def get_bprop_conv3d_transpose(self):
"""Grad definition for `Conv3DTranspose` operation."""
input_grad = nps.Conv3D(out_channel=self.in_channel,
kernel_size=self.kernel_size,
mode=self.mode,
pad_mode="pad",
pad=self.pad,
stride=self.stride,
dilation=self.dilation,
group=self.group,
data_format=self.data_format)
filter_grad = G.Conv3DBackpropFilter(out_channel=self.in_channel,
kernel_size=self.kernel_size,
mode=self.mode,
pad_mode="pad",
pad=self.pad,
stride=self.stride,
dilation=self.dilation,
group=self.group,
data_format=self.data_format)
def bprop(x, w, out, dout):
dx = input_grad(dout, w)
dw = filter_grad(dout, x, F.shape(w))
return dx, dw, zeros_like(out)
return bprop
def __init__(self,
in_channel,
out_channel,
kernel_size,
mode=1,
pad_mode="valid",
pad=0,
stride=1,
dilation=1,
group=1,
data_format="NCDHW",
bias_init="zeros",
has_bias=True):
super().__init__()
self.weight_shape = (out_channel, in_channel, kernel_size[0],
kernel_size[1], kernel_size[2])
self.weight = weight_variable(self.weight_shape)
self.conv = nps.Conv3D(out_channel=out_channel, kernel_size=kernel_size, mode=mode, \
pad_mode=pad_mode, pad=pad, stride=stride, dilation=dilation, \
group=group, data_format=data_format)
self.bias_init = bias_init
self.has_bias = has_bias
self.bias_add = P.BiasAdd(data_format=data_format)
if self.has_bias:
self.bias = Parameter(initializer(self.bias_init, [out_channel]),
name='bias')
