backward

paddle.Tensor. backward ( self: Tensor, grad_tensor: Tensor | None = None, retain_graph: bool = False, *, dump_backward_graph_path: str | None = None ) → None

Run backward of current Graph which starts from current Tensor.

The new gradient will accumulate on previous gradient.

You can clear gradient by Tensor.clear_grad() .

Parameters

grad_tensor (Tensor|None, optional) – initial gradient values of the current Tensor. If grad_tensor is None, the initial gradient values of the current Tensor would be Tensor filled with 1.0; if grad_tensor is not None, it must have the same length as the current Tensor. The default value is None.
retain_graph (bool, optional) – If False, the graph used to compute grads will be freed. If you would like to add more ops to the built graph after calling this method( backward ), set the parameter retain_graph to True, then the grads will be retained. Thus, setting it to False is much more memory-efficient. Defaults to False.
dump_backward_graph_path (str, optional) – Specifies the directory path for storing the debug file. If this parameter is specified, the backward-related graph (in dot format) and the debugging call stack information will be generated in this directory.

Returns

None

Examples

           >>> import paddle
>>> x = paddle.to_tensor(5., stop_gradient=False)
>>> for i in range(5):
...     y = paddle.pow(x, 4.0)
...     y.backward()
...     print("{}: {}".format(i, x.grad))
0: 500.0
1: 1000.0
2: 1500.0
3: 2000.0
4: 2500.0

>>> x.clear_grad()
>>> print("{}".format(x.grad))
0.0

>>> grad_tensor=paddle.to_tensor(2.)
>>> for i in range(5):
...     y = paddle.pow(x, 4.0)
...     y.backward(grad_tensor)
...     print("{}: {}".format(i, x.grad))
0: 1000.0
1: 2000.0
2: 3000.0
3: 4000.0
4: 5000.0

          