neon.backends.autodiff.Autodiff

neon.backends.autodiff.Autodiff(op_tree, be, next_error=None)[source]

Automatic differentiation given an op-tree.

Parameters:
  • op_tree (OpTreeNode) – the op-tree to take gradient of
  • be (Backend) – computation backend used
  • next_error (Tensor or OpTreeNode, optional) – next layer’s error, usually self.delta in a layer. If set to None, then automatically the default value is tensor ones() in output shape