def init(self,
in_shape,
out_channels,
kernel_size,
stride=1,
global_args=None,
bias=True,
padding=0,
out_padding=0,
activation="ReLU",
**kargs):
self.prev = in_shape
self.in_channels = in_shape[0]
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
self.out_padding = out_padding
self.activation = activation
self.use_softplus = h.default(global_args, 'use_softplus', False)
weights_shape = (self.in_channels, self.out_channels, kernel_size,
kernel_size)
self.weight = torch.nn.Parameter(torch.Tensor(*weights_shape))
if bias:
self.bias = torch.nn.Parameter(torch.Tensor(weights_shape[0]))
else:
self.bias = None # h.zeros(weights_shape[0])
outshape = getShapeConvTranspose(
in_shape, (out_channels, kernel_size, kernel_size), stride,
padding, out_padding)
return outshape
def init(self, prev, out_channels, kernel_size, stride = 1, global_args = None, bias=True, padding = 0, **kargs):
self.prev = prev
self.in_channels = prev[0]
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
self.use_softplus = h.default(global_args, 'use_softplus', False)
weights_shape = (self.out_channels, self.in_channels, kernel_size, kernel_size)
self.weight = torch.nn.Parameter(torch.Tensor(*weights_shape))
if bias:
self.bias = torch.nn.Parameter(torch.Tensor(weights_shape[0]))
else:
self.bias = h.zeros(weights_shape[0])
outshape = getShapeConv(prev, (out_channels, kernel_size, kernel_size), stride, padding)
return outshape
def init(self, prev, global_args = None, **kargs):
self.use_softplus = h.default(global_args, 'use_softplus', False)
return prev