def create_link(self, initializers):
initial_bias, = initializers
if self.in_channels == 'omit':
link = deconvolution_nd.DeconvolutionND(self.ndim,
self.out_channels,
self.ksize,
stride=self.stride,
pad=self.pad,
outsize=self.outsize,
initial_bias=initial_bias,
nobias=self.nobias,
groups=self.groups)
else:
link = deconvolution_nd.DeconvolutionND(self.ndim,
self.in_channels,
self.out_channels,
self.ksize,
stride=self.stride,
pad=self.pad,
outsize=self.outsize,
initial_bias=initial_bias,
nobias=self.nobias,
groups=self.groups)
return link
def test_deconv3d(self):
in_c, out_c, x = self._get_data(3)
link_nd = deconvolution_nd.DeconvolutionND(3,
in_c,
out_c,
2,
initialW=1)
link_3d = deconvolution_nd.Deconvolution3D(in_c, out_c, 2, initialW=1)
testing.assert_allclose(link_nd(x).data, link_3d(x).data)
def setUp(self):
N = 2
in_channels = 3
out_channels = 2
ndim = len(self.dims)
ksize = (3, ) * ndim
stride = (2, ) * ndim
pad = (1, ) * ndim
if self.used_outsize == 'case1' or self.used_outsize == 'None':
# Use output size determined with get_deconv_outsize.
outs = tuple(
conv.get_deconv_outsize(d, k, s, p)
for (d, k, s, p) in zip(self.dims, ksize, stride, pad))
elif self.used_outsize == 'case2':
# Use possible output size other than the one determined with
# get_deconv_outsize.
outs = tuple(
conv.get_deconv_outsize(d, k, s, p) + 1
for (d, k, s, p) in zip(self.dims, ksize, stride, pad))
if self.used_outsize != 'None':
outsize = outs
else:
outsize = None
if not self.nobias:
initial_bias = initializers.Uniform(scale=1, dtype=self.dtype)
else:
initial_bias = None
self.link = deconvolution_nd.DeconvolutionND(ndim,
in_channels,
out_channels,
ksize,
stride=stride,
pad=pad,
outsize=outsize,
initial_bias=initial_bias,
nobias=self.nobias)
self.link.cleargrads()
x_shape = (N, in_channels) + self.dims
self.x = numpy.random.uniform(-1, 1, x_shape).astype(self.dtype)
gy_shape = (N, out_channels) + outs
self.gy = numpy.random.uniform(-1, 1, gy_shape).astype(self.dtype)
self.check_forward_options = {}
self.check_backward_options = {'eps': 1e-2, 'atol': 1e-4, 'rtol': 1e-3}
if self.dtype == numpy.float16:
self.check_forward_options = {'atol': 5e-3, 'rtol': 5e-2}
self.check_backward_options = {
'eps': 2**-3,
'atol': 1e-2,
'rtol': 1e-1
}
def test_no_initial_bias(self):
ndim = 3
ksize = 3
link = deconvolution_nd.DeconvolutionND(ndim, 3, 2, ksize, nobias=True)
self.assertIsNone(link.b)
