mindspore.ops.ApplyAddSign

class mindspore.ops.ApplyAddSign[source]

Updates relevant entries according to the AddSign algorithm.

\[\begin{split}\begin{array}{ll} \\ m_{t+1} = \beta * m_{t} + (1 - \beta) * g \\ \text{update} = (\alpha + \text{sign_decay} * sign(g) * sign(m)) * g \\ var = var - lr_{t+1} * \text{update} \end{array}\end{split}\]

\(t\) represents updating step while \(m\) represents the 1st moment vector, \(m_{t}\) is the last momentent of \(m_{t+1}\), \(lr\) represents scaling factor lr, \(g\) represents grad, \(\alpha\) represents alpha, \(\beta\) represents beta.

Inputs of var, accum and grad comply with the implicit type conversion rules to make the data types consistent. If they have different data types, lower priority data type will be converted to relatively highest priority data type. RuntimeError exception will be thrown when the data type conversion of Parameter is required.

Inputs:
  • var (Parameter) - Variable tensor to be updated. With float32 or float16 data type. The shape is \((N, *)\) where \(*\) means, any number of additional dimensions.

  • m (Parameter) - Variable tensor to be updated, has the same shape and data type as var.

  • lr (Union[Number, Tensor]) - The learning rate value, must be a scalar. With float32 or float16 data type.

  • alpha (Union[Number, Tensor]) - Must be a scalar. With float32 or float16 data type.

  • sign_decay (Union[Number, Tensor]) - Must be a scalar. With float32 or float16 data type.

  • beta (Union[Number, Tensor]) - The exponential decay rate, must be a scalar. With float32 or float16 data type.

  • grad (Tensor) - A tensor of the same shape and data type as var, for the gradient.

Outputs:

Tuple of 2 Tensors, the updated parameters.

  • var (Tensor) - The same shape and data type as var.

  • m (Tensor) - The same shape and data type as m.

Raises
  • TypeError – If dtype of var, lr, alpha, sign_decay or beta is neither float16 nor float32.

  • TypeError – If lr, alpha or sign_decay is neither a Number nor a Tensor.

  • TypeError – If grad is not a Tensor.

Supported Platforms:

Ascend

Examples

>>> class Net(nn.Cell):
...     def __init__(self):
...         super(Net, self).__init__()
...         self.apply_add_sign = ops.ApplyAddSign()
...         self.var = Parameter(Tensor(np.array([[0.6, 0.4],
...                                               [0.1, 0.5]]).astype(np.float32)), name="var")
...         self.m = Parameter(Tensor(np.array([[0.6, 0.5],
...                                             [0.2, 0.6]]).astype(np.float32)), name="m")
...         self.lr = 0.001
...         self.alpha = 1.0
...         self.sign_decay = 0.99
...         self.beta = 0.9
...     def construct(self, grad):
...         out = self.apply_add_sign(self.var, self.m, self.lr, self.alpha, self.sign_decay, self.beta, grad)
...         return out
...
>>> net = Net()
>>> grad = Tensor(np.array([[0.3, 0.7], [0.1, 0.8]]).astype(np.float32))
>>> output = net(grad)
>>> print(output)
(Tensor(shape=[2, 2], dtype=Float32, value=
[[ 5.99403024e-01,  3.98607016e-01],
 [ 9.98010039e-02,  4.98407990e-01]]), Tensor(shape=[2, 2], dtype=Float32, value=
[[ 5.70000052e-01,  5.19999981e-01],
 [ 1.89999998e-01,  6.20000064e-01]]))