mindspore.nn.ForwardValueAndGrad

class mindspore.nn.ForwardValueAndGrad(network, weights=None, get_all=False, get_by_list=False, sens_param=False)[source]

Network training package class.

Including the network and a gradient function. The resulting Cell is trained with input ‘*inputs’. The backward graph will be created in the gradient function to calculating gradient.

Parameters
  • network (Cell) – The training network.

  • weights (ParameterTuple) – The parameters of the training network that need to calculate the gradient.

  • get_all (bool) – If True, get all the gradients with respect to inputs. Default: False.

  • get_by_list (bool) – If True, get all the gradients with respect to Parameter variables. If get_all and get_by_list are both False, get the gradient with respect to first input. If get_all and get_by_list are both True, get the gradients with respect to inputs and Parameter variables at the same time in the form of ((gradients with respect to inputs), (gradients with respect to parameters)). Default: False.

  • sens_param (bool) – Whether to append sensitivity (gradient with respect to output) as input. If sens_param is False, a ‘ones_like(outputs)’ sensitivity will be attached automatically. Default: False. If the sens_param is True, a sensitivity (gradient with respect to output) needs to be transferred through the input parameter.

Inputs:
  • (*inputs) (Tuple(Tensor…)) - Tuple of inputs with shape \((N, \ldots)\).

  • (sens) - A sensitivity (gradient with respect to output) as the input of backpropagation. If network has single output, the sens is a tensor. If network has multiple outputs, the sens is the tuple(tensor).

Outputs:
  • forward value - The result of network forward running.

  • gradients (tuple(tensor)) - The gradients of network parameters and inputs.

Supported Platforms:

Ascend GPU CPU

Examples

>>> class Net(nn.Cell):
...    def __init__(self):
...        super(Net, self).__init__()
...        self.weight = Parameter(Tensor(np.ones([2, 2]).astype(np.float32)), name="weight")
...        self.matmul = P.MatMul()
...
...    def construct(self, x):
...        out = self.matmul(x, self.weight)
...        return out
...
>>> net = Net()
>>> criterion = nn.SoftmaxCrossEntropyWithLogits()
>>> net_with_criterion = nn.WithLossCell(net, criterion)
>>> weight = ParameterTuple(net.trainable_params())
>>> train_network = nn.ForwardValueAndGrad(net_with_criterion, weights=weight, get_all=True, get_by_list=True)
>>> inputs = Tensor(np.ones([1, 2]).astype(np.float32))
>>> labels = Tensor(np.zeros([1, 2]).astype(np.float32))
>>> result = train_network(inputs, labels)
>>> print(result)
(Tensor(shape=[1], dtype=Float32, value=[0.00000000e+00]), ((Tensor(shape=[1, 2], dtype=Float32, value=
[[1.00000000e+00, 1.00000000e+00]]), Tensor(shape=[1, 2], dtype=Float32, value=
[[0.00000000e+00, 0.00000000e+00]])), (Tensor(shape=[2, 2], dtype=Float32, value=
[[5.00000000e-01, 5.00000000e-01],
 [5.00000000e-01, 5.00000000e-01]]),)))