backward(tensors, grad_tensors=None, retain_graph=None, create_graph=False, grad_variables=None, inputs=None)¶
Computes the sum of gradients of given tensors with respect to graph leaves.
The graph is differentiated using the chain rule. If any of
tensorsare non-scalar (i.e. their data has more than one element) and require gradient, then the Jacobian-vector product would be computed, in this case the function additionally requires specifying
grad_tensors. It should be a sequence of matching length, that contains the “vector” in the Jacobian-vector product, usually the gradient of the differentiated function w.r.t. corresponding tensors (
Noneis an acceptable value for all tensors that don’t need gradient tensors).
This function accumulates gradients in the leaves - you might need to zero
.gradattributes or set them to
Nonebefore calling it. See Default gradient layouts for details on the memory layout of accumulated gradients.
Using this method with
create_graph=Truewill create a reference cycle between the parameter and its gradient which can cause a memory leak. We recommend using
autograd.gradwhen creating the graph to avoid this. If you have to use this function, make sure to reset the
.gradfields of your parameters to
Noneafter use to break the cycle and avoid the leak.
If you run any forward ops, create
grad_tensors, and/or call
backwardin a user-specified CUDA stream context, see Stream semantics of backward passes.
inputsare provided and a given input is not a leaf, the current implementation will call its grad_fn (even though it is not strictly needed to get this gradients). It is an implementation detail on which the user should not rely. See https://github.com/pytorch/pytorch/pull/60521#issuecomment-867061780 for more details.
grad_tensors (Sequence[Tensor or None] or Tensor, optional) – The “vector” in the Jacobian-vector product, usually gradients w.r.t. each element of corresponding tensors. None values can be specified for scalar Tensors or ones that don’t require grad. If a None value would be acceptable for all grad_tensors, then this argument is optional.
retain_graph (bool, optional) – If
False, the graph used to compute the grad will be freed. Note that in nearly all cases setting this option to
Trueis not needed and often can be worked around in a much more efficient way. Defaults to the value of
create_graph (bool, optional) – If
True, graph of the derivative will be constructed, allowing to compute higher order derivative products. Defaults to
inputs (Sequence[Tensor] or Tensor, optional) – Inputs w.r.t. which the gradient be will accumulated into
.grad. All other Tensors will be ignored. If not provided, the gradient is accumulated into all the leaf Tensors that were used to compute the attr::tensors.