def _backward_cpu(self, context, dy):
if isinstance(self.attrs._x, Node):
N = len(dy)
col = np.zeros((N, self.attrs._in_shape[0], self.attrs._kernel[0],
self.attrs._kernel[1], self.attrs._out_shape[1], self.attrs._out_shape[2]))
col_k = np.rollaxis(col.reshape(
N, self.attrs._in_shape[0], -1, self.attrs._out_shape[1], self.attrs._out_shape[2]), 2)
col_k[:] = dy / float(len(col_k))
dx = col2im(col, self.attrs._in_shape[1:], self.attrs._stride, self.attrs._padding)
self.attrs._x._update_diff(context, dx)
def _backward_cpu(self, context, dy):
if isinstance(self.attrs._x, Node):
N = len(dy)
index = self.attrs._index
col = np.zeros((N, self.attrs._in_shape[0], self.attrs._kernel[0],
self.attrs._kernel[1], self.attrs._out_shape[1], self.attrs._out_shape[2]))
col_k = np.rollaxis(col.reshape(
N, self.attrs._in_shape[0], -1, self.attrs._out_shape[1], self.attrs._out_shape[2]), 2)
for i in np.ndindex(N, self.attrs._in_shape[0], self.attrs._out_shape[1], self.attrs._out_shape[2]):
col_k[index[i]][i] = dy[i]
dx = col2im(col, self.attrs._in_shape[1:], self.attrs._stride, self.attrs._padding)
self.attrs._x._update_diff(context, dx)
def _oper_cpu(cls, x, w, b, in_shape, out_shape, kernel, stride, padding):
z = np.tensordot(w, x, (0, 1))
z = np.rollaxis(z, 3)
z = col2im(z, out_shape[1:], stride, padding) + b
ret = cls._create_node(z)
ret.attrs._x = x
ret.attrs._w = w
ret.attrs._b = b
ret.attrs._in_shape = in_shape
ret.attrs._kernel = kernel
ret.attrs._stride = stride
ret.attrs._padding = padding
return ret
def _backward_cpu(self, context, dy):
dy = to_value(dy)
if isinstance(self.attrs._x, Node):
dx = np.tensordot(self.attrs._w, dy, (0, 1))
dx = np.rollaxis(dx, 3)
dx = col2im(dx, self.attrs._in_shape[1:],
self.attrs._stride, self.attrs._padding)
self.attrs._x._update_diff(context, dx)
if isinstance(self.attrs._w, Node):
self.attrs._w._update_diff(context, np.tensordot(
dy, self.attrs._col, ([0, 2, 3], [0, 4, 5])))
if isinstance(self.attrs._b, Node):
self.attrs._b._update_diff(context, np.sum(dy, (0, 2, 3), keepdims=True))
def _backward_cpu(self, context, dy, **kwargs):
dy = to_value(dy)
N, in_channels, in_h, in_w = self.attrs._x.shape
groups = self.attrs._groups
oCg = self.attrs._oCg
iCg = self.attrs._iCg
out_h, out_w = self.attrs._out_shape[-2:]
k_h, k_w = self.attrs._kernel
if isinstance(self.attrs._x, Node):
dy_temp = dy.transpose(1, 0, 2, 3)
dy_temp = dy_temp.reshape(groups, oCg, N * out_h * out_w)
w_temp = self.attrs._w.reshape(groups, oCg, iCg * k_h * k_w)
w_temp = w_temp.transpose(0, 2, 1)
dx = np.matmul(w_temp, dy_temp)
dx = dx.reshape(groups * iCg, k_h, k_w, N, out_h, out_w)
dx = np.rollaxis(dx, 3)
dx = col2im(dx, self.attrs._in_shape[1:], self.attrs._stride,
self.attrs._padding, self.attrs._dilation)
self.attrs._x._update_diff(context, dx, **kwargs)
if isinstance(self.attrs._w, Node):
col_temp = self.attrs._col
col_temp = col_temp.transpose(0, 2, 1)
dy_temp = dy.transpose(1, 0, 2, 3)
dy_temp = dy_temp.reshape(groups, oCg, N * out_h * out_w)
dw = np.matmul(dy_temp, col_temp)
dw = dw.reshape(groups * oCg, iCg, k_h, k_w)
self.attrs._w._update_diff(context, dw, **kwargs)
if isinstance(self.attrs._b, Node):
self.attrs._b._update_diff(context,
np.sum(dy, (0, 2, 3), keepdims=True),
**kwargs)