mindspore.ops.SparseApplyAdagradV2

class mindspore.ops.SparseApplyAdagradV2(*args, **kwargs)[source]

Updates relevant entries according to the adagrad scheme, one more epsilon attribute than SparseApplyAdagrad.

\[accum += grad * grad\]
\[var -= lr * grad * \frac{1}{\sqrt{accum} + \epsilon}\]

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.

Parameters

  • lr (float) – Learning rate.

  • epsilon (float) – A small value added for numerical stability.

  • use_locking (bool) – If True, the var and accum tensors will be protected from being updated. Default: False.

  • update_slots (bool) – If True, the computation logic will be different to False. Default: True.

Inputs:

  • var (Parameter) - Variable to be updated. The data type must be float16 or float32.

  • accum (Parameter) - Accumulation to be updated. The shape and data type must be the same as var.

  • grad (Tensor) - Gradient. The shape must be the same as var’s shape except the first dimension. Gradients has the same data type as var.

  • indices (Tensor) - A vector of indices into the first dimension of var and accum. The shape of indices must be the same as grad in first dimension, the type must be int32.

Outputs:

Tuple of 2 tensors, the updated parameters.

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

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

Raises

  • TypeError – If neither lr nor epsilon is a float.

  • TypeError – If neither update_slots nor use_locking is a bool.

  • TypeError – If dtype of var, accum or grad is neither float16 nor float32.

  • TypeError – If dtype of indices is not int32.

Supported Platforms:

Ascend

Examples

>>> import numpy as np
>>> import mindspore.nn as nn
>>> from mindspore import Tensor
>>> from mindspore import Parameter
>>> from mindspore.ops import operations as ops
>>> import mindspore.common.dtype as mstype
>>> class Net(nn.Cell):
...     def __init__(self):
...         super(Net, self).__init__()
...         self.sparse_apply_adagrad_v2 = ops.SparseApplyAdagradV2(lr=1e-8, epsilon=1e-6)
...         self.var = Parameter(Tensor(np.array([[0.2]]).astype(np.float32)), name="var")
...         self.accum = Parameter(Tensor(np.array([[0.1]]).astype(np.float32)), name="accum")
...
...     def construct(self, grad, indices):
...         out = self.sparse_apply_adagrad_v2(self.var, self.accum, grad, indices)
...         return out
...
>>> net = Net()
>>> grad = Tensor(np.array([[0.7]]).astype(np.float32))
>>> indices = Tensor(np.ones([1]), mstype.int32)
>>> output = net(grad, indices)
>>> print(output)
(Tensor(shape=[1, 1], dtype=Float32, value=
[[ 2.00000003e-01]]), Tensor(shape=[1, 1], dtype=Float32, value=
[[ 1.00000001e-01]]))