mindspore.ops.ApplyAdagradDA

class mindspore.ops.ApplyAdagradDA(use_locking=False)

Update var according to the proximal adagrad scheme. The Adagrad algorithm was proposed in Adaptive Subgradient Methods for Online Learning and Stochastic Optimization.

\[\begin{split}\begin{array}{ll} \\ grad\_accum += grad \\ grad\_squared\_accum += grad * grad \\ tmp\_val= \begin{cases} sign(grad\_accum) * max\left \{|grad\_accum|-l1*global\_step, 0\right \} & \text{ if } l1>0 \\ grad\_accum & \text{ otherwise } \\ \end{cases} \\ x\_value = -1 * lr * tmp\_val \\ y\_value = l2 * global\_step * lr + \sqrt{grad\_squared\_accum} \\ var = \frac{ x\_value }{ y\_value } \end{array}\end{split}\]

Inputs of var, gradient_accumulator, gradient_squared_accumulator and grad comply with the implicit type conversion rules to make the data types consistent. If they have different data types, the lower priority data type will be converted to the relatively highest priority data type.

Parameters

use_locking (bool) – If True , updating of the var and accum tensors will be protected by a lock. Otherwise the behavior is undefined, but may exhibit less contention. Default: False .

Inputs:

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

• gradient_accumulator (Parameter) - The dict of mutable tensor \(grad\_accum\). Must have the same shape as var.

• gradient_squared_accumulator (Parameter) - The dict of mutable tensor \(grad\_squared\_accum\). Must have the same shape as var.

• grad (Tensor) - A tensor for gradient. Must have the same shape as var.

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

• l1 ([Number, Tensor]) - L1 regularization. Must be a scalar. With float32 or float16 data type.

• l2 ([Number, Tensor]) - L2 regularization. Must be a scalar. With float32 or float16 data type.

• global_step ([Number, Tensor]) - Training step number. Must be a scalar. With int32 or int64 data type.

Outputs:

Tuple of 3 Tensors, the updated parameters.

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

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

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

Raises

• TypeError – If var, gradient_accumulator or gradient_squared_accumulator is not a Parameter.

• TypeError – If grad is not a Tensor.

• TypeError – If lr, l1, l2 or global_step is neither a Number nor a Tensor.

• TypeError – If use_locking is not a bool.

• TypeError – If dtype of var, gradient_accumulator, gradient_squared_accumulator, grad, lr, l1 or l2 is neither float16 nor float32.

• TypeError – If dtype of gradient_accumulator, gradient_squared_accumulator or grad is not same as var.

• TypeError – If dtype of global_step is not int32 nor int64.

• ValueError – If the shape size of lr, l1, l2 and global_step is not 0.

• TypeError – If the data type of var, gradient_accumulator, gradient_squared_accumulator and grad conversion of Parameter is not supported.

Supported Platforms:

Ascend GPU CPU

Examples

>>> import numpy as np
>>> from mindspore import dtype as mstype
>>> from mindspore import Tensor, nn, ops, Parameter
>>> class ApplyAdagradDANet(nn.Cell):
...     def __init__(self, use_locking=False):
...         super(ApplyAdagradDANet, self).__init__()
...         self.apply_adagrad_d_a = ops.ApplyAdagradDA(use_locking)
...         self.var = Parameter(Tensor(np.array([[0.6, 0.4], [0.1, 0.5]]).astype(np.float32)), name="var")
...         self.gradient_accumulator = Parameter(Tensor(np.array([[0.1, 0.3],
...                                                                [0.1, 0.5]]).astype(np.float32)),
...                                               name="gradient_accumulator")
...         self.gradient_squared_accumulator = Parameter(Tensor(np.array([[0.2, 0.1],
...                                                                        [0.1, 0.2]]).astype(np.float32)),
...                                                       name="gradient_squared_accumulator")
...         self.gradient_accumulator = Parameter(Tensor(np.array([[0.1, 0.3],
...                                                                [0.1, 0.5]]).astype(np.float32)),
...                                               name="gradient_accumulator")
...     def construct(self, grad, lr, l1, l2, global_step):
...         out = self.apply_adagrad_d_a(self.var, self.gradient_accumulator,
...                                      self.gradient_squared_accumulator, grad, lr, l1, l2, global_step)
...         return out
...
>>> net = ApplyAdagradDANet()
>>> grad = Tensor(np.array([[0.3, 0.4], [0.1, 0.2]]).astype(np.float32))
>>> lr = Tensor(0.001, mstype.float32)
>>> l1 = Tensor(0.001, mstype.float32)
>>> l2 = Tensor(0.001, mstype.float32)
>>> global_step = Tensor(2, mstype.int32)
>>> output = net(grad, lr, l1, l2, global_step)
>>> print(output)
(Tensor(shape=[2, 2], dtype=Float32, value=
[[-7.39064650e-04, -1.36888528e-03],
 [-5.96988888e-04, -1.42478070e-03]]), Tensor(shape=[2, 2], dtype=Float32, value=
[[ 4.00000006e-01,  7.00000048e-01],
 [ 2.00000003e-01,  6.99999988e-01]]), Tensor(shape=[2, 2], dtype=Float32, value=
[[ 2.90000021e-01,  2.60000020e-01],
 [ 1.09999999e-01,  2.40000010e-01]]))