mindspore.mint.optim.AdamW

class mindspore.mint.optim.AdamW(params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=1e-2, amsgrad=False, *, maximize=False)

Implements Adam Weight Decay algorithm.

\[\begin{split}\begin{array}{l} &\newline &\hline \\ &\textbf{input} : \gamma \text{(lr)}, \: \beta_1, \beta_2 \text{(betas)}, \: \theta_0 \text{(params)}, \: f(\theta) \text{(objective)}, \: \epsilon \text{ (epsilon)} \\ &\hspace{13mm} \lambda \text{(weight decay)}, \: \textit{amsgrad}, \: \textit{maximize} \\ &\textbf{initialize} : m_0 \leftarrow 0 \text{ (first moment)}, v_0 \leftarrow 0 \text{ ( second moment)}, \: \widehat{v_0}^{max}\leftarrow 0 \\[-1.ex] &\newline &\hline \\ &\textbf{for} \: t=1 \: \textbf{to} \: \ldots \: \textbf{do} \\ &\hspace{6mm}\textbf{if} \: \textit{maximize}: \\ &\hspace{11mm}g_t \leftarrow -\nabla_{\theta} f_t (\theta_{t-1}) \\ &\hspace{6mm}\textbf{else} \\ &\hspace{11mm}g_t \leftarrow \nabla_{\theta} f_t (\theta_{t-1}) \\ &\hspace{6mm} \theta_t \leftarrow \theta_{t-1} - \gamma \lambda \theta_{t-1} \\ &\hspace{6mm}m_t \leftarrow \beta_1 m_{t-1} + (1 - \beta_1) g_t \\ &\hspace{6mm}v_t \leftarrow \beta_2 v_{t-1} + (1-\beta_2) g^2_t \\ &\hspace{6mm}\widehat{m_t} \leftarrow m_t/\big(1-\beta_1^t \big) \\ &\hspace{6mm}\widehat{v_t} \leftarrow v_t/\big(1-\beta_2^t \big) \\ &\hspace{6mm}\textbf{if} \: amsgrad \\ &\hspace{11mm}\widehat{v_t}^{max} \leftarrow \mathrm{max}(\widehat{v_t}^{max}, \widehat{v_t}) \\ &\hspace{11mm}\theta_t \leftarrow \theta_t - \gamma \widehat{m_t}/ \big(\sqrt{\widehat{v_t}^{max}} + \epsilon \big) \\ &\hspace{6mm}\textbf{else} \\ &\hspace{11mm}\theta_t \leftarrow \theta_t - \gamma \widehat{m_t}/ \big(\sqrt{\widehat{v_t}} + \epsilon \big) \\ &\newline &\hline \\[-1.ex] &\bf{return} \: \theta_t \\[-1.ex] &\newline &\hline \\[-1.ex] \end{array}\end{split}\]

More details of the AdamW algorithm can be found in the paper Decoupled Weight Decay Regularization and On the Convergence of Adam and Beyond.

Warning

• This is an experimental optimizer API that is subject to change. This module must be used with lr scheduler module in LRScheduler Class .

• For Ascend, it is only supported on platforms above Atlas A2.

Parameters

• params (Union[list(Parameter), list(dict)]) – list of parameters to optimize or dicts defining parameter groups.

• lr (float, optional) – learning rate. Default: 1e-3.

• betas (Tuple[float, float], optional) – The exponential decay rate for the moment estimations. Default: (0.9, 0.999).

• eps (float, optional) – term added to the denominator to improve numerical stability. Must be greater than 0. Default: 1e-8.

• weight_decay (float, optional) – weight decay (L2 penalty). Default: 1e-2..

• amsgrad (bool, optional) – whether to use the AMSGrad algorithm. Default: False.

Keyword Arguments

maximize (bool, optional) – maximize the params based on the objective, instead of minimizing. Default: False.

Inputs:

• gradients (tuple[Tensor]) - The gradients of params.

Raises

• ValueError – If the learning rate is not float.

• ValueError – If the learning rate is less than 0.

• ValueError – If the eps is less than 0.

• ValueError – If the betas not in the range of [0, 1).

• ValueError – If the weight_decay is less than 0.

Supported Platforms:

Ascend

Examples

>>> import mindspore
>>> from mindspore import mint
>>> from mindspore.mint import optim
>>> # Define the network structure of LeNet5. Refer to
>>> # https://gitee.com/mindspore/docs/blob/master/docs/mindspore/code/lenet.py
>>> net = LeNet5()
>>> loss_fn = nn.SoftmaxCrossEntropyWithLogits(sparse=True)
>>> optimizer = optim.AdamW(net.trainable_params(), lr=0.1)
>>> def forward_fn(data, label):
...     logits = net(data)
...     loss = loss_fn(logits, label)
...     return loss, logits
>>> grad_fn = mindspore.value_and_grad(forward_fn, None, optimizer.parameters, has_aux=True)
>>> def train_step(data, label):
...     (loss, _), grads = grad_fn(data, label)
...     optimizer(grads)
...     return loss