class mindspore.ops.ApplyCenteredRMSProp(use_locking=False)[source]

Optimizer that implements the centered RMSProp algorithm. Please refer to the usage in source code of mindspore.nn.RMSProp.

The updating formulas of ApplyCenteredRMSProp algorithm are as follows,

\[\begin{split}\begin{array}{ll} \\ g_{t+1} = \rho g_{t} + (1 - \rho)\nabla Q_{i}(w) \\ s_{t+1} = \rho s_{t} + (1 - \rho)(\nabla Q_{i}(w))^2 \\ m_{t+1} = \beta m_{t} + \frac{\eta} {\sqrt{s_{t+1} - g_{t+1}^2 + \epsilon}} \nabla Q_{i}(w) \\ w = w - m_{t+1} \end{array}\end{split}\]

where \(w\) represents var, which will be updated. \(g_{t+1}\) represents mean_gradient, \(g_{t}\) is the last moment of \(g_{t+1}\). \(s_{t+1}\) represents mean_square, \(s_{t}\) is the last moment of \(s_{t+1}\), \(m_{t+1}\) represents moment, \(m_{t}\) is the last moment of \(m_{t+1}\). \(\rho\) represents decay. \(\beta\) is the momentum term, represents momentum. \(\epsilon\) is a smoothing term to avoid division by zero, represents epsilon. \(\eta\) represents learning_rate. \(\nabla Q_{i}(w)\) represents grad.


The difference between ApplyCenteredRMSProp and ApplyRMSProp is that the former uses the centered RMSProp algorithm, and the centered RRMSProp algorithm uses an estimate of the centered second moment(i.e., the variance) for normalization, as opposed to regular RMSProp, which uses the (uncertained) second moment. This often helps with training, but is slightly more expensive in terms of computation and memory.


In dense implementation of this algorithm, mean_gradient, mean_square, and moment will update even if the grad is zero. But in this sparse implementation, mean_gradient, mean_square, and moment will not update in iterations during which the grad is zero.


use_locking (bool) – Whether to enable a lock to protect the variable and accumulation tensors from being updated. Default: False.

  • var (Tensor) - Weights to be updated.

  • mean_gradient (Tensor) - Mean gradients, must be the same type as var.

  • mean_square (Tensor) - Mean square gradients, must be the same type as var.

  • moment (Tensor) - Delta of var, must be the same type as var.

  • grad (Tensor) - Gradient, must be the same type as var.

  • learning_rate (Union[Number, Tensor]) - Learning rate. Must be a float number or a scalar tensor with float16 or float32 data type.

  • decay (float) - Decay rate.

  • momentum (float) - Momentum.

  • epsilon (float) - Ridge term.


Tensor, parameters to be updated.

  • TypeError – If use_locking is not a bool.

  • TypeError – If var, mean_gradient, mean_square, moment or grad is not a Tensor.

  • TypeError – If learing_rate is neither a Number nor a Tensor.

  • TypeError – If dtype of learing_rate is neither float16 nor float32.

  • TypeError – If decay, momentum or epsilon is not a float.

Supported Platforms:

Ascend GPU CPU


>>> class Net(nn.Cell):
...     def __init__(self):
...         super(Net, self).__init__()
...         self.apply_centerd_rms_prop = ops.ApplyCenteredRMSProp()
...         self.var = Parameter(Tensor(np.ones([2, 2]).astype(np.float32)), name="var")
...     def construct(self, mean_grad, mean_square, moment, grad, decay, momentum, epsilon, lr):
...         out = self.apply_centerd_rms_prop(self.var, mean_grad, mean_square, moment, grad,
...                                           lr, decay, momentum, epsilon)
...         return out
>>> net = Net()
>>> mean_grad = Tensor(np.ones([2, 2]).astype(np.float32))
>>> mean_square = Tensor(np.ones([2, 2]).astype(np.float32))
>>> moment = Tensor(np.ones([2, 2]).astype(np.float32))
>>> grad = Tensor(np.ones([2, 2]).astype(np.float32))
>>> output = net(mean_grad, mean_square, moment, grad, 0.0, 1e-10, 0.001, 0.01)
>>> print(net.var.asnumpy())
[[0.68377227  0.68377227]
 [0.68377227  0.68377227]]