mindspore.nn.WarmUpLR

class mindspore.nn.WarmUpLR(learning_rate, warmup_steps)

Gets learning rate warming up.

For current step, the formula of computing warmup learning rate is:

$$warmup\mathrm{\_}learning\mathrm{\_}rate=learning\mathrm{\_}rate\ast tmp\mathrm{\_}step/warmup\mathrm{\_}steps$$

Where

$$tmp\mathrm{\_}step=min(current\mathrm{\_}step,warmup\mathrm{\_}steps)$$

Parameters

• learning_rate (float) – The initial value of learning rate. The value of learning_rate must be greater than 0.

• warmup_steps (int) – The warm up steps of learning rate. The value of warmup_steps must be greater than or equal to 1.

Inputs:

• global_step (Tensor) - The current step number. Shape is $()$.

Outputs:

Tensor. The learning rate value for the current step with shape $()$.

Raises

• TypeError – If learning_rate is not a float.

• TypeError – If warmup_steps is not an int.

• ValueError – If warmup_steps is less than 1.

• ValueError – If learning_rate is less than or equal to 0.

Supported Platforms:

Ascend GPU CPU

Examples

>>> import mindspore
>>> from mindspore import Tensor, nn
>>>
>>> learning_rate = 0.1
>>> warmup_steps = 2
>>> global_step = Tensor(2, mindspore.int32)
>>> warmup_lr = nn.WarmUpLR(learning_rate, warmup_steps)
>>> lr = warmup_lr(global_step)
>>> net = nn.Dense(2, 3)
>>> optim = nn.SGD(net.trainable_params(), learning_rate=lr)