def forward(self, *args, **kwargs):
self.args = [self._convert2array(arg) for arg in args]
if self.enable_auto_broadcast:
self.args = self._autobroadcast(self.args)
self.kwargs = kwargs
out = self._forward(*tuple(arg.value for arg in self.args), **kwargs)
if config.enable_backprop:
return Array(out, parent=self)
else:
return Array(out, parent=None)
def __init__(self, kernel, stride, pad):
super().__init__()
self.in_ch = kernel.shape[-2]
self.out_ch = kernel.shape[-1]
self.kernel_size = kernel.shape[:2]
self.stride = stride
self.pad = pad
kernel = kernel.value
with self.set_parameter():
self.w = Array(kernel.reshape(-1, kernel.shape[-1]))
class Convolve2d(Network):
def __init__(self, kernel, stride, pad):
super().__init__()
self.in_ch = kernel.shape[-2]
self.out_ch = kernel.shape[-1]
self.kernel_size = kernel.shape[:2]
self.stride = stride
self.pad = pad
kernel = kernel.value
with self.set_parameter():
self.w = Array(kernel.reshape(-1, kernel.shape[-1]))
@property
def kernel(self):
return self.w.reshape(*self.kernel_size, self.in_ch, self.out_ch)
def __call__(self, x):
func = Convolve2dFunction(self.kernel_size, self.stride, self.pad)
return func.forward(x, self.w)
def zeros(size):
return Array(np.zeros(size, dtype=config.dtype))
def ones(size):
return Array(np.ones(size, dtype=config.dtype))