kind = algo
return pp
def split(x, indices_or_sections, axis=0):
if all_ready((x, )):
offsets = intVector()
if numpy.isscalar(indices_or_sections):
if indices_or_sections == 0:
raise ValueError('integer division or modulo by zero')
elif x.shape[axis] % indices_or_sections:
raise ValueError(
'array split does not result in an equal division')
for i in range(x.shape[axis] / indices_or_sections, x.shape[axis],
x.shape[axis] / indices_or_sections):
offsets.push_back(int(i))
else:
# FIXME
# bypass python3 issue
for i in indices_or_sections:
offsets.push_back(int(i))
ys = concat.Backward(x, offsets, axis)
if ys:
# indices_or_sections = [0, ...]
# axis = 0
if not numpy.isscalar(indices_or_sections) and \
axis == 0 and indices_or_sections[0] == 0:
shape = x.shape
shape = (0, ) + shape[1:]
y1 = numpy.ndarray(shape, dtype=x.dtype)
ys = list((y1, ) + ys)
else:
# For performance improvement
# indices_or_sections = 1
if numpy.isscalar(indices_or_sections) and \
indices_or_sections == 1:
ys = [x]
# indices_or_sections = [0]
# axis = 0
elif axis == 0 and indices_or_sections[0] == 0 \
and len(indices_or_sections) == 1:
shape = x.shape
shape = (0, ) + shape[1:]
y1 = numpy.ndarray(shape, dtype=x.dtype)
ys = list((y1, ) + (x, ))
# other not support scenarios
else:
ys = numpy.split(x, indices_or_sections, axis)
else:
ys = numpy.split(x, indices_or_sections, axis)
return ys