forked from mindspore-Ecosystem/mindspore
215 lines
4.5 KiB
ReStructuredText
215 lines
4.5 KiB
ReStructuredText
mindspore.nn
|
||
=============
|
||
|
||
神经网络Cell。
|
||
|
||
用于构建神经网络中的预定义构建块或计算单元。
|
||
|
||
MindSpore中 `mindspore.nn` 接口与上一版本相比，新增、删除和支持平台的变化信息请参考 `API Updates <https://gitee.com/mindspore/docs/blob/master/resource/api_updates/ops_api_updates.md>`_。
|
||
|
||
基本构成单元
|
||
------------
|
||
|
||
.. cnmsplatformautosummary::
|
||
:toctree: nn
|
||
|
||
mindspore.nn.Cell
|
||
|
||
容器
|
||
-----------
|
||
|
||
.. cnmsplatformautosummary::
|
||
:toctree: nn
|
||
|
||
mindspore.nn.CellList
|
||
mindspore.nn.SequentialCell
|
||
|
||
卷积层
|
||
--------------------
|
||
|
||
.. cnmsplatformautosummary::
|
||
:toctree: nn
|
||
|
||
mindspore.nn.Conv1d
|
||
mindspore.nn.Conv1dTranspose
|
||
mindspore.nn.Conv2d
|
||
mindspore.nn.Conv2dTranspose
|
||
mindspore.nn.Conv3d
|
||
mindspore.nn.Conv3dTranspose
|
||
|
||
梯度
|
||
-----
|
||
|
||
.. cnmsplatformautosummary::
|
||
:toctree: nn
|
||
|
||
mindspore.nn.Jvp
|
||
mindspore.nn.Vjp
|
||
|
||
非线性激活函数
|
||
--------------
|
||
|
||
.. cnmsplatformautosummary::
|
||
:toctree: nn
|
||
|
||
mindspore.nn.ELU
|
||
mindspore.nn.FastGelu
|
||
mindspore.nn.GELU
|
||
mindspore.nn.HShrink
|
||
mindspore.nn.HSigmoid
|
||
mindspore.nn.HSwish
|
||
mindspore.nn.LeakyReLU
|
||
mindspore.nn.LogSigmoid
|
||
mindspore.nn.LogSoftmax
|
||
mindspore.nn.ReLU
|
||
mindspore.nn.Sigmoid
|
||
mindspore.nn.Softmax
|
||
mindspore.nn.Tanh
|
||
|
||
工具
|
||
-----
|
||
|
||
.. cnmsplatformautosummary::
|
||
:toctree: nn
|
||
|
||
mindspore.nn.Flatten
|
||
mindspore.nn.Tril
|
||
|
||
损失函数
|
||
--------
|
||
|
||
.. cnmsplatformautosummary::
|
||
:toctree: nn
|
||
|
||
mindspore.nn.L1Loss
|
||
mindspore.nn.MSELoss
|
||
mindspore.nn.SmoothL1Loss
|
||
|
||
优化器
|
||
-------
|
||
|
||
.. cnmsplatformautosummary::
|
||
:toctree: nn
|
||
|
||
mindspore.nn.Adagrad
|
||
mindspore.nn.Adam
|
||
mindspore.nn.AdamOffload
|
||
mindspore.nn.AdamWeightDecay
|
||
mindspore.nn.FTRL
|
||
mindspore.nn.Lamb
|
||
mindspore.nn.LARS
|
||
mindspore.nn.LazyAdam
|
||
mindspore.nn.Momentum
|
||
mindspore.nn.Optimizer
|
||
mindspore.nn.ProximalAdagrad
|
||
mindspore.nn.RMSProp
|
||
mindspore.nn.SGD
|
||
mindspore.nn.thor
|
||
|
||
封装函数
|
||
---------
|
||
|
||
.. cnmsplatformautosummary::
|
||
:toctree: nn
|
||
|
||
mindspore.nn.DistributedGradReducer
|
||
mindspore.nn.DynamicLossScaleUpdateCell
|
||
mindspore.nn.FixedLossScaleUpdateCell
|
||
mindspore.nn.ForwardValueAndGrad
|
||
mindspore.nn.PipelineCell
|
||
mindspore.nn.TrainOneStepCell
|
||
mindspore.nn.TrainOneStepWithLossScaleCell
|
||
mindspore.nn.WithEvalCell
|
||
mindspore.nn.WithLossCell
|
||
|
||
数学函数
|
||
--------
|
||
|
||
.. cnmsplatformautosummary::
|
||
:toctree: nn
|
||
|
||
mindspore.nn.Moments
|
||
|
||
评估指标
|
||
--------
|
||
|
||
.. cnmsautosummary::
|
||
:toctree: nn
|
||
|
||
mindspore.nn.Accuracy
|
||
mindspore.nn.F1
|
||
mindspore.nn.Fbeta
|
||
mindspore.nn.get_metric_fn
|
||
mindspore.nn.Loss
|
||
mindspore.nn.MAE
|
||
mindspore.nn.Metric
|
||
mindspore.nn.MSE
|
||
mindspore.nn.names
|
||
mindspore.nn.Precision
|
||
mindspore.nn.Recall
|
||
mindspore.nn.rearrange_inputs
|
||
mindspore.nn.Top1CategoricalAccuracy
|
||
mindspore.nn.Top5CategoricalAccuracy
|
||
mindspore.nn.TopKCategoricalAccuracy
|
||
|
||
动态学习率
|
||
-----------
|
||
|
||
LearningRateSchedule类
|
||
^^^^^^^^^^^^^^^^^^^^^^^
|
||
|
||
本模块中的动态学习率都是LearningRateSchedule的子类，将LearningRateSchedule的实例传递给优化器。在训练过程中，优化器以当前step为输入调用该实例，得到当前的学习率。
|
||
|
||
.. code-block::
|
||
|
||
import mindspore.nn as nn
|
||
|
||
min_lr = 0.01
|
||
max_lr = 0.1
|
||
decay_steps = 4
|
||
cosine_decay_lr = nn.CosineDecayLR(min_lr, max_lr, decay_steps)
|
||
|
||
net = Net()
|
||
optim = nn.Momentum(net.trainable_params(), learning_rate=cosine_decay_lr, momentum=0.9)
|
||
|
||
.. cnmsplatformautosummary::
|
||
:toctree: nn
|
||
|
||
mindspore.nn.CosineDecayLR
|
||
mindspore.nn.ExponentialDecayLR
|
||
mindspore.nn.InverseDecayLR
|
||
mindspore.nn.NaturalExpDecayLR
|
||
mindspore.nn.PolynomialDecayLR
|
||
mindspore.nn.WarmUpLR
|
||
|
||
Dynamic LR函数
|
||
^^^^^^^^^^^^^^
|
||
|
||
本模块中的动态学习率都是function，调用function并将结果传递给优化器。在训练过程中，优化器将result[current step]作为当前学习率。
|
||
|
||
.. code-block::
|
||
|
||
import mindspore.nn as nn
|
||
|
||
min_lr = 0.01
|
||
max_lr = 0.1
|
||
total_step = 6
|
||
step_per_epoch = 1
|
||
decay_epoch = 4
|
||
|
||
lr= nn.cosine_decay_lr(min_lr, max_lr, total_step, step_per_epoch, decay_epoch)
|
||
|
||
net = Net()
|
||
optim = nn.Momentum(net.trainable_params(), learning_rate=lr, momentum=0.9)
|
||
|
||
.. cnmsautosummary::
|
||
:toctree: nn
|
||
|
||
mindspore.nn.cosine_decay_lr
|
||
mindspore.nn.exponential_decay_lr
|
||
mindspore.nn.inverse_decay_lr
|
||
mindspore.nn.natural_exp_decay_lr
|
||
mindspore.nn.piecewise_constant_lr
|
||
mindspore.nn.polynomial_decay_lr
|
||
mindspore.nn.warmup_lr
|