def _init_from_numpy(self, array):
self._static_shape = Size(array.shape)
# We use the scope of ``numpy`` instead of ``leaf``
# As it is costly to switch memory between ``copy`` and ``zero-copy``
self._dg_tensor = tensor_utils.FromPyArray(array,
name=TPool.get('numpy'))
self._ignored_grads = {self.name +
'_grad'} if not self._requires_grad else None
def _init_from_shape(self, shape):
if isinstance(shape, int): shape = [shape]
self._static_shape = Size(shape)
self._dg_tensor = tensor_utils.FromShape(
shape,
self._dtype,
ctx=CTX_TO_DEVICE_OPTION[tuple(self._ctx)],
name=TPool.get('leaf'))
self._ignored_grads = {self.name +
'_grad'} if not self._requires_grad else None
def register_buffers(self, n_buffers):
"""Apply for n buffers from TensorPool.
Buffers will be released after backward pass.
Parameters
----------
n_buffers : int
The number of buffers.
"""
return [TPool.get() for i in range(n_buffers)]