Method GetBackwardPass

caffe2/python/core.py:1046–1104

Gets the backward pass that computes the derivatives of given blobs. Inputs: ys: a list or a dictionary specifying what blobs we want to compute derivatives of. If the input is a list, we will automatically generate their gradients with all-one values;

(self, ys)

        return new_input_to_grad, gradient_ops

    def GetBackwardPass(self, ys):
        """Gets the backward pass that computes the derivatives of given blobs.

        Inputs:
          ys: a list or a dictionary specifying what blobs we want to compute
              derivatives of. If the input is a list, we will automatically
              generate their gradients with all-one values; if the input is a
              dictionary, for any dictionary entries that are not None, we will
              take the corresponding blobs as their gradients; for all those
              that are None, we will auto-fill them with 1.
        """
        if isinstance(ys, list):
            ys = dict((y, None) for y in ys)
        elif not isinstance(ys, dict):
            raise TypeError("ys should either be a list or a dict.")

        # Set the gradient frontier with the initialized external
        # gradients.
        for y in ys.keys():
            self.gradient_frontier[y] = self.frontier[y]
            self.input_usages[str(y)][self.frontier[str(y)]].append(
                len(self.ssa))

        all_input_to_grad, all_gradient_ops = self._GetInitGradients(ys)

        # (2) Now, after having the virtual play above, we now play the ops
        # backwards, creating the gradients along the path. Note that although
        # we are playing it backwards, we cannot refer to variables that are
        # at a version older than current_versions because it is already been
        # overwritten.
        for forward_op_idx in reversed(range(len(self.ssa))):
            input_to_grad, gradient_ops = self._GenerateGradientsForForwardOp(
                forward_op_idx, all_input_to_grad)
            all_input_to_grad.update(input_to_grad)
            all_gradient_ops += gradient_ops

            # If there are multiple use blobs, do gradient accumulation.
            additional_sum_ops, grad_map = self.DoGradientAccumulation(
                forward_op_idx)
            # This line is so that if in an accumulation some of the operators
            # have not produced gradients, they still do not overwrite the
            # general all_input_to_grad map.
            all_input_to_grad.update(grad_map)
            all_gradient_ops += additional_sum_ops

        # (3) Post-processing.
        # After we have done computation for each op, we now have the gradient
        # operators ready. For the output map, we will convert everything to
        # BlobReferences for easier handling in python.
        all_input_to_grad_out = {}
        for key, val in all_input_to_grad.items():
            if val is not None:
                if isinstance(val, (bytes, str)):
                    grad_out = BlobReference(val)
                else:
                    grad_out = GradientSlice(BlobReference(val[0]),
                                             BlobReference(val[1]))
                all_input_to_grad_out[BlobReference(key)] = grad_out