tinyms.layers

Layer module contains pre-defined building blocks or computing units to construct neural networks.

The high-level components (Layers) used to construct the neural network.

class tinyms.layers.Layer(auto_prefix=True, flags=None)[源代码]

Base class for all neural networks.

A ‘Layer’ could be a single neural network layer, such as conv2d, relu, batch_norm, etc. or a composition of cells to constructing a network.

注解

In general, the autograd algorithm will automatically generate the implementation of the gradient function, but if back-propagation(bprop) method is implemented, the gradient function will be replaced by the bprop. The bprop implementation will receive a Tensor dout containing the gradient of the loss w.r.t. the output, and a Tensor out containing the forward result. The bprop needs to compute the gradient of the loss w.r.t. the inputs, gradient of the loss w.r.t. Parameter variables are not supported currently. The bprop method must contain the self parameter.

参数

auto_prefix (bool) – Recursively generate namespaces. Default: True.

实际案例

>>> from tinyms import layers, primitives as P
>>>
>>> class MyNet(layers.Layer):
...    def __init__(self):
...        super(MyNet, self).__init__()
...        self.relu = P.ReLU()
...
...    def construct(self, x):
...        return self.relu(x)
property bprop_debug

Get whether cell custom bprop debug is enabled.

cast_param(param)

Cast parameter according to auto mix precision level in pynative mode.

参数

param (Parameter) – The parameter to cast.

cells()

Returns an iterator over immediate cells.

cells_and_names(cells=None, name_prefix='')

Returns an iterator over all cells in the network.

Includes the cell’s name and itself.

参数

  • cells (str) – Cells to iterate over. Default: None.

  • name_prefix (str) – Namespace. Default: ‘’.

实际案例

>>> n = Net()
>>> names = []
>>> for m in n.cells_and_names():
...     if m[0]:
...         names.append(m[0])
compile(*inputs)

Compiles cell.

参数

inputs (tuple) – Input parameters.

compile_and_run(*inputs)

Compiles and runs cell.

参数

inputs (tuple) – Input parameters.

返回

Object, the result of executing.

construct(*inputs, **kwargs)

Defines the computation to be performed. This method must be overridden by all subclasses.

返回

Tensor, returns the computed result.

exec_checkpoint_graph()

Executes saving checkpoint graph operation.

extend_repr()

Sets the extended representation of the Cell.

To print customized extended information, re-implement this method in your own cells.

generate_scope()

Generate the scope for each cell object in the network.

get_func_graph_proto()

Return graph binary proto.

get_parameters(expand=True)

Returns an iterator over cell parameters.

Yields parameters of this cell. If expand is True, yield parameters of this cell and all subcells.

参数

expand (bool) – If true, yields parameters of this cell and all subcells. Otherwise, only yield parameters that are direct members of this cell. Default: True.

实际案例

>>> net = Net()
>>> parameters = []
>>> for item in net.get_parameters():
...     parameters.append(item)
get_scope()

Returns the scope of a cell object in one network.

init_parameters_data(auto_parallel_mode=False)

Initialize all parameters and replace the original saved parameters in cell.

注解

trainable_params() and other similar interfaces may return different parameter instance after init_parameters_data, do not save these result.

参数

auto_parallel_mode (bool) – If running in auto_parallel_mode.

返回

Dict[Parameter, Parameter], returns a dict of original parameter and replaced parameter.

insert_child_to_cell(child_name, child_cell)

Adds a child cell to the current cell with a given name.

参数

  • child_name (str) – Name of the child cell.

  • child_cell (Cell) – The child cell to be inserted.

引发

  • KeyError – Child Cell’s name is incorrect or duplicated with the other child name.

  • TypeError – Child Cell’s type is incorrect.

insert_param_to_cell(param_name, param, check_name=True)

Adds a parameter to the current cell.

Inserts a parameter with given name to the cell. Please refer to the usage in source code of mindspore.nn.Cell.__setattr__.

参数

  • param_name (str) – Name of the parameter.

  • param (Parameter) – Parameter to be inserted to the cell.

  • check_name (bool) – Determines whether the name input is compatible. Default: True.

引发

  • KeyError – If the name of parameter is null or contains dot.

  • AttributeError – If user did not call init() first.

  • TypeError – If the type of parameter is not Parameter.

load_parameter_slice(params)

Replace parameters with sliced tensors by parallel strategies.

Please refer to the usage in source code of mindspore.common._Executor.compile.

参数

params (dict) – The parameters dictionary used for initializing the data graph.

name_cells()

Returns an iterator over all cells in the network.

Include name of the cell and cell itself.

property param_prefix

Param prefix is the prefix of current cell’s direct child parameter.

parameters_and_names(name_prefix='', expand=True)

Returns an iterator over cell parameters.

Includes the parameter’s name and itself.

参数

  • name_prefix (str) – Namespace. Default: ‘’.

  • expand (bool) – If true, yields parameters of this cell and all subcells. Otherwise, only yield parameters that are direct members of this cell. Default: True.

实际案例

>>> n = Net()
>>> names = []
>>> for m in n.parameters_and_names():
...     if m[0]:
...         names.append(m[0])
parameters_dict(recurse=True)

Gets parameters dictionary.

Gets the parameters dictionary of this cell.

参数

recurse (bool) – Whether contains the parameters of subcells. Default: True.

返回

OrderedDict, return parameters dictionary.

recompute(mode=True)

Set the cell recomputed. All the primitive in the cell will be set recomputed. If a primitive set recomputed feeds into some backward nodes for computing gradient, rather than storing the intermediate activation computed in forward pass, we will recompute it in backward pass.

注解

  • If the computation involves something like randomization or global variable, the equivalence is not guaranteed currently.

  • If the recompute api of a primitive in this cell is also called, the recompute mode of this primitive is subject to the recompute api of the primitive.

参数

mode (bool) – Specifies whether the cell is recomputed. Default: True.

register_backward_hook(fn)

Set the cell backward hook function. Note that this function is only supported in Pynative Mode.

注解

fn must be defined as the following code. cell_name is the name of registered cell. grad_input is gradient passed to the cell. grad_output is the gradient computed and passed to the next cell or primitive, which may be modified and returned. hook_fn(cell_name, grad_input, grad_output) -> Tensor or None.

参数

fn (function) – Specifies the hook function with grad as input.

remove_redundant_parameters()

Remove the redundant parameters

set_auto_parallel()

Set the cell to auto parallel mode.

注解

If a cell needs to use the auto parallel or semi auto parallel mode for training, evaluation or prediction, this interface needs to be called by the cell.

set_broadcast_flag(mode=True)

Set the cell to data_parallel mode.

The cell can be accessed as an attribute using the given name.

参数

mode (bool) – Specifies whether the model is data_parallel. Default: True.

set_comm_fusion(fusion_type, recurse=True)

Set comm_fusion for all the parameters in the Net. Please refer to the description of mindspore.common.parameter.comm_fusion.

注解

The value of attribute will be overwritten when the function is called multiply.

参数

  • fusion_type (int) – The value of comm_fusion.

  • recurse (bool) – Whether sets the trainable parameters of subcells. Default: True.

set_grad(requires_grad=True)

Sets the cell flag for gradient. In pynative mode, this parameter specifies whether the network require gradients. If True, the backward network needed to compute the gradients will be generated when the forward network is executed.

参数

requires_grad (bool) – Specifies if the net need to grad, if it is True, cell will construct backward network in pynative mode. Default: True.

set_parallel_input_with_inputs(*inputs)

Slice inputs tensors by parallel strategies, and set the sliced inputs to _parallel_input_run

参数

inputs (tuple) – inputs of construct method.

set_param_ps(recurse=True, init_in_server=False)

Set whether the trainable parameters are updated by parameter server and whether the trainable parameters are initialized on server.

注解

It only works when a running task is in the parameter server mode.

参数

  • recurse (bool) – Whether sets the trainable parameters of subcells. Default: True.

  • init_in_server (bool) – Whether trainable parameters updated by parameter server are initialized on server. Default: False.

set_train(mode=True)

Sets the cell to training mode.

The cell itself and all children cells will be set to training mode. Layers that have different constructions for training and predicting, such as BatchNorm, will distinguish between the branches by this attribute. If set to True, the training branch will be executed, otherwise another branch.

参数

mode (bool) – Specifies whether the model is training. Default: True.

to_float(dst_type)

Add cast on all inputs of cell and child cells to run with certain float type.

If dst_type is mindspore.dtype.float16, all the inputs of Cell including input, Parameter, Tensor as const will be cast to float16. Please refer to the usage in source code of mindspore.train.amp.build_train_network.

注解

Multiple calls will overwrite.

参数

dst_type (mindspore.dtype) – Transfer Cell to Run with dst_type. dst_type can be mindspore.dtype.float16 or mindspore.dtype.float32.

引发

ValueError – If dst_type is not float32 nor float16.

trainable_params(recurse=True)

Returns all trainable parameters.

Returns a list of all trainable parameters.

参数

recurse (bool) – Whether contains the trainable parameters of subcells. Default: True.

返回

List, the list of trainable parameters.

untrainable_params(recurse=True)

Returns all untrainable parameters.

Returns a list of all untrainable parameters.

参数

recurse (bool) – Whether contains the untrainable parameters of subcells. Default: True.

返回

List, the list of untrainable parameters.

update_cell_prefix()

Update the all child cells’ self.param_prefix.

After being invoked, it can get all the cell’s children’s name prefix by ‘_param_prefix’.

update_cell_type(cell_type)

The current cell type is updated when a quantization aware training network is encountered.

After being invoked, it can set the cell type to ‘cell_type’.

update_parameters_name(prefix='', recurse=True)

Updates the names of parameters with given prefix string.

Adds the given prefix to the names of parameters.

参数

  • prefix (str) – The prefix string.

  • recurse (bool) – Whether contains the parameters of subcells. Default: True.

Layer module contains pre-defined building blocks or computing units to construct neural networks.

The high-level components (Layers) used to construct the neural network.

class tinyms.layers.SequentialLayer(*args)[源代码]

Sequential layer container.

A list of Layers will be added to it in the order they are passed in the constructor. Alternatively, an ordered dict of cells can also be passed in.

参数

args (Union[list, OrderedDict]) – List of subclass of Layer.

引发

TypeError – If the type of the argument is not list or OrderedDict.

Inputs:

  • input (Tensor) - Tensor with shape according to the first Cell in the sequence.

Outputs:

Tensor, the output Tensor with shape depending on the input and defined sequence of Layers.

实际案例

>>> import tinyms as ts
>>> from tinyms.layers import SequentialLayer, Conv2d, ReLU
>>>
>>> seq_layer = SequentialLayer([Conv2d(3, 2, 3, pad_mode='valid', weight_init="ones"), ReLU()])
>>> x = ts.ones([1, 3, 4, 4])
>>> print(seq_layer(x))
[[[[27. 27.]
   [27. 27.]]
  [[27. 27.]
   [27. 27.]]]]
class tinyms.layers.LayerList(*args, **kwargs)[源代码]

Holds Layers in a list.

LayerList can be used like a regular Python list, support ‘__getitem__’, ‘__setitem__’, ‘__delitem__’, ‘__len__’, ‘__iter__’ and ‘__iadd__’, but layers it contains are properly registered, and will be visible by all Layer methods.

参数

args (list, optional) – List of subclass of Layer.

实际案例

>>> from tinyms.layers import LayerList, Conv2d, BatchNorm2d, ReLU
>>>
>>> conv = nn.Conv2d(100, 20, 3)
>>> layers = LayerList([BatchNorm2d(20)])
>>> layers.insert(0, Conv2d(100, 20, 3))
>>> layers.append(ReLU())
>>> layers
LayerList<
  (0): Conv2d<input_channels=100, ..., bias_init=None>
  (1): BatchNorm2d<num_features=20, ..., moving_variance=Parameter (name=variance)>
  (2): ReLU<>
  >
class tinyms.layers.Softmax(axis=-1)[源代码]

Softmax activation function.

Applies the Softmax function to an n-dimensional input Tensor.

The input is a Tensor of logits transformed with exponential function and then normalized to lie in range [0, 1] and sum up to 1.

Softmax is defined as:

\[\text{softmax}(x_{i}) = \frac{\exp(x_i)}{\sum_{j=0}^{n-1}\exp(x_j)},\]

where \(x_{i}\) is the \(i\)-th slice in the given dimension of the input Tensor.

参数

axis (Union[int, tuple[int]]) – The axis to apply Softmax operation, -1 means the last dimension. Default: -1.

Inputs:

  • x (Tensor) - The input of Softmax with data type of float16 or float32.

Outputs:

Tensor, which has the same type and shape as x with values in the range[0,1].

引发

  • TypeError – If axis is neither an int nor a tuple.

  • TypeError – If dtype of x is neither float16 nor float32.

  • ValueError – If axis is a tuple whose length is less than 1.

  • ValueError – If axis is a tuple whose elements are not all in range [-len(x), len(x)).

Supported Platforms:

Ascend GPU CPU

实际案例

>>> input_x = Tensor(np.array([-1, -2, 0, 2, 1]), mindspore.float16)
>>> softmax = nn.Softmax()
>>> output = softmax(input_x)
>>> print(output)
[0.03168 0.01166 0.0861  0.636   0.2341 ]
class tinyms.layers.LogSoftmax(axis=-1)[源代码]

LogSoftmax activation function.

Applies the LogSoftmax function to n-dimensional input tensor.

The input is transformed by the Softmax function and then by the log function to lie in range[-inf,0).

Logsoftmax is defined as:

\[\text{logsoftmax}(x_i) = \log \left(\frac{\exp(x_i)}{\sum_{j=0}^{n-1} \exp(x_j)}\right),\]

where \(x_{i}\) is the \(i\)-th slice in the given dimension of the input Tensor.

参数

axis (int) – The axis to apply LogSoftmax operation, -1 means the last dimension. Default: -1.

Inputs:

  • x (Tensor) - The input of LogSoftmax, with float16 or float32 data type.

Outputs:

Tensor, which has the same type and shape as the input as x with values in the range[-inf,0).

引发

  • TypeError – If axis is not an int.

  • TypeError – If dtype of x is neither float16 nor float32.

  • ValueError – If axis is not in range [-len(x), len(x)).

Supported Platforms:

Ascend GPU

实际案例

>>> input_x = Tensor(np.array([[-1.0, 4.0, -8.0], [2.0, -5.0, 9.0]]), mindspore.float32)
>>> log_softmax = nn.LogSoftmax()
>>> output = log_softmax(input_x)
>>> print(output)
[[-5.00672150e+00 -6.72150636e-03 -1.20067215e+01]
 [-7.00091219e+00 -1.40009127e+01 -9.12250078e-04]]
class tinyms.layers.ReLU[源代码]

Rectified Linear Unit activation function.

Applies the rectified linear unit function element-wise.

\[\text{ReLU}(x) = (x)^+ = \max(0, x),\]

It returns element-wise \(\max(0, x)\), specially, the neurons with the negative output will be suppressed and the active neurons will stay the same.

The picture about ReLU looks like this ReLU.

Inputs:

  • input_data (Tensor) - The input of ReLU.

Outputs:

Tensor, with the same type and shape as the input_data.

引发

TypeError – If dtype of input_data is not a number.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> input_x = Tensor(np.array([-1, 2, -3, 2, -1]), mindspore.float16)
>>> relu = nn.ReLU()
>>> output = relu(input_x)
>>> print(output)
[0. 2. 0. 2. 0.]
class tinyms.layers.ReLU6[源代码]

Compute ReLU6 activation function.

ReLU6 is similar to ReLU with a upper limit of 6, which if the inputs are greater than 6, the outputs will be suppressed to 6. It computes element-wise as

\[\min(\max(0, x), 6).\]

The input is a Tensor of any valid shape.

Inputs:

  • input_data (Tensor) - The input of ReLU6 with data type of float16 or float32.

Outputs:

Tensor, which has the same type as input_data.

引发

TypeError – If dtype of input_data is neither float16 nor float32.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> input_x = Tensor(np.array([-1, -2, 0, 2, 1]), mindspore.float16)
>>> relu6 = nn.ReLU6()
>>> output = relu6(input_x)
>>> print(output)
[0. 0. 0. 2. 1.]
class tinyms.layers.Tanh[源代码]

Tanh activation function.

Applies the Tanh function element-wise, returns a new tensor with the hyperbolic tangent of the elements of input, The input is a Tensor with any valid shape.

Tanh function is defined as:

\[tanh(x_i) = \frac{\exp(x_i) - \exp(-x_i)}{\exp(x_i) + \exp(-x_i)} = \frac{\exp(2x_i) - 1}{\exp(2x_i) + 1},\]

where \(x_i\) is an element of the input Tensor.

Inputs:

  • input_data (Tensor) - The input of Tanh with data type of float16 or float32.

Outputs:

Tensor, with the same type and shape as the input_data.

引发

TypeError – If dtype of input_data is neither float16 nor float32.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> input_x = Tensor(np.array([1, 2, 3, 2, 1]), mindspore.float16)
>>> tanh = nn.Tanh()
>>> output = tanh(input_x)
>>> print(output)
[0.7617 0.964  0.995  0.964  0.7617]
class tinyms.layers.GELU[源代码]

Gaussian error linear unit activation function.

Applies GELU function to each element of the input. The input is a Tensor with any valid shape.

GELU is defined as:

\[GELU(x_i) = x_i*P(X < x_i),\]

where \(P\) is the cumulative distribution function of standard Gaussian distribution and \(x_i\) is the element of the input.

The picture about GELU looks like this GELU.

Inputs:

  • input_data (Tensor) - The input of GELU with data type of float16 or float32.

Outputs:

Tensor, with the same type and shape as the input_data.

引发

TypeError – If dtype of input_data is neither float16 nor float32.

Supported Platforms:

Ascend GPU

实际案例

>>> input_x = Tensor(np.array([[-1.0, 4.0, -8.0], [2.0, -5.0, 9.0]]), mindspore.float32)
>>> gelu = nn.GELU()
>>> output = gelu(input_x)
>>> print(output)
[[-1.5880802e-01  3.9999299e+00 -3.1077917e-21]
 [ 1.9545976e+00 -2.2918017e-07  9.0000000e+00]]
class tinyms.layers.FastGelu[源代码]

Fast Gaussian error linear unit activation function.

Applies FastGelu function to each element of the input. The input is a Tensor with any valid shape.

FastGelu is defined as:

\[FastGelu(x_i) = \frac {x_i} {1 + \exp(-1.702 * \left| x_i \right|)} * \exp(0.851 * (x_i - \left| x_i \right|))\]

where \(x_i\) is the element of the input.

Inputs:

  • input_data (Tensor) - The input of FastGelu with data type of float16 or float32.

Outputs:

Tensor, with the same type and shape as the input_data.

引发

TypeError – If dtype of input_data is neither float16 nor float32.

Supported Platforms:

Ascend

实际案例

>>> input_x = Tensor(np.array([[-1.0, 4.0, -8.0], [2.0, -5.0, 9.0]]), mindspore.float32)
>>> fast_gelu = nn.FastGelu()
>>> output = fast_gelu(input_x)
>>> print(output)
[[-1.5420423e-01  3.9955850e+00 -9.7664279e-06]
 [ 1.9356586e+00 -1.0070159e-03  8.9999981e+00]]
class tinyms.layers.Sigmoid[源代码]

Sigmoid activation function.

Applies sigmoid-type activation element-wise.

Sigmoid function is defined as:

\[\text{sigmoid}(x_i) = \frac{1}{1 + \exp(-x_i)},\]

where \(x_i\) is the element of the input.

The picture about Sigmoid looks like this Sigmoid.

Inputs:

  • input_data (Tensor) - The input of Sigmoid with data type of float16 or float32.

Outputs:

Tensor, with the same type and shape as the input_data.

引发

TypeError – If dtype of input_data is neither float16 nor float32.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> input_x = Tensor(np.array([-1, -2, 0, 2, 1]), mindspore.float16)
>>> sigmoid = nn.Sigmoid()
>>> output = sigmoid(input_x)
>>> print(output)
[0.2688  0.11914 0.5     0.881   0.7305 ]
class tinyms.layers.PReLU(channel=1, w=0.25)[源代码]

PReLU activation function.

Applies the PReLU function element-wise.

PReLU is defined as:

\[prelu(x_i)= \max(0, x_i) + w * \min(0, x_i),\]

where \(x_i\) is an element of an channel of the input.

Here \(w\) is a learnable parameter with a default initial value 0.25. Parameter \(w\) has dimensionality of the argument channel. If called without argument channel, a single parameter \(w\) will be shared across all channels.

The picture about PReLU looks like this PReLU.

参数

  • channel (int) – The dimension of input. Default: 1.

  • w (Union[float, list, Tensor]) – The initial value of w. Default: 0.25.

Inputs:

  • input_data (Tensor) - The input of PReLU with data type of float16 or float32.

Outputs:

Tensor, with the same type and shape as the input_data.

引发

  • TypeError – If channel is not an int.

  • TypeError – If w is not one of float, list, Tensor.

  • TypeError – If dtype of input_data is neither float16 nor float32.

  • ValueError – If channel is less than 1.

  • ValueError – If length of shape of input_data is equal to 1.

Supported Platforms:

Ascend

实际案例

>>> input_x = Tensor(np.array([[[[0.1, 0.6], [0.9, 0.9]]]]), mindspore.float32)
>>> prelu = nn.PReLU()
>>> output = prelu(input_x)
>>> print(output)
[[[[0.1 0.6]
   [0.9 0.9]]]]
tinyms.layers.get_activation(name)[源代码]

Gets the activation function.

参数

name (str) – The name of the activation function.

返回

Function, the activation function.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> sigmoid = nn.get_activation('sigmoid')
class tinyms.layers.LeakyReLU(alpha=0.2)[源代码]

Leaky ReLU activation function.

LeakyReLU is similar to ReLU, but LeakyReLU has a slope that makes it not equal to 0 at x < 0. The activation function is defined as:

\[\text{leaky_relu}(x) = \begin{cases}x, &\text{if } x \geq 0; \cr \text{alpha} * x, &\text{otherwise.}\end{cases}\]

See https://ai.stanford.edu/~amaas/papers/relu_hybrid_icml2013_final.pdf

参数

alpha (Union[int, float]) – Slope of the activation function at x < 0. Default: 0.2.

Inputs:

  • input_x (Tensor) - The input of LeakyReLU.

Outputs:

Tensor, has the same type and shape as the input_x.

引发

TypeError – If alpha is not a float or an int.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> input_x = Tensor(np.array([[-1.0, 4.0, -8.0], [2.0, -5.0, 9.0]]), mindspore.float32)
>>> leaky_relu = nn.LeakyReLU()
>>> output = leaky_relu(input_x)
>>> print(output)
[[-0.2  4.  -1.6]
 [ 2.  -1.   9. ]]
class tinyms.layers.HSigmoid[源代码]

Hard sigmoid activation function.

Applies hard sigmoid activation element-wise. The input is a Tensor with any valid shape.

Hard sigmoid is defined as:

\[\text{hsigmoid}(x_{i}) = max(0, min(1, \frac{x_{i} + 3}{6})),\]

where \(x_{i}\) is the \(i\)-th slice in the given dimension of the input Tensor.

Inputs:

  • input_data (Tensor) - The input of HSigmoid, data type must be float16 or float32.

Outputs:

Tensor, with the same type and shape as the input_data.

引发

TypeError – If dtype of input_data is neither float16 nor float32.

Supported Platforms:

GPU CPU

实际案例

>>> input_x = Tensor(np.array([-1, -2, 0, 2, 1]), mindspore.float16)
>>> hsigmoid = nn.HSigmoid()
>>> result = hsigmoid(input_x)
>>> print(result)
[0.3333  0.1666  0.5  0.833  0.6665]
class tinyms.layers.HSwish[源代码]

Hard swish activation function.

Applies hswish-type activation element-wise. The input is a Tensor with any valid shape.

Hard swish is defined as:

\[\text{hswish}(x_{i}) = x_{i} * \frac{ReLU6(x_{i} + 3)}{6},\]

where \(x_{i}\) is the \(i\)-th slice in the given dimension of the input Tensor.

Inputs:

  • input_data (Tensor) - The input of HSwish, data type must be float16 or float32.

Outputs:

Tensor, with the same type and shape as the input_data.

引发

TypeError – If dtype of input_data is neither float16 nor float32.

Supported Platforms:

GPU CPU

实际案例

>>> input_x = Tensor(np.array([-1, -2, 0, 2, 1]), mindspore.float16)
>>> hswish = nn.HSwish()
>>> result = hswish(input_x)
>>> print(result)
[-0.3333  -0.3333  0  1.666  0.6665]
class tinyms.layers.ELU(alpha=1.0)[源代码]

Exponential Linear Uint activation function.

Applies the exponential linear unit function element-wise. The activation function is defined as:

\[E_{i} = \begin{cases} x, &\text{if } x \geq 0; \cr \text{alpha} * (\exp(x_i) - 1), &\text{otherwise.} \end{cases}\]

The picture about ELU looks like this ELU.

参数

alpha (float) – The coefficient of negative factor whose type is float. Default: 1.0.

Inputs:

  • input_data (Tensor) - The input of ELU with data type of float16 or float32.

Outputs:

Tensor, with the same type and shape as the input_data.

引发

  • TypeError – If alpha is not a float.

  • TypeError – If dtype of input_data is neither float16 nor float32.

  • ValueError – If alpha is not equal to 1.0.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> input_x = Tensor(np.array([-1, -2, 0, 2, 1]), mindspore.float32)
>>> elu = nn.ELU()
>>> result = elu(input_x)
>>> print(result)
[-0.63212055  -0.86466473  0.  2.  1.]
class tinyms.layers.LogSigmoid[源代码]

Logsigmoid activation function.

Applies logsigmoid activation element-wise. The input is a Tensor with any valid shape.

Logsigmoid is defined as:

\[\text{logsigmoid}(x_{i}) = log(\frac{1}{1 + \exp(-x_i)}),\]

where \(x_{i}\) is the element of the input.

Inputs:

  • input_data (Tensor) - The input of LogSigmoid with data type of float16 or float32.

Outputs:

Tensor, with the same type and shape as the input_data.

引发

TypeError – If dtype of input_data is neither float16 nor float32.

Supported Platforms:

Ascend GPU

实际案例

>>> net = nn.LogSigmoid()
>>> input_x = Tensor(np.array([1.0, 2.0, 3.0]), mindspore.float32)
>>> output = net(input_x)
>>> print(output)
[-0.31326166 -0.12692806 -0.04858734]
class tinyms.layers.BatchNorm1d(num_features, eps=1e-05, momentum=0.9, affine=True, gamma_init='ones', beta_init='zeros', moving_mean_init='zeros', moving_var_init='ones', use_batch_statistics=None)[源代码]

Batch normalization layer over a 2D input.

Batch Normalization is widely used in convolutional networks. This layer applies Batch Normalization over a 2D input (a mini-batch of 1D inputs) to reduce internal covariate shift as described in the paper Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. It rescales and recenters the feature using a mini-batch of data and the learned parameters which can be described in the following formula.

\[y = \frac{x - \mathrm{E}[x]}{\sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta\]

注解

The implementation of BatchNorm is different in graph mode and pynative mode, therefore the mode is not recommended to be changed after net was initialized.

参数

  • num_features (int) – C from an expected input of size (N, C).

  • eps (float) – A value added to the denominator for numerical stability. Default: 1e-5.

  • momentum (float) – A floating hyperparameter of the momentum for the running_mean and running_var computation. Default: 0.9.

  • affine (bool) – A bool value. When set to True, gamma and beta can be learned. Default: True.

  • gamma_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the gamma weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘ones’.

  • beta_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the beta weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘zeros’.

  • moving_mean_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the moving mean. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘zeros’.

  • moving_var_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the moving variance. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘ones’.

  • use_batch_statistics (bool) – If true, use the mean value and variance value of current batch data. If false, use the mean value and variance value of specified value. If None, the training process will use the mean and variance of current batch data and track the running mean and variance, the evaluation process will use the running mean and variance. Default: None.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in})\).

Outputs:

Tensor, the normalized, scaled, offset tensor, of shape \((N, C_{out})\).

Supported Platforms:

Ascend

引发

实际案例

>>> net = nn.BatchNorm1d(num_features=4)
>>> np.random.seed(0)
>>> input = Tensor(np.random.randint(0, 255, [2, 4]), mindspore.float32)
>>> output = net(input)
>>> print(output)
[[171.99915   46.999763  116.99941  191.99904 ]
 [ 66.999664 250.99875   194.99902  102.99948 ]]
class tinyms.layers.BatchNorm2d(num_features, eps=1e-05, momentum=0.9, affine=True, gamma_init='ones', beta_init='zeros', moving_mean_init='zeros', moving_var_init='ones', use_batch_statistics=None, data_format='NCHW')[源代码]

Batch normalization layer over a 4D input.

Batch Normalization is widely used in convolutional networks. This layer applies Batch Normalization over a 4D input (a mini-batch of 2D inputs with additional channel dimension) to avoid internal covariate shift as described in the paper Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. It rescales and recenters the feature using a mini-batch of data and the learned parameters which can be described in the following formula.

\[y = \frac{x - \mathrm{E}[x]}{\sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta\]

注解

The implementation of BatchNorm is different in graph mode and pynative mode, therefore that mode can not be changed after net was initialized. Note that the formula for updating the running_mean and running_var is \(\hat{x}_\text{new} = (1 - \text{momentum}) \times x_t + \text{momentum} \times \hat{x}\), where \(\hat{x}\) is the estimated statistic and \(x_t\) is the new observed value.

参数

  • num_features (int) – C from an expected input of size (N, C, H, W).

  • eps (float) – A value added to the denominator for numerical stability. Default: 1e-5.

  • momentum (float) – A floating hyperparameter of the momentum for the running_mean and running_var computation. Default: 0.9.

  • affine (bool) – A bool value. When set to True, gamma and beta can be learned. Default: True.

  • gamma_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the gamma weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘ones’.

  • beta_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the beta weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘zeros’.

  • moving_mean_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the moving mean. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘zeros’.

  • moving_var_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the moving variance. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘ones’.

  • use_batch_statistics (bool) – If true, use the mean value and variance value of current batch data. If false, use the mean value and variance value of specified value. If None, the training process will use the mean and variance of current batch data and track the running mean and variance, the evaluation process will use the running mean and variance. Default: None.

  • data_format (str) – The optional value for data format, is ‘NHWC’ or ‘NCHW’. Default: ‘NCHW’.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in}, H_{in}, W_{in})\).

Outputs:

Tensor, the normalized, scaled, offset tensor, of shape \((N, C_{out}, H_{out}, W_{out})\).

引发

  • TypeError – If num_features is not an int.

  • TypeError – If eps is not a float.

  • ValueError – If num_features is less than 1.

  • ValueError – If momentum is not in range [0, 1].

  • ValueError – If data_format is neither ‘NHWC’ not ‘NCHW’.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> net = nn.BatchNorm2d(num_features=3)
>>> np.random.seed(0)
>>> input = Tensor(np.random.randint(0, 255, [1, 3, 2, 2]), mindspore.float32)
>>> output = net(input)
>>> print(output)
[[[[171.99915   46.999763 ]
   [116.99941  191.99904  ]]
  [[ 66.999664 250.99875  ]
   [194.99902  102.99948  ]]
  [[  8.999955 210.99895  ]
   [ 20.999895 241.9988   ]]]]
class tinyms.layers.BatchNorm3d(num_features, eps=1e-05, momentum=0.9, affine=True, gamma_init='ones', beta_init='zeros', moving_mean_init='zeros', moving_var_init='ones', use_batch_statistics=None, data_format='NCDHW')[源代码]

Batch normalization layer over a 5D input.

Batch Normalization is widely used in convolutional networks. This layer applies Batch Normalization over a 5D input (a mini-batch of 3D inputs with additional channel dimension) to avoid internal covariate shift.

\[y = \frac{x - \mathrm{E}[x]}{\sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta\]

注解

The implementation of BatchNorm is different in graph mode and pynative mode, therefore that mode can not be changed after net was initialized. Note that the formula for updating the running_mean and running_var is \(\hat{x}_\text{new} = (1 - \text{momentum}) \times x_t + \text{momentum} \times \hat{x}\), where \(\hat{x}\) is the estimated statistic and \(x_t\) is the new observed value.

参数

  • num_features (int) – C from an expected input of size (N, C, D, H, W).

  • eps (float) – A value added to the denominator for numerical stability. Default: 1e-5.

  • momentum (float) – A floating hyperparameter of the momentum for the running_mean and running_var computation. Default: 0.9.

  • affine (bool) – A bool value. When set to True, gamma and beta can be learned. Default: True.

  • gamma_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the gamma weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘ones’.

  • beta_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the beta weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘zeros’.

  • moving_mean_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the moving mean. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘zeros’.

  • moving_var_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the moving variance. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘ones’.

  • use_batch_statistics (bool) – If true, use the mean value and variance value of current batch data. If false, use the mean value and variance value of specified value. If None, the training process will use the mean and variance of current batch data and track the running mean and variance, the evaluation process will use the running mean and variance. Default: None.

  • data_format (str) – The optional value for data format is ‘NCDHW’. Default: ‘NCDHW’.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in}, D_{in}, H_{in}, W_{in})\).

Outputs:

Tensor, the normalized, scaled, offset tensor, of shape \((N, C_{out}, D_{out},H_{out}, W_{out})\).

引发
Supported Platforms:

Ascend GPU CPU

实际案例

>>> net = nn.BatchNorm3d(num_features=3)
>>> np.random.seed(0)
>>> input = Tensor(np.random.randint(0, 255, [16, 3, 10, 32, 32]), mindspore.float32)
>>> output = net(input)
>>> print(output.shape)
(16, 3, 10, 32, 32)
class tinyms.layers.LayerNorm(normalized_shape, begin_norm_axis=-1, begin_params_axis=-1, gamma_init='ones', beta_init='zeros', epsilon=1e-07)[源代码]

Applies Layer Normalization over a mini-batch of inputs.

Layer normalization is widely used in recurrent neural networks. It applies normalization on a mini-batch of inputs for each single training case as described in the paper Layer Normalization. Unlike batch normalization, layer normalization performs exactly the same computation at training and testing time. It can be described using the following formula. It is applied across all channels and pixel but only one batch size.

\[y = \frac{x - \mathrm{E}[x]}{\sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta\]
参数

  • normalized_shape (Union(tuple[int], list[int]) – The normalization is performed over axis begin_norm_axis … R - 1.

  • begin_norm_axis (int) – The first normalization dimension: normalization will be performed along dimensions begin_norm_axis: rank(inputs), the value should be in [-1, rank(input)). Default: -1.

  • begin_params_axis (int) – The first parameter(beta, gamma)dimension: scale and centering parameters will have dimensions begin_params_axis: rank(inputs) and will be broadcast with the normalized inputs accordingly, the value should be in [-1, rank(input)). Default: -1.

  • gamma_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the gamma weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, ‘xavier_uniform’, ‘he_uniform’, etc. Default: ‘ones’.

  • beta_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the beta weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, ‘xavier_uniform’, ‘he_uniform’, etc. Default: ‘zeros’.

  • epsilon (float) – A value added to the denominator for numerical stability. Default: 1e-7.

Inputs:

  • input_x (Tensor) - The shape of ‘input_x’ is \((x_1, x_2, ..., x_R)\), and input_shape[begin_norm_axis:] is equal to normalized_shape.

Outputs:

Tensor, the normalized and scaled offset tensor, has the same shape and data type as the input_x.

引发

  • TypeError – If normalized_shape is neither a list nor tuple.

  • TypeError – If begin_norm_axis or begin_params_axis is not an int.

  • TypeError – If epsilon is not a float.

Supported Platforms:

Ascend GPU

实际案例

>>> x = Tensor(np.ones([20, 5, 10, 10]), mindspore.float32)
>>> shape1 = x.shape[1:]
>>> m = nn.LayerNorm(shape1,  begin_norm_axis=1, begin_params_axis=1)
>>> output = m(x).shape
>>> print(output)
(20, 5, 10, 10)
extend_repr()[源代码]

Display instance object as string.

class tinyms.layers.GroupNorm(num_groups, num_channels, eps=1e-05, affine=True, gamma_init='ones', beta_init='zeros')[源代码]

Group Normalization over a mini-batch of inputs.

Group normalization is widely used in recurrent neural networks. It applies normalization on a mini-batch of inputs for each single training case as described in the paper Group Normalization. Group normalization divides the channels into groups and computes within each group the mean and variance for normalization, and it performs very stable over a wide range of batch size. It can be described using the following formula.

\[y = \frac{x - \mathrm{E}[x]}{\sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta\]
参数

  • num_groups (int) – The number of groups to be divided along the channel dimension.

  • num_channels (int) – The number of channels per group.

  • eps (float) – A value added to the denominator for numerical stability. Default: 1e-5.

  • affine (bool) – A bool value, this layer will have learnable affine parameters when set to true. Default: True.

  • gamma_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the gamma weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, ‘xavier_uniform’, ‘he_uniform’, etc. Default: ‘ones’. If gamma_init is a Tensor, the shape must be [num_channels].

  • beta_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the beta weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, ‘xavier_uniform’, ‘he_uniform’, etc. Default: ‘zeros’. If beta_init is a Tensor, the shape must be [num_channels].

Inputs:

  • input_x (Tensor) - The input feature with shape [N, C, H, W].

Outputs:

Tensor, the normalized and scaled offset tensor, has the same shape and data type as the input_x.

引发

  • TypeError – If num_groups or num_channels is not an int.

  • TypeError – If eps is not a float.

  • TypeError – If affine is not a bool.

  • ValueError – If num_groups or num_channels is less than 1.

  • ValueError – If num_channels is not divided by num_groups.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> goup_norm_op = nn.GroupNorm(2, 2)
>>> x = Tensor(np.ones([1, 2, 4, 4], np.float32))
>>> output = goup_norm_op(x)
>>> print(output)
[[[[0. 0. 0. 0.]
   [0. 0. 0. 0.]
   [0. 0. 0. 0.]
   [0. 0. 0. 0.]]
  [[0. 0. 0. 0.]
   [0. 0. 0. 0.]
   [0. 0. 0. 0.]
   [0. 0. 0. 0.]]]]
extend_repr()[源代码]

Display instance object as string.

class tinyms.layers.GlobalBatchNorm(**kwargs)[源代码]

Global normalization layer over a N-dimension input.

Global Normalization is cross device synchronized batch normalization. The implementation of Batch Normalization only normalizes the data within each device. Global normalization will normalize the input within the group. It has been described in the paper Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. It rescales and recenters the feature using a mini-batch of data and the learned parameters which can be described in the following formula.

\[y = \frac{x - \mathrm{E}[x]}{\sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta\]

注解

Currently, GlobalBatchNorm only supports 2D and 4D inputs.

参数

  • num_features (int) – C from an expected input of size (N, C, H, W).

  • device_num_each_group (int) – The number of devices in each group. Default: 2.

  • eps (float) – A value added to the denominator for numerical stability. Default: 1e-5.

  • momentum (float) – A floating hyperparameter of the momentum for the running_mean and running_var computation. Default: 0.9.

  • affine (bool) – A bool value. When set to True, gamma and beta can be learned. Default: True.

  • gamma_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the gamma weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, ‘xavier_uniform’, ‘he_uniform’, etc. Default: ‘ones’.

  • beta_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the beta weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, ‘xavier_uniform’, ‘he_uniform’, etc. Default: ‘zeros’.

  • moving_mean_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the moving mean. The values of str refer to the function initializer including ‘zeros’, ‘ones’, ‘xavier_uniform’, ‘he_uniform’, etc. Default: ‘zeros’.

  • moving_var_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the moving variance. The values of str refer to the function initializer including ‘zeros’, ‘ones’, ‘xavier_uniform’, ‘he_uniform’, etc. Default: ‘ones’.

  • use_batch_statistics (bool) – If true, use the mean value and variance value of current batch data. If false, use the mean value and variance value of specified value. If None, training process will use the mean and variance of current batch data and track the running mean and variance, eval process will use the running mean and variance. Default: None.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in}, H_{in}, W_{in})\).

Outputs:

Tensor, the normalized, scaled, offset tensor, of shape \((N, C_{out}, H_{out}, W_{out})\).

引发

  • TypeError – If num_features or device_num_each_group is not an int.

  • TypeError – If eps is not a float.

  • ValueError – If num_features is less than 1.

  • ValueError – If momentum is not in range [0, 1].

  • ValueError – If device_num_each_group is less than 2.

Supported Platforms:

Ascend

实际案例

>>> # This example should be run with multiple processes.
>>> # Please refer to the tutorial > Distributed Training on mindspore.cn.
>>> import numpy as np
>>> from mindspore.communication import init
>>> from mindspore import context
>>> from mindspore.context import ParallelMode
>>> from mindspore import nn, Tensor
>>> from mindspore.common import dtype as mstype
>>>
>>> context.set_context(mode=context.GRAPH_MODE)
>>> init()
>>> context.reset_auto_parallel_context()
>>> context.set_auto_parallel_context(parallel_mode=ParallelMode.DATA_PARALLEL)
>>> np.random.seed(0)
>>> global_bn_op = nn.GlobalBatchNorm(num_features=3, device_num_each_group=2)
>>> input = Tensor(np.random.randint(0, 255, [1, 3, 2, 2]), mstype.float32)
>>> output = global_bn_op(input)
>>> print(output)
[[[[171.99915    46.999763]
   [116.99941   191.99904 ]]
  [[ 66.999664  250.99875 ]
   [194.99902   102.99948 ]]
  [[  8.999955  210.99895 ]
   [ 20.9999895 241.9988  ]]]]
class tinyms.layers.SyncBatchNorm(num_features, eps=1e-05, momentum=0.9, affine=True, gamma_init='ones', beta_init='zeros', moving_mean_init='zeros', moving_var_init='ones', use_batch_statistics=None, process_groups=None)[源代码]

Sync Batch normalization layer over a N-dimension input.

Sync Batch Normalization is cross device synchronized batch normalization. The implementation of Batch Normalization only normalizes the data within each device. Sync Batch normalization will normalize the input within the group. It has been described in the paper Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. It rescales and recenters the feature using a mini-batch of data and the learned parameters which can be described in the following formula.

\[y = \frac{x - \mathrm{E}[x]}{\sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta\]

注解

Currently, SyncBatchNorm only supports 2D and 4D inputs.

参数

  • num_features (int) – C from an expected input of size (N, C, H, W).

  • eps (float) – A value added to the denominator for numerical stability. Default: 1e-5.

  • momentum (float) – A floating hyperparameter of the momentum for the running_mean and running_var computation. Default: 0.9.

  • affine (bool) – A bool value. When set to True, gamma and beta can be learned. Default: True.

  • gamma_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the gamma weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, ‘xavier_uniform’, ‘he_uniform’, etc. Default: ‘ones’.

  • beta_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the beta weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, ‘xavier_uniform’, ‘he_uniform’, etc. Default: ‘zeros’.

  • moving_mean_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the moving mean. The values of str refer to the function initializer including ‘zeros’, ‘ones’, ‘xavier_uniform’, ‘he_uniform’, etc. Default: ‘zeros’.

  • moving_var_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the moving variance. The values of str refer to the function initializer including ‘zeros’, ‘ones’, ‘xavier_uniform’, ‘he_uniform’, etc. Default: ‘ones’.

  • use_batch_statistics (bool) – If true, use the mean value and variance value of current batch data. If false, use the mean value and variance value of specified value. If None, training process will use the mean and variance of current batch data and track the running mean and variance, eval process will use the running mean and variance. Default: None.

  • process_groups (list) – A list to divide devices into different sync groups, containing N subtraction lists. Each subtraction list contains int numbers identifying rank ids which need to be synchronized in the same group. All int values must be in [0, rank_size) and different from each other. Default: None, indicating synchronization across all devices.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in}, H_{in}, W_{in})\).

Outputs:

Tensor, the normalized, scaled, offset tensor, of shape \((N, C_{out}, H_{out}, W_{out})\).

引发

  • TypeError – If num_features is not an int.

  • TypeError – If eps is not a float.

  • TypeError – If process_groups is not a list.

  • ValueError – If num_features is less than 1.

  • ValueError – If momentum is not in range [0, 1].

  • ValueError – If rank_id in process_groups is not in range [0, rank_size).

Supported Platforms:

Ascend

实际案例

>>> # This example should be run with multiple processes.
>>> # Please refer to the tutorial > Distributed Training on mindspore.cn.
>>> import numpy as np
>>> from mindspore.communication import init
>>> from mindspore import context
>>> from mindspore.context import ParallelMode
>>> from mindspore import nn, Tensor
>>> from mindspore.common import dtype as mstype
>>>
>>> context.set_context(mode=context.GRAPH_MODE)
>>> init()
>>> context.reset_auto_parallel_context()
>>> context.set_auto_parallel_context(parallel_mode=ParallelMode.DATA_PARALLEL)
>>> np.random.seed(0)
>>> sync_bn_op = nn.SyncBatchNorm(num_features=3, process_groups=[[0, 1], [2, 3]])
>>> input = Tensor(np.random.randint(0, 255, [1, 3, 2, 2]), mstype.float32)
>>> output = sync_bn_op(input)
>>> print(output)
[[[[171.99915    46.999763]
   [116.99941   191.99904 ]]
  [[ 66.999664  250.99875 ]
   [194.99902   102.99948 ]]
  [[  8.999955  210.99895 ]
   [ 20.9999895 241.9988  ]]]]
class tinyms.layers.InstanceNorm2d(num_features, eps=1e-05, momentum=0.1, affine=True, gamma_init='ones', beta_init='zeros', moving_mean_init='zeros', moving_var_init='ones', use_batch_statistics=True)[源代码]

Instance normalization layer over a 4D input.

This layer applies Instance Normalization over a 4D input (a mini-batch of 2D inputs with additional channel dimension) as described in the paper Instance Normalization: The Missing Ingredient for Fast Stylization. It rescales and recenters the feature using a mini-batch of data and the learned parameters which can be described in the following formula.

\[y = \frac{x - \mathrm{E}[x]}{\sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta\]

gamma and beta are learnable parameter vectors of size num_features if affine is True. The standard-deviation is calculated via the biased estimator.

By default, this layer uses instance statistics computed from input data in both training and evaluation modes.

If use_batch_statistics is set to True, it means training phases, and this layer keeps running estimates of its computed mean and variance, which are then used for normalization during evaluation. The running estimates are kept with a default momentum of 0.1.

InstanceNorm2d and BatchNorm2d are very similar, but have some differences. InstanceNorm2d is applied on each channel of channeled data like RGB images, but BatchNorm2d is usually applied on each batch of batched data.

注解

Note that the formula for updating the running_mean and running_var is \(\hat{x}_\text{new} = (1 - \text{momentum}) \times x_t + \text{momentum} \times \hat{x}\), where \(\hat{x}\) is the estimated statistic and \(x_t\) is the new observed value.

参数

  • num_features (int) – C from an expected input of size (N, C, H, W).

  • eps (float) – A value added to the denominator for numerical stability. Default: 1e-5.

  • momentum (float) – A floating hyperparameter of the momentum for the running_mean and running_var computation. Default: 0.1.

  • affine (bool) – A bool value. When set to True, gamma and beta can be learned. Default: True.

  • gamma_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the gamma weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘ones’.

  • beta_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the beta weight. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘zeros’.

  • moving_mean_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the moving mean. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘zeros’.

  • moving_var_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the moving variance. The values of str refer to the function initializer including ‘zeros’, ‘ones’, etc. Default: ‘ones’.

  • use_batch_statistics (bool) – If true, use the mean value and variance value of current batch data. If false, use the mean value and variance value of specified value. Default: True.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C, H, W)\). Data type: float16 or float32.

Outputs:

Tensor, the normalized, scaled, offset tensor, of shape \((N, C, H, W)\). Same type and shape as the input_x.

Supported Platforms:

GPU

引发

  • TypeError – If num_features is not an int.

  • TypeError – If eps is not a float.

  • TypeError – If momentum is not a float.

  • TypeError – If affine is not a bool.

  • TypeError – If the type of gamma_init/beta_init/moving_mean_init/moving_var_init is not same, or if the initialized element type is not float32.

  • ValueError – If num_features is less than 1.

  • ValueError – If momentum is not in range [0, 1].

  • KeyError – If any of gamma_init/beta_init/moving_mean_init/moving_var_init is str and the homonymous class inheriting from Initializer not exists.

实际案例

>>> net = nn.InstanceNorm2d(3)
>>> np.random.seed(0)
>>> input = Tensor(np.random.randint(0, 255, [2, 3, 2, 2]), mindspore.float32)
>>> output = net(input)
>>> print(output.shape)
(2, 3, 2, 2)
class tinyms.layers.Conv2d(in_channels, out_channels, kernel_size, stride=1, pad_mode='same', padding=0, dilation=1, group=1, has_bias=False, weight_init='normal', bias_init='zeros', data_format='NCHW')[源代码]

2D convolution layer.

Applies a 2D convolution over an input tensor which is typically of shape \((N, C_{in}, H_{in}, W_{in})\), where \(N\) is batch size, \(C_{in}\) is channel number, and \(H_{in}, W_{in})\) are height and width. For each batch of shape \((C_{in}, H_{in}, W_{in})\), the formula is defined as:

\[out_j = \sum_{i=0}^{C_{in} - 1} ccor(W_{ij}, X_i) + b_j,\]

where \(ccor\) is the cross-correlation operator, \(C_{in}\) is the input channel number, \(j\) ranges from \(0\) to \(C_{out} - 1\), \(W_{ij}\) corresponds to the \(i\)-th channel of the \(j\)-th filter and \(out_{j}\) corresponds to the \(j\)-th channel of the output. \(W_{ij}\) is a slice of kernel and it has shape \((\text{ks_h}, \text{ks_w})\), where \(\text{ks_h}\) and \(\text{ks_w}\) are the height and width of the convolution kernel. The full kernel has shape \((C_{out}, C_{in} // \text{group}, \text{ks_h}, \text{ks_w})\), where group is the group number to split the input in the channel dimension.

If the ‘pad_mode’ is set to be “valid”, the output height and width will be \(\left \lfloor{1 + \frac{H_{in} + 2 \times \text{padding} - \text{ks_h} - (\text{ks_h} - 1) \times (\text{dilation} - 1) }{\text{stride}}} \right \rfloor\) and \(\left \lfloor{1 + \frac{W_{in} + 2 \times \text{padding} - \text{ks_w} - (\text{ks_w} - 1) \times (\text{dilation} - 1) }{\text{stride}}} \right \rfloor\) respectively.

The first introduction can be found in paper Gradient Based Learning Applied to Document Recognition.

参数

  • in_channels (int) – The number of input channel \(C_{in}\).

  • out_channels (int) – The number of output channel \(C_{out}\).

  • kernel_size (Union[int, tuple[int]]) – The data type is int or a tuple of 2 integers. Specifies the height and width of the 2D convolution window. Single int means the value is for both the height and the width of the kernel. A tuple of 2 ints means the first value is for the height and the other is for the width of the kernel.

  • stride (Union[int, tuple[int]]) – The distance of kernel moving, an int number that represents the height and width of movement are both strides, or a tuple of two int numbers that represent height and width of movement respectively. Default: 1.

  • pad_mode (str) –

    Specifies padding mode. The optional values are “same”, “valid”, “pad”. Default: “same”.

    • same: Adopts the way of completion. The height and width of the output will be the same as the input. The total number of padding will be calculated in horizontal and vertical directions and evenly distributed to top and bottom, left and right if possible. Otherwise, the last extra padding will be done from the bottom and the right side. If this mode is set, padding must be 0.

    • valid: Adopts the way of discarding. The possible largest height and width of output will be returned without padding. Extra pixels will be discarded. If this mode is set, padding must be 0.

    • pad: Implicit paddings on both sides of the input. The number of padding will be padded to the input Tensor borders. padding must be greater than or equal to 0.

  • padding (Union[int, tuple[int]]) – Implicit paddings on both sides of the input. If padding is one integer, the paddings of top, bottom, left and right are the same, equal to padding. If padding is a tuple with four integers, the paddings of top, bottom, left and right will be equal to padding[0], padding[1], padding[2], and padding[3] accordingly. Default: 0.

  • dilation (Union[int, tuple[int]]) – The data type is int or a tuple of 2 integers. Specifies the dilation rate to use for dilated convolution. If set to be \(k > 1\), there will be \(k - 1\) pixels skipped for each sampling location. Its value must be greater or equal to 1 and bounded by the height and width of the input. Default: 1.

  • group (int) – Splits filter into groups, in_ channels and out_channels must be divisible by the number of groups. If the group is equal to in_channels and out_channels, this 2D convolution layer also can be called 2D depthwise convolution layer. Default: 1.

  • has_bias (bool) – Specifies whether the layer uses a bias vector. Default: False.

  • weight_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the convolution kernel. It can be a Tensor, a string, an Initializer or a number. When a string is specified, values from ‘TruncatedNormal’, ‘Normal’, ‘Uniform’, ‘HeUniform’ and ‘XavierUniform’ distributions as well as constant ‘One’ and ‘Zero’ distributions are possible. Alias ‘xavier_uniform’, ‘he_uniform’, ‘ones’ and ‘zeros’ are acceptable. Uppercase and lowercase are both acceptable. Refer to the values of Initializer for more details. Default: ‘normal’.

  • bias_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the bias vector. Possible Initializer and string are the same as ‘weight_init’. Refer to the values of Initializer for more details. Default: ‘zeros’.

  • data_format (str) – The optional value for data format, is ‘NHWC’ or ‘NCHW’. Default: ‘NCHW’.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in}, H_{in}, W_{in})\) or \((N, H_{in}, W_{in}, C_{in})\).

Outputs:

Tensor of shape \((N, C_{out}, H_{out}, W_{out})\) or \((N, H_{out}, W_{out}, C_{out})\).

引发

  • TypeError – If in_channels, out_channels or group is not an int.

  • TypeError – If kernel_size, stride, padding or dilation is neither an int not a tuple.

  • ValueError – If in_channels, out_channels, kernel_size, stride or dilation is less than 1.

  • ValueError – If padding is less than 0.

  • ValueError – If pad_mode is not one of ‘same’, ‘valid’, ‘pad’.

  • ValueError – If padding is a tuple whose length is not equal to 4.

  • ValueError – If pad_mode is not equal to ‘pad’ and padding is not equal to (0, 0, 0, 0).

  • ValueError – If data_format is neither ‘NCHW’ not ‘NHWC’.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> net = nn.Conv2d(120, 240, 4, has_bias=False, weight_init='normal')
>>> input = Tensor(np.ones([1, 120, 1024, 640]), mindspore.float32)
>>> output = net(input).shape
>>> print(output)
(1, 240, 1024, 640)
class tinyms.layers.Conv2dTranspose(in_channels, out_channels, kernel_size, stride=1, pad_mode='same', padding=0, dilation=1, group=1, has_bias=False, weight_init='normal', bias_init='zeros')[源代码]

2D transposed convolution layer.

Compute a 2D transposed convolution, which is also known as a deconvolution (although it is not an actual deconvolution).

Input is typically of shape \((N, C, H, W)\), where \(N\) is batch size and \(C\) is channel number.

If the ‘pad_mode’ is set to be “pad”, the height and width of output are defined as:

\[ \begin{align}\begin{aligned}H_{out} = (H_{in} - 1) \times \text{stride} - 2 \times \text{padding} + \text{dilation} \times (\text{ks_h} - 1) + 1\\W_{out} = (W_{in} - 1) \times \text{stride} - 2 \times \text{padding} + \text{dilation} \times (\text{ks_w} - 1) + 1\end{aligned}\end{align} \]

where \(\text{ks_h}\) is the height of the convolution kernel and \(\text{ks_w}\) is the width of the convolution kernel.

参数

  • in_channels (int) – The number of channels in the input space.

  • out_channels (int) – The number of channels in the output space.

  • kernel_size (Union[int, tuple]) – int or a tuple of 2 integers, which specifies the height and width of the 2D convolution window. Single int means the value is for both the height and the width of the kernel. A tuple of 2 ints means the first value is for the height and the other is for the width of the kernel.

  • stride (Union[int, tuple[int]]) – The distance of kernel moving, an int number that represents the height and width of movement are both strides, or a tuple of two int numbers that represent height and width of movement respectively. Its value must be equal to or greater than 1. Default: 1.

  • pad_mode (str) –

    Select the mode of the pad. The optional values are “pad”, “same”, “valid”. Default: “same”.

    • pad: Implicit paddings on both sides of the input.

    • same: Adopted the way of completion.

    • valid: Adopted the way of discarding.

  • padding (Union[int, tuple[int]]) – Implicit paddings on both sides of the input. If padding is one integer, the paddings of top, bottom, left and right are the same, equal to padding. If padding is a tuple with four integers, the paddings of top, bottom, left and right will be equal to padding[0], padding[1], padding[2], and padding[3] accordingly. Default: 0.

  • dilation (Union[int, tuple[int]]) – The data type is int or a tuple of 2 integers. Specifies the dilation rate to use for dilated convolution. If set to be \(k > 1\), there will be \(k - 1\) pixels skipped for each sampling location. Its value must be greater than or equal to 1 and bounded by the height and width of the input. Default: 1.

  • group (int) – Splits filter into groups, in_channels and out_channels must be divisible by the number of groups. This does not support for Davinci devices when group > 1. Default: 1.

  • has_bias (bool) – Specifies whether the layer uses a bias vector. Default: False.

  • weight_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the convolution kernel. It can be a Tensor, a string, an Initializer or a number. When a string is specified, values from ‘TruncatedNormal’, ‘Normal’, ‘Uniform’, ‘HeUniform’ and ‘XavierUniform’ distributions as well as constant ‘One’ and ‘Zero’ distributions are possible. Alias ‘xavier_uniform’, ‘he_uniform’, ‘ones’ and ‘zeros’ are acceptable. Uppercase and lowercase are both acceptable. Refer to the values of Initializer for more details. Default: ‘normal’.

  • bias_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the bias vector. Possible Initializer and string are the same as ‘weight_init’. Refer to the values of Initializer for more details. Default: ‘zeros’.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in}, H_{in}, W_{in})\).

Outputs:

Tensor of shape \((N, C_{out}, H_{out}, W_{out})\).

引发

  • TypeError – If in_channels, out_channels or group is not an int.

  • TypeError – If kernel_size, stride, padding or dilation is neither an int not a tuple.

  • ValueError – If in_channels, out_channels, kernel_size, stride or dilation is less than 1.

  • ValueError – If padding is less than 0.

  • ValueError – If pad_mode is not one of ‘same’, ‘valid’, ‘pad’.

  • ValueError – If padding is a tuple whose length is not equal to 4.

  • ValueError – If pad_mode is not equal to ‘pad’ and padding is not equal to (0, 0, 0, 0).

Supported Platforms:

Ascend GPU

实际案例

>>> net = nn.Conv2dTranspose(3, 64, 4, has_bias=False, weight_init='normal', pad_mode='pad')
>>> input = Tensor(np.ones([1, 3, 16, 50]), mindspore.float32)
>>> output = net(input).shape
>>> print(output)
(1, 64, 19, 53)
class tinyms.layers.Conv1d(in_channels, out_channels, kernel_size, stride=1, pad_mode='same', padding=0, dilation=1, group=1, has_bias=False, weight_init='normal', bias_init='zeros')[源代码]

1D convolution layer.

Applies a 1D convolution over an input tensor which is typically of shape \((N, C_{in}, W_{in})\), where \(N\) is batch size and \(C_{in}\) is channel number. For each batch of shape \((C_{in}, W_{in})\), the formula is defined as:

\[out_j = \sum_{i=0}^{C_{in} - 1} ccor(W_{ij}, X_i) + b_j,\]

where \(ccor\) is the cross correlation operator, \(C_{in}\) is the input channel number, \(j\) ranges from \(0\) to \(C_{out} - 1\), \(W_{ij}\) corresponds to the \(i\)-th channel of the \(j\)-th filter and \(out_{j}\) corresponds to the \(j\)-th channel of the output. \(W_{ij}\) is a slice of kernel and it has shape \((\text{ks_w})\), where \(\text{ks_w}\) is the width of the convolution kernel. The full kernel has shape \((C_{out}, C_{in} // \text{group}, \text{ks_w})\), where group is the group number to split the input in the channel dimension.

If the ‘pad_mode’ is set to be “valid”, the output width will be \(\left \lfloor{1 + \frac{W_{in} + 2 \times \text{padding} - \text{ks_w} - (\text{ks_w} - 1) \times (\text{dilation} - 1) }{\text{stride}}} \right \rfloor\) respectively.

The first introduction of convolution layer can be found in paper Gradient Based Learning Applied to Document Recognition.

参数

  • in_channels (int) – The number of input channel \(C_{in}\).

  • out_channels (int) – The number of output channel \(C_{out}\).

  • kernel_size (int) – The data type is int. Specifies the width of the 1D convolution window.

  • stride (int) – The distance of kernel moving, an int number that represents the width of movement. Default: 1.

  • pad_mode (str) –

    Specifies padding mode. The optional values are “same”, “valid”, “pad”. Default: “same”.

    • same: Adopts the way of completion. The output width will be the same as the input. The total number of padding will be calculated in the horizontal direction and evenly distributed to left and right if possible. Otherwise, the last extra padding will be done from the bottom and the right side. If this mode is set, padding must be 0.

    • valid: Adopts the way of discarding. The possible largest width of the output will be returned without padding. Extra pixels will be discarded. If this mode is set, padding must be 0.

    • pad: Implicit paddings on both sides of the input. The number of padding will be padded to the input Tensor borders. padding must be greater than or equal to 0.

  • padding (int) – Implicit paddings on both sides of the input. Default: 0.

  • dilation (int) – The data type is int. Specifies the dilation rate to use for dilated convolution. If set to be \(k > 1\), there will be \(k - 1\) pixels skipped for each sampling location. Its value must be greater or equal to 1 and bounded by the height and width of the input. Default: 1.

  • group (int) – Splits filter into groups, in_ channels and out_channels must be divisible by the number of groups. Default: 1.

  • has_bias (bool) – Specifies whether the layer uses a bias vector. Default: False.

  • weight_init (Union[Tensor, str, Initializer, numbers.Number]) – An initializer for the convolution kernel. It can be a Tensor, a string, an Initializer or a number. When a string is specified, values from ‘TruncatedNormal’, ‘Normal’, ‘Uniform’, ‘HeUniform’ and ‘XavierUniform’ distributions as well as constant ‘One’ and ‘Zero’ distributions are possible. Alias ‘xavier_uniform’, ‘he_uniform’, ‘ones’ and ‘zeros’ are acceptable. Uppercase and lowercase are both acceptable. Refer to the values of Initializer for more details. Default: ‘normal’.

  • bias_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the bias vector. Possible Initializer and string are the same as ‘weight_init’. Refer to the values of Initializer for more details. Default: ‘zeros’.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in}, W_{in})\).

Outputs:

Tensor of shape \((N, C_{out}, W_{out})\).

引发

  • TypeError – If in_channels, out_channels, kernel_size, stride, padding or dilation is not an int.

  • ValueError – If in_channels, out_channels, kernel_size, stride or dilation is less than 1.

  • ValueError – If padding is less than 0.

  • ValueError – If pad_mode is not one of ‘same’, ‘valid’, ‘pad’.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> net = nn.Conv1d(120, 240, 4, has_bias=False, weight_init='normal')
>>> input = Tensor(np.ones([1, 120, 640]), mindspore.float32)
>>> output = net(input).shape
>>> print(output)
(1, 240, 640)
class tinyms.layers.Conv1dTranspose(in_channels, out_channels, kernel_size, stride=1, pad_mode='same', padding=0, dilation=1, group=1, has_bias=False, weight_init='normal', bias_init='zeros')[源代码]

1D transposed convolution layer.

Compute a 1D transposed convolution, which is also known as a deconvolution (although it is not an actual deconvolution).

Input is typically of shape \((N, C, W)\), where \(N\) is batch size and \(C\) is channel number.

If the ‘pad_mode’ is set to be “pad”, the width of output is defined as:

\[W_{out} = (W_{in} - 1) \times \text{stride} - 2 \times \text{padding} + \text{dilation} \times (\text{ks_w} - 1) + 1\]

where \(\text{ks_w}\) is the width of the convolution kernel.

参数

  • in_channels (int) – The number of channels in the input space.

  • out_channels (int) – The number of channels in the output space.

  • kernel_size (int) – int, which specifies the width of the 1D convolution window.

  • stride (int) – The distance of kernel moving, an int number that represents the width of movement. Default: 1.

  • pad_mode (str) –

    Select the mode of the pad. The optional values are “pad”, “same”, “valid”. Default: “same”.

    • pad: Implicit paddings on both sides of the input.

    • same: Adopted the way of completion.

    • valid: Adopted the way of discarding.

  • padding (int) – Implicit paddings on both sides of the input. Default: 0.

  • dilation (int) – The data type is int. Specifies the dilation rate to use for dilated convolution. If set to be \(k > 1\), there will be \(k - 1\) pixels skipped for each sampling location. Its value must be greater or equal to 1 and bounded by the width of the input. Default: 1.

  • group (int) – Splits filter into groups, in_channels and out_channels must be divisible by the number of groups. This is not support for Davinci devices when group > 1. Default: 1.

  • has_bias (bool) – Specifies whether the layer uses a bias vector. Default: False.

  • weight_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the convolution kernel. It can be a Tensor, a string, an Initializer or a numbers.Number. When a string is specified, values from ‘TruncatedNormal’, ‘Normal’, ‘Uniform’, ‘HeUniform’ and ‘XavierUniform’ distributions as well as constant ‘One’ and ‘Zero’ distributions are possible. Alias ‘xavier_uniform’, ‘he_uniform’, ‘ones’ and ‘zeros’ are acceptable. Uppercase and lowercase are both acceptable. Refer to the values of Initializer for more details. Default: ‘normal’.

  • bias_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the bias vector. Possible Initializer and string are the same as ‘weight_init’. Refer to the values of Initializer for more details. Default: ‘zeros’.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in}, W_{in})\).

Outputs:

Tensor of shape \((N, C_{out}, W_{out})\).

引发

  • TypeError – If in_channels, out_channels, kernel_size, stride, padding or dilation is not an int.

  • ValueError – If in_channels, out_channels, kernel_size, stride or dilation is less than 1.

  • ValueError – If padding is less than 0.

  • ValueError – If pad_mode is not one of ‘same’, ‘valid’, ‘pad’.

Supported Platforms:

Ascend GPU

实际案例

>>> net = nn.Conv1dTranspose(3, 64, 4, has_bias=False, weight_init='normal', pad_mode='pad')
>>> input = Tensor(np.ones([1, 3, 50]), mindspore.float32)
>>> output = net(input).shape
>>> print(output)
(1, 64, 53)
class tinyms.layers.Conv3d(in_channels, out_channels, kernel_size, stride=1, pad_mode='same', padding=0, dilation=1, group=1, has_bias=False, weight_init='normal', bias_init='zeros', data_format='NCDHW')[源代码]

3D convolution layer.

Applies a 3D convolution over an input tensor which is typically of shape \((N, C_{in}, D_{in}, H_{in}, W_{in})\) and output shape \((N, C_{out}, D_{out}, H_{out}, W_{out})\). where \(N\) is batch size. \(C\) is channel number. the formula is defined as:

\[\operatorname{out}\left(N_{i}, C_{\text {out}_j}\right)=\operatorname{bias}\left(C_{\text {out}_j}\right)+ \sum_{k=0}^{C_{in}-1} ccor(\text {weight}\left(C_{\text {out}_j}, k\right), \operatorname{input}\left(N_{i}, k\right))\]

where \(ccor\) is the cross-correlation operator.

If the ‘pad_mode’ is set to be “valid”, the output height and width will be \(\left \lfloor{1 + \frac{D_{in} + 2 \times \text{padding} - \text{ks_d} - (\text{ks_d} - 1) \times (\text{dilation} - 1) }{\text{stride}}} \right \rfloor\) and \(\left \lfloor{1 + \frac{H_{in} + 2 \times \text{padding} - \text{ks_h} - (\text{ks_h} - 1) \times (\text{dilation} - 1) }{\text{stride}}} \right \rfloor\) and \(\left \lfloor{1 + \frac{W_{in} + 2 \times \text{padding} - \text{ks_w} - (\text{ks_w} - 1) \times (\text{dilation} - 1) }{\text{stride}}} \right \rfloor\) respectively.

参数

  • in_channels (int) – The number of input channel \(C_{in}\).

  • out_channels (int) – The number of output channel \(C_{out}\).

  • kernel_size (Union[int, tuple[int]]) – The data type is int or a tuple of 3 integers. Specifies the depth, height and width of the 3D convolution window. Single int means the value is for the depth, height and the width of the kernel. A tuple of 3 ints means the first value is for the depth, second value is for height and the other is for the width of the kernel.

  • stride (Union[int, tuple[int]]) – The distance of kernel moving, an int number that represents the depth, height and width of movement are both strides, or a tuple of three int numbers that represent depth, height and width of movement respectively. Default: 1.

  • pad_mode (str) –

    Specifies padding mode. The optional values are “same”, “valid”, “pad”. Default: “same”.

    • same: Adopts the way of completion. The depth, height and width of the output will be the same as the input. The total number of padding will be calculated in depth, horizontal and vertical directions and evenly distributed to head and tail, top and bottom, left and right if possible. Otherwise, the last extra padding will be done from the tail, bottom and the right side. If this mode is set, padding must be 0.

    • valid: Adopts the way of discarding. The possible largest depth, height and width of output will be returned without padding. Extra pixels will be discarded. If this mode is set, padding must be 0.

    • pad: Implicit paddings on both sides of the input in depth, height, width. The number of padding will be padded to the input Tensor borders. padding must be greater than or equal to 0.

  • padding (Union(int, tuple[int])) – Implicit paddings on both sides of the input. The data type is int or a tuple of 6 integers. Default: 0. If padding is an integer, the paddings of head, tail, top, bottom, left and right are the same, equal to padding. If paddings is a tuple of six integers, the padding of head, tail, top, bottom, left and right equal to padding[0], padding[1], padding[2], padding[3], padding[4] and padding[5] correspondingly.

  • dilation (Union[int, tuple[int]]) – The data type is int or a tuple of 3 integers : math:(dilation_d, dilation_h, dilation_w). Currently, dilation on depth only supports the case of 1. Specifies the dilation rate to use for dilated convolution. If set to be \(k > 1\), there will be \(k - 1\) pixels skipped for each sampling location. Its value must be greater or equal to 1 and bounded by the height and width of the input. Default: 1.

  • group (int) – Splits filter into groups, in_ channels and out_channels must be divisible by the number of groups. Default: 1. Only 1 is currently supported.

  • has_bias (bool) – Specifies whether the layer uses a bias vector. Default: False.

  • weight_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the convolution kernel. It can be a Tensor, a string, an Initializer or a number. When a string is specified, values from ‘TruncatedNormal’, ‘Normal’, ‘Uniform’, ‘HeUniform’ and ‘XavierUniform’ distributions as well as constant ‘One’ and ‘Zero’ distributions are possible. Alias ‘xavier_uniform’, ‘he_uniform’, ‘ones’ and ‘zeros’ are acceptable. Uppercase and lowercase are both acceptable. Refer to the values of Initializer for more details. Default: ‘normal’.

  • bias_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the bias vector. Possible Initializer and string are the same as ‘weight_init’. Refer to the values of Initializer for more details. Default: ‘zeros’.

  • data_format (str) – The optional value for data format. Currently only support “NCDHW”.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in}, D_{in}, H_{in}, W_{in})\). Currently input data type only support float16 and float32.

Outputs:

Tensor, the value that applied 3D convolution. The shape is \((N, C_{out}, D_{out}, H_{out}, W_{out})\).

引发

  • TypeError – If in_channels, out_channels or group is not an int.

  • TypeError – If kernel_size, stride, padding or dilation is neither an int nor a tuple of six.

  • ValueError – If out_channels, kernel_size, stride or dilation is less than 1.

  • ValueError – If padding is less than 0.

  • ValueError – If pad_mode is not one of ‘same’, ‘valid’, ‘pad’.

  • ValueError – If padding is a tuple whose length is not equal to 6.

  • ValueError – If pad_mode is not equal to ‘pad’ and padding is not equal to (0, 0, 0, 0, 0, 0).

  • ValueError – If data_format is not ‘NCDHW’.

Supported Platforms:

Ascend

实际案例

>>> input = Tensor(np.ones([16, 3, 10, 32, 32]), mindspore.float32)
>>> conv3d = nn.Conv3d(in_channels=3, out_channels=32, kernel_size=(4, 3, 3))
>>> output = conv3d(input)
>>> print(output.shape)
(16, 32, 10, 32, 32)
class tinyms.layers.Conv3dTranspose(in_channels, out_channels, kernel_size, stride=1, pad_mode='same', padding=0, dilation=1, group=1, output_padding=0, has_bias=False, weight_init='normal', bias_init='zeros', data_format='NCDHW')[源代码]

Compute a 3D transposed convolution, which is also known as a deconvolution (although it is not an actual deconvolution).

Input is typically of shape \((N, C, D, H, W)\), where \(N\) is batch size and \(C\) is channel number.

If the ‘pad_mode’ is set to be “pad”, the height and width of output are defined as:

\[ \begin{align}\begin{aligned}D_{out} = (D_{in} - 1) \times \text{stride_d} - 2 \times \text{padding_d} + \text{dilation_d} \times (\text{kernel_size_d} - 1) + \text{output_padding_d} + 1\\H_{out} = (H_{in} - 1) \times \text{stride_h} - 2 \times \text{padding_h} + \text{dilation_h} \times (\text{kernel_size_h} - 1) + \text{output_padding_h} + 1\\W_{out} = (W_{in} - 1) \times \text{stride_w} - 2 \times \text{padding_w} + \text{dilation_w} \times (\text{kernel_size_w} - 1) + \text{output_padding_w} + 1\end{aligned}\end{align} \]
参数

  • in_channels (int) – The number of input channel \(C_{in}\).

  • out_channels (int) – The number of output channel \(C_{out}\).

  • kernel_size (Union[int, tuple[int]]) – The kernel size of the 3D convolution.

  • stride (Union[int, tuple[int]]) – The distance of kernel moving, an int number that represents the depth, height and width of movement are both strides, or a tuple of three int numbers that represent depth, height and width of movement respectively. Its value must be equal to or greater than 1. Default: 1.

  • pad_mode (str) –

    Select the mode of the pad. The optional values are “pad”, “same”, “valid”. Default: “same”.

    • same: Adopts the way of completion. The depth, height and width of the output will be the same as the input. The total number of padding will be calculated in depth, horizontal and vertical directions and evenly distributed to head and tail, top and bottom, left and right if possible. Otherwise, the last extra padding will be done from the tail, bottom and the right side. If this mode is set, padding and output_padding must be 0.

    • valid: Adopts the way of discarding. The possible largest depth, height and width of output will be returned without padding. Extra pixels will be discarded. If this mode is set, padding and output_padding must be 0.

    • pad: Implicit paddings on both sides of the input in depth, height, width. The number of pad will be padded to the input Tensor borders. padding must be greater than or equal to 0.

  • padding (Union(int, tuple[int])) – The pad value to be filled. Default: 0. If padding is an integer, the paddings of head, tail, top, bottom, left and right are the same, equal to padding. If padding is a tuple of six integers, the padding of head, tail, top, bottom, left and right equal to padding[0], padding[1], padding[2], padding[3], padding[4] and padding[5] correspondingly.

  • dilation (Union(int, tuple[int])) – The data type is int or a tuple of 3 integers : math:(dilation_d, dilation_h, dilation_w). Currently, dilation on depth only supports the case of 1. Specifies the dilation rate to use for dilated convolution. If set to be \(k > 1\), there will be \(k - 1\) pixels skipped for each sampling location. Its value must be greater or equal to 1 and bounded by the height and width of the input. Default: 1.

  • group (int) – Splits filter into groups, in_ channels and out_channels must be divisible by the number of groups. Default: 1. Only 1 is currently supported.

  • output_padding (Union(int, tuple[int])) – Add extra size to each dimension of the output. Default: 0. Must be greater than or equal to 0.

  • has_bias (bool) – Specifies whether the layer uses a bias vector. Default: False.

  • weight_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the convolution kernel. It can be a Tensor, a string, an Initializer or a number. When a string is specified, values from ‘TruncatedNormal’, ‘Normal’, ‘Uniform’, ‘HeUniform’ and ‘XavierUniform’ distributions as well as constant ‘One’ and ‘Zero’ distributions are possible. Alias ‘xavier_uniform’, ‘he_uniform’, ‘ones’ and ‘zeros’ are acceptable. Uppercase and lowercase are both acceptable. Refer to the values of Initializer for more details. Default: ‘normal’.

  • bias_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the bias vector. Possible Initializer and string are the same as ‘weight_init’. Refer to the values of Initializer for more details. Default: ‘zeros’.

  • data_format (str) – The optional value for data format. Currently only support ‘NCDHW’.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in}, D_{in}, H_{in}, W_{in})\). Currently input data type only support float16 and float32.

Outputs:

Tensor, the shape is \((N, C_{out}, D_{out}, H_{out}, W_{out})\).

Supported Platforms:

Ascend

引发

  • TypeError – If in_channels, out_channels or group is not an int.

  • TypeError – If kernel_size, stride, padding , dilation or output_padding is neither an int not a tuple of three.

  • TypeError – If input data type is not float16 or float32.

  • ValueError – If in_channels, out_channels, kernel_size, stride or dilation is less than 1.

  • ValueError – If padding is less than 0.

  • ValueError – If pad_mode is not one of ‘same’, ‘valid’, ‘pad’.

  • ValueError – If padding is a tuple whose length is not equal to 6.

  • ValueError – If pad_mode is not equal to ‘pad’ and padding is not equal to (0, 0, 0, 0, 0, 0).

  • ValueError – If data_format is not ‘NCDHW’.

实际案例

>>> input = Tensor(np.ones([32, 16, 10, 32, 32]), mindspore.float32)
>>> conv3d_transpose = nn.Conv3dTranspose(in_channels=16, out_channels=3, kernel_size=(4, 6, 2), pad_mode='pad')
>>> output = conv3d_transpose(input)
>>> print(output.shape)
(32, 3, 13, 37, 33)
class tinyms.layers.LSTM(input_size, hidden_size, num_layers=1, has_bias=True, batch_first=False, dropout=0, bidirectional=False)[源代码]

Stacked LSTM (Long Short-Term Memory) layers.

Apply LSTM layer to the input.

There are two pipelines connecting two consecutive cells in a LSTM model; one is cell state pipeline and the other is hidden state pipeline. Denote two consecutive time nodes as \(t-1\) and \(t\). Given an input \(x_t\) at time \(t\), an hidden state \(h_{t-1}\) and an cell state \(c_{t-1}\) of the layer at time \({t-1}\), the cell state and hidden state at time \(t\) is computed using an gating mechanism. Input gate \(i_t\) is designed to protect the cell from perturbation by irrelevant inputs. Forget gate \(f_t\) affords protection of the cell by forgetting some information in the past, which is stored in \(h_{t-1}\). Output gate \(o_t\) protects other units from perturbation by currently irrelevant memory contents. Candidate cell state \(\tilde{c}_t\) is calculated with the current input, on which the input gate will be applied. Finally, current cell state \(c_{t}\) and hidden state \(h_{t}\) are computed with the calculated gates and cell states. The complete formulation is as follows.

\[\begin{split}\begin{array}{ll} \\ i_t = \sigma(W_{ix} x_t + b_{ix} + W_{ih} h_{(t-1)} + b_{ih}) \\ f_t = \sigma(W_{fx} x_t + b_{fx} + W_{fh} h_{(t-1)} + b_{fh}) \\ \tilde{c}_t = \tanh(W_{cx} x_t + b_{cx} + W_{ch} h_{(t-1)} + b_{ch}) \\ o_t = \sigma(W_{ox} x_t + b_{ox} + W_{oh} h_{(t-1)} + b_{oh}) \\ c_t = f_t * c_{(t-1)} + i_t * \tilde{c}_t \\ h_t = o_t * \tanh(c_t) \\ \end{array}\end{split}\]

Here \(\sigma\) is the sigmoid function, and \(*\) is the Hadamard product. \(W, b\) are learnable weights between the output and the input in the formula. For instance, \(W_{ix}, b_{ix}\) are the weight and bias used to transform from input \(x\) to \(i\). Details can be found in paper LONG SHORT-TERM MEMORY and Long Short-Term Memory Recurrent Neural Network Architectures for Large Scale Acoustic Modeling.

参数

  • input_size (int) – Number of features of input.

  • hidden_size (int) – Number of features of hidden layer.

  • num_layers (int) – Number of layers of stacked LSTM . Default: 1.

  • has_bias (bool) – Whether the cell has bias b_ih and b_hh. Default: True.

  • batch_first (bool) – Specifies whether the first dimension of input is batch_size. Default: False.

  • dropout (float, int) – If not 0, append Dropout layer on the outputs of each LSTM layer except the last layer. Default 0. The range of dropout is [0.0, 1.0].

  • bidirectional (bool) – Specifies whether it is a bidirectional LSTM. Default: False.

Inputs:

  • input (Tensor) - Tensor of shape (seq_len, batch_size, input_size) or (batch_size, seq_len, input_size).

  • hx (tuple) - A tuple of two Tensors (h_0, c_0) both of data type mindspore.float32 or mindspore.float16 and shape (num_directions * num_layers, batch_size, hidden_size). Data type of hx must be the same as input.

Outputs:

Tuple, a tuple contains (output, (h_n, c_n)).

  • output (Tensor) - Tensor of shape (seq_len, batch_size, num_directions * hidden_size).

  • hx_n (tuple) - A tuple of two Tensor (h_n, c_n) both of shape (num_directions * num_layers, batch_size, hidden_size).

引发

  • TypeError – If input_size, hidden_size or num_layers is not an int.

  • TypeError – If has_bias, batch_first or bidirectional is not a bool.

  • TypeError – If dropout is neither a float nor an int.

  • ValueError – If dropout is not in range [0.0, 1.0].

Supported Platforms:

Ascend GPU

实际案例

>>> net = nn.LSTM(10, 16, 2, has_bias=True, batch_first=True, bidirectional=False)
>>> input = Tensor(np.ones([3, 5, 10]).astype(np.float32))
>>> h0 = Tensor(np.ones([1 * 2, 3, 16]).astype(np.float32))
>>> c0 = Tensor(np.ones([1 * 2, 3, 16]).astype(np.float32))
>>> output, (hn, cn) = net(input, (h0, c0))
>>> print(output.shape)
(3, 5, 16)
class tinyms.layers.LSTMCell(input_size, hidden_size, has_bias=True, batch_first=False, dropout=0, bidirectional=False)[源代码]

LSTM (Long Short-Term Memory) layer.

Apply LSTM layer to the input.

There are two pipelines connecting two consecutive cells in a LSTM model; one is cell state pipeline and the other is hidden state pipeline. Denote two consecutive time nodes as \(t-1\) and \(t\). Given an input \(x_t\) at time \(t\), an hidden state \(h_{t-1}\) and an cell state \(c_{t-1}\) of the layer at time \({t-1}\), the cell state and hidden state at time \(t\) is computed using an gating mechanism. Input gate \(i_t\) is designed to protect the cell from perturbation by irrelevant inputs. Forget gate \(f_t\) affords protection of the cell by forgetting some information in the past, which is stored in \(h_{t-1}\). Output gate \(o_t\) protects other units from perturbation by currently irrelevant memory contents. Candidate cell state \(\tilde{c}_t\) is calculated with the current input, on which the input gate will be applied. Finally, current cell state \(c_{t}\) and hidden state \(h_{t}\) are computed with the calculated gates and cell states. The complete formulation is as follows.

\[\begin{split}\begin{array}{ll} \\ i_t = \sigma(W_{ix} x_t + b_{ix} + W_{ih} h_{(t-1)} + b_{ih}) \\ f_t = \sigma(W_{fx} x_t + b_{fx} + W_{fh} h_{(t-1)} + b_{fh}) \\ \tilde{c}_t = \tanh(W_{cx} x_t + b_{cx} + W_{ch} h_{(t-1)} + b_{ch}) \\ o_t = \sigma(W_{ox} x_t + b_{ox} + W_{oh} h_{(t-1)} + b_{oh}) \\ c_t = f_t * c_{(t-1)} + i_t * \tilde{c}_t \\ h_t = o_t * \tanh(c_t) \\ \end{array}\end{split}\]

Here \(\sigma\) is the sigmoid function, and \(*\) is the Hadamard product. \(W, b\) are learnable weights between the output and the input in the formula. For instance, \(W_{ix}, b_{ix}\) are the weight and bias used to transform from input \(x\) to \(i\). Details can be found in paper LONG SHORT-TERM MEMORY and Long Short-Term Memory Recurrent Neural Network Architectures for Large Scale Acoustic Modeling.

LSTMCell is a single-layer RNN, you can achieve multi-layer RNN by stacking LSTMCell.

参数

  • input_size (int) – Number of features of input.

  • hidden_size (int) – Number of features of hidden layer.

  • has_bias (bool) – Whether the cell has bias b_ih and b_hh. Default: True.

  • batch_first (bool) – Specifies whether the first dimension of input is batch_size. Default: False.

  • dropout (float, int) – If not 0, append Dropout layer on the outputs of each LSTM layer except the last layer. Default 0. The range of dropout is [0.0, 1.0].

  • bidirectional (bool) – Specifies whether this is a bidirectional LSTM. If set True, number of directions will be 2 otherwise number of directions is 1. Default: False.

Inputs:

  • input (Tensor) - Tensor of shape (seq_len, batch_size, input_size).

  • h - data type mindspore.float32 or mindspore.float16 and shape (num_directions, batch_size, hidden_size).

  • c - data type mindspore.float32 or mindspore.float16 and shape (num_directions, batch_size, hidden_size). Data type of h’ and ‘c’ must be the same of `input.

  • w - data type mindspore.float32 or mindspore.float16 and shape (weight_size, 1, 1). The value of weight_size depends on input_size, hidden_size and bidirectional

Outputs:

output, h_n, c_n, ‘reserve’, ‘state’.

  • output (Tensor) - Tensor of shape (seq_len, batch_size, num_directions * hidden_size).

  • h - A Tensor with shape (num_directions, batch_size, hidden_size).

  • c - A Tensor with shape (num_directions, batch_size, hidden_size).

  • reserve - reserved

  • state - reserved

引发

  • TypeError – If input_size or hidden_size or num_layers is not an int.

  • TypeError – If has_bias or batch_first or bidirectional is not a bool.

  • TypeError – If dropout is neither a float nor an int.

  • ValueError – If dropout is not in range [0.0, 1.0].

Supported Platforms:

GPU CPU

实际案例

>>> net = nn.LSTMCell(10, 12, has_bias=True, batch_first=True, bidirectional=False)
>>> input = Tensor(np.ones([3, 5, 10]).astype(np.float32))
>>> h = Tensor(np.ones([1, 3, 12]).astype(np.float32))
>>> c = Tensor(np.ones([1, 3, 12]).astype(np.float32))
>>> w = Tensor(np.ones([1152, 1, 1]).astype(np.float32))
>>> output, h, c, _, _ = net(input, h, c, w)
>>> print(output.shape)
(3, 5, 12)
class tinyms.layers.Dropout(keep_prob=0.5, dtype=mindspore.float32)[源代码]

Dropout layer for the input.

Randomly set some elements of the input tensor to zero with probability \(1 - keep\_prob\) during training using samples from a Bernoulli distribution.

The outputs are scaled by a factor of \(\frac{1}{keep\_prob}\) during training so that the output layer remains at a similar scale. During inference, this layer returns the same tensor as the input.

This technique is proposed in paper Dropout: A Simple Way to Prevent Neural Networks from Overfitting and proved to be effective to reduce over-fitting and prevents neurons from co-adaptation. See more details in Improving neural networks by preventing co-adaptation of feature detectors.

注解

Each channel will be zeroed out independently on every construct call.

参数

  • keep_prob (float) – The keep rate, greater than 0 and less equal than 1. E.g. rate=0.9, dropping out 10% of input units. Default: 0.5.

  • dtype (mindspore.dtype) – Data type of input. Default: mindspore.float32.

Inputs:

  • input (Tensor) - The input of Dropout with data type of float16 or float32.

Outputs:

Tensor, output tensor with the same shape as the input.

引发

  • TypeError – If keep_prob is not a float.

  • TypeError – If dtype of input is not neither float16 nor float32.

  • ValueError – If keep_prob is not in range (0, 1].

  • ValueError – If length of shape of input is less than 1.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> x = Tensor(np.ones([2, 2, 3]), mindspore.float32)
>>> net = nn.Dropout(keep_prob=0.8)
>>> net.set_train()
Dropout<keep_prob=0.8>
>>> output = net(x)
>>> print(output.shape)
(2, 2, 3)
class tinyms.layers.Flatten[源代码]

Flatten layer for the input.

Flattens a tensor without changing dimension of batch size on the 0-th axis.

Inputs:

  • input (Tensor) - Tensor of shape \((N, \ldots)\) to be flattened.

Outputs:

Tensor, the shape of the output tensor is \((N, X)\), where \(X\) is the product of the remaining dimensions.

引发

TypeError – If input is not a subclass of Tensor.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> input = Tensor(np.array([[[1.2, 1.2], [2.1, 2.1]], [[2.2, 2.2], [3.2, 3.2]]]), mindspore.float32)
>>> net = nn.Flatten()
>>> output = net(input)
>>> print(output)
[[1.2 1.2 2.1 2.1]
 [2.2 2.2 3.2 3.2]]
class tinyms.layers.Dense(in_channels, out_channels, weight_init='normal', bias_init='zeros', has_bias=True, activation=None)[源代码]

The dense connected layer.

Applies dense connected layer for the input. This layer implements the operation as:

\[\text{outputs} = \text{activation}(\text{inputs} * \text{kernel} + \text{bias}),\]

where \(\text{activation}\) is the activation function passed as the activation argument (if passed in), \(\text{kernel}\) is a weight matrix with the same data type as the inputs created by the layer, and \(\text{bias}\) is a bias vector with the same data type as the inputs created by the layer (only if has_bias is True).

参数

  • in_channels (int) – The number of channels in the input space.

  • out_channels (int) – The number of channels in the output space.

  • weight_init (Union[Tensor, str, Initializer, numbers.Number]) – The trainable weight_init parameter. The dtype is same as input x. The values of str refer to the function initializer. Default: ‘normal’.

  • bias_init (Union[Tensor, str, Initializer, numbers.Number]) – The trainable bias_init parameter. The dtype is same as input x. The values of str refer to the function initializer. Default: ‘zeros’.

  • has_bias (bool) – Specifies whether the layer uses a bias vector. Default: True.

  • activation (Union[str, Cell, Primitive]) – activate function applied to the output of the fully connected layer, eg. ‘ReLU’.Default: None.

Inputs:

  • input (Tensor) - Tensor of shape \((*, in\_channels)\).

Outputs:

Tensor of shape \((*, out\_channels)\).

引发

  • TypeError – If in_channels or out_channels is not an int.

  • TypeError – If has_bias is not a bool.

  • TypeError – If activation is not one of str, Cell, Primitive, None.

  • ValueError – If length of shape of weight_init is not equal to 2 or shape[0] of weight_init is not equal to out_channels or shape[1] of weight_init is not equal to in_channels.

  • ValueError – If length of shape of bias_init is not equal to 1 or shape[0] of bias_init is not equal to out_channels.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> input = Tensor(np.array([[180, 234, 154], [244, 48, 247]]), mindspore.float32)
>>> net = nn.Dense(3, 4)
>>> output = net(input)
>>> print(output.shape)
(2, 4)
class tinyms.layers.ClipByNorm(axis=None)[源代码]

Clips tensor values to a maximum \(L_2\)-norm.

The output of this layer remains the same if the \(L_2\)-norm of the input tensor is not greater than the argument clip_norm. Otherwise the tensor will be normalized as:

\[\text{output}(X) = \frac{\text{clip_norm} * X}{L_2(X)},\]

where \(L_2(X)\) is the \(L_2\)-norm of \(X\).

参数

axis (Union[None, int, tuple(int)]) – Compute the L2-norm along the Specific dimension. Default: None, all dimensions to calculate.

Inputs:

  • input (Tensor) - Tensor of shape N-D. The type must be float32 or float16.

  • clip_norm (Tensor) - A scalar Tensor of shape \(()\) or \((1)\). Or a tensor shape can be broadcast to input shape.

Outputs:

Tensor, clipped tensor with the same shape as the input, whose type is float32.

引发

  • TypeError – If axis is not one of None, int, tuple.

  • TypeError – If dtype of input is neither float32 nor float16.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> net = nn.ClipByNorm()
>>> input = Tensor(np.random.randint(0, 10, [4, 16]), mindspore.float32)
>>> clip_norm = Tensor(np.array([100]).astype(np.float32))
>>> output = net(input, clip_norm)
>>> print(output.shape)
(4, 16)
class tinyms.layers.Norm(axis=(), keep_dims=False)[源代码]

Computes the norm of vectors, currently including Euclidean norm, i.e., \(L_2\)-norm.

\[norm(x) = \sqrt{\sum_{i=1}^{n} (x_i^2)}\]
参数

  • axis (Union[tuple, int]) – The axis over which to compute vector norms. Default: ().

  • keep_dims (bool) – If true, the axis indicated in axis are kept with size 1. Otherwise, the dimensions in axis are removed from the output shape. Default: False.

Inputs:

  • input (Tensor) - Tensor which is not empty.

Outputs:

Tensor, output tensor with dimensions in ‘axis’ reduced to 1 will be returned if ‘keep_dims’ is True; otherwise a Tensor with dimensions in ‘axis’ removed is returned.

引发

  • TypeError – If axis is neither an int nor tuple.

  • TypeError – If keep_dims is not a bool.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> net = nn.Norm(axis=0)
>>> input = Tensor(np.array([[4, 4, 9, 1], [2, 1, 3, 6]]), mindspore.float32)
>>> output = net(input)
>>> print(output)
[4.472136 4.1231055 9.486833 6.0827627]
class tinyms.layers.OneHot(axis=-1, depth=1, on_value=1.0, off_value=0.0, dtype=mindspore.float32)[源代码]

Returns a one-hot tensor.

The locations represented by indices in argument indices take value on_value, while all other locations take value off_value.

注解

If the input indices is rank \(N\), the output will have rank \(N+1\). The new axis is created at dimension axis.

If indices is a scalar, the output shape will be a vector of length depth.

If indices is a vector of length features, the output shape will be:

features * depth if axis == -1

depth * features if axis == 0

If indices is a matrix with shape [batch, features], the output shape will be:

batch * features * depth if axis == -1

batch * depth * features if axis == 1

depth * batch * features if axis == 0
参数

  • axis (int) – Features x depth if axis is -1, depth x features if axis is 0. Default: -1.

  • depth (int) – A scalar defining the depth of the one hot dimension. Default: 1.

  • on_value (float) – A scalar defining the value to fill in output[i][j] when indices[j] = i. Default: 1.0.

  • off_value (float) – A scalar defining the value to fill in output[i][j] when indices[j] != i. Default: 0.0.

  • dtype (mindspore.dtype) – Data type of ‘on_value’ and ‘off_value’, not the data type of indices. Default: mindspore.float32.

Inputs:

  • indices (Tensor) - A tensor of indices with data type of int32 or int64 and arbitrary shape.

Outputs:

Tensor, the one-hot tensor of data type dtype with dimension at axis expanded to depth and filled with on_value and off_value.

引发

  • TypeError – If axis or depth is not an int.

  • TypeError – If dtype of indices is neither int32 nor int64.

  • ValueError – If axis is not in range [-1, len(indices_shape)].

  • ValueError – If depth is less than 0.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> net = nn.OneHot(depth=4, axis=1)
>>> indices = Tensor([[1, 3], [0, 2]], dtype=mindspore.int32)
>>> output = net(indices)
>>> print(output)
[[[0. 0.]
  [1. 0.]
  [0. 0.]
  [0. 1.]]
 [[1. 0.]
  [0. 0.]
  [0. 1.]
  [0. 0.]]]
class tinyms.layers.Pad(paddings, mode='CONSTANT')[源代码]

Pads the input tensor according to the paddings and mode.

参数

  • paddings (tuple) – The shape of parameter paddings is (N, 2). N is the rank of input data. All elements of paddings are int type. For D th dimension of input, paddings[D, 0] indicates how many sizes to be extended ahead of the D th dimension of the input tensor, and paddings[D, 1] indicates how many sizes to be extended behind of the D th dimension of the input tensor. The padded size of each dimension D of the output is: \(paddings[D, 0] + input\_x.dim\_size(D) + paddings[D, 1]\)

  • mode (str) – Specifies padding mode. The optional values are “CONSTANT”, “REFLECT”, “SYMMETRIC”. Default: “CONSTANT”.

Inputs:

  • input_x (Tensor) - The input tensor.

Outputs:

Tensor, the tensor after padding.

  • If mode is “CONSTANT”, it fills the edge with 0, regardless of the values of the input_x. If the input_x is [[1,2,3], [4,5,6], [7,8,9]] and paddings is [[1,1], [2,2]], then the Outputs is [[0,0,0,0,0,0,0], [0,0,1,2,3,0,0], [0,0,4,5,6,0,0], [0,0,7,8,9,0,0], [0,0,0,0,0,0,0]].

  • If mode is “REFLECT”, it uses a way of symmetrical copying through the axis of symmetry to fill in. If the input_x is [[1,2,3], [4,5,6], [7,8,9]] and paddings is [[1,1], [2,2]], then the Outputs is [[6,5,4,5,6,5,4], [3,2,1,2,3,2,1], [6,5,4,5,6,5,4], [9,8,7,8,9,8,7], [6,5,4,5,6,5,4]].

  • If mode is “SYMMETRIC”, the filling method is similar to the “REFLECT”. It is also copied according to the symmetry axis, except that it includes the symmetry axis. If the input_x is [[1,2,3], [4,5,6], [7,8,9]] and paddings is [[1,1], [2,2]], then the Outputs is [[2,1,1,2,3,3,2], [2,1,1,2,3,3,2], [5,4,4,5,6,6,5], [8,7,7,8,9,9,8], [8,7,7,8,9,9,8]].

引发

  • TypeError – If paddings is not a tuple.

  • ValueError – If length of paddings is more than 4 or its shape is not (n, 2).

  • ValueError – If mode is not one of ‘CONSTANT’, ‘REFLECT’, ‘SYMMETRIC’.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> from mindspore import Tensor
>>> from mindspore.ops import operations as P
>>> import mindspore.nn as nn
>>> import numpy as np
>>> class Net(nn.Cell):
...     def __init__(self):
...         super(Net, self).__init__()
...         self.pad = nn.Pad(paddings=((1, 1), (2, 2)), mode="CONSTANT")
...     def construct(self, x):
...         return self.pad(x)
>>> x = np.array([[0.3, 0.5, 0.2], [0.5, 0.7, 0.3]], dtype=np.float32)
>>> pad = Net()
>>> output = pad(Tensor(x))
>>> print(output)
[[0.         0.         0.         0.         0.         0.        0.         ]
 [0.         0.         0.3        0.5        0.2        0.        0.         ]
 [0.         0.         0.5        0.7        0.3        0.        0.         ]
 [0.         0.         0.         0.         0.         0.        0.         ]]
class tinyms.layers.Unfold(ksizes, strides, rates, padding='valid')[源代码]

Extracts patches from images. The input tensor must be a 4-D tensor and the data format is NCHW.

参数

  • ksizes (Union[tuple[int], list[int]]) – The size of sliding window, must be a tuple or a list of integers, and the format is [1, ksize_row, ksize_col, 1].

  • strides (Union[tuple[int], list[int]]) – Distance between the centers of the two consecutive patches, must be a tuple or list of int, and the format is [1, stride_row, stride_col, 1].

  • rates (Union[tuple[int], list[int]]) – In each extracted patch, the gap between the corresponding dimension pixel positions, must be a tuple or a list of integers, and the format is [1, rate_row, rate_col, 1].

  • padding (str) –

    The type of padding algorithm, is a string whose value is “same” or “valid”, not case sensitive. Default: “valid”.

    • same: Means that the patch can take the part beyond the original image, and this part is filled with 0.

    • valid: Means that the taken patch area must be completely covered in the original image.

Inputs:

  • input_x (Tensor) - A 4-D tensor whose shape is [in_batch, in_depth, in_row, in_col] and data type is number.

Outputs:

Tensor, a 4-D tensor whose data type is same as input_x, and the shape is [out_batch, out_depth, out_row, out_col] where out_batch is the same as the in_batch.

\(out\_depth = ksize\_row * ksize\_col * in\_depth\)

\(out\_row = (in\_row - (ksize\_row + (ksize\_row - 1) * (rate\_row - 1))) // stride\_row + 1\)

\(out\_col = (in\_col - (ksize\_col + (ksize\_col - 1) * (rate\_col - 1))) // stride\_col + 1\)

引发

  • TypeError – If ksizes, strides or rates is neither a tuple nor list.

  • ValueError – If shape of ksizes, strides or rates is not (1, x_row, x_col, 1).

  • ValueError – If the second and third element of ksizes, strides or rates is less than 1.

Supported Platforms:

Ascend

实际案例

>>> net = Unfold(ksizes=[1, 2, 2, 1], strides=[1, 2, 2, 1], rates=[1, 2, 2, 1])
>>> image = Tensor(np.ones([2, 3, 6, 6]), dtype=mstype.float16)
>>> output = net(image)
>>> print(output.shape)
(2, 12, 2, 2)
class tinyms.layers.Tril[源代码]

Returns a tensor with elements above the kth diagonal zeroed.

Inputs:

  • x (Tensor) - The input tensor.

  • k (Int) - The index of diagonal. Default: 0

Outputs:

Tensor, has the same type as input x.

引发
Supported Platforms:

Ascend GPU CPU

实际案例

>>> x = Tensor(np.array([[1, 2], [3, 4]]))
>>> tril = nn.Tril()
>>> result = tril(x)
>>> print(result)
[[1   0]
 [3   4]]
class tinyms.layers.Triu[源代码]

Returns a tensor with elements below the kth diagonal zeroed.

Inputs:

  • x (Tensor) - The input tensor.

  • k (Int) - The index of diagonal. Default: 0

Outputs:

Tensor, has the same type as input x.

引发
Supported Platforms:

Ascend GPU CPU

实际案例

>>> x = Tensor(np.array([[1, 2], [3, 4]]))
>>> triu = nn.Triu()
>>> result = triu(x)
>>> print(result)
[[1 2]
 [0 4]]
class tinyms.layers.ResizeBilinear[源代码]

Samples the input tensor to the given size or scale_factor by using bilinear interpolate.

Inputs:

  • x (Tensor) - Tensor to be resized. Input tensor must be a 4-D tensor with shape: math:(batch, channels, height, width), with data type of float16 or float32.

  • size (Union[tuple[int], list[int]]): A tuple or list of 2 int elements ‘(new_height, new_width)’, the new size of the tensor. One and only one of size and scale_factor can be set to None. Default: None.

  • scale_factor (int): The scale factor of new size of the tensor. The value should be positive integer. One and only one of size and scale_factor can be set to None. Default: None.

  • align_corners (bool): If true, rescale input by ‘(new_height - 1) / (height - 1)’, which exactly aligns the 4 corners of images and resized images. If false, rescale by ‘new_height / height’. Default: False.

Outputs:

Resized tensor. If size is set, the result is 4-D tensor with shape:math:(batch, channels, new_height, new_width) in float32. If scale is set, the result is 4-D tensor with shape:math:(batch, channels, scale_factor * height, scale_factor * width) in float32

引发

  • TypeError – If size is not one of tuple, list, None.

  • TypeError – If scale_factor is neither int nor None.

  • TypeError – If align_corners is not a bool.

  • TypeError – If dtype of x is neither float16 nor float32.

  • ValueError – If size and scale_factor are both None or not None.

  • ValueError – If length of shape of x is not equal to 4.

  • ValueError – If scale_factor is an int which is less than 1.

  • ValueError – If size is a list or tuple whose length is not equal to 2.

Supported Platforms:

Ascend CPU

实际案例

>>> tensor = Tensor([[[[1, 2, 3, 4], [5, 6, 7, 8]]]], mindspore.float32)
>>> resize_bilinear = nn.ResizeBilinear()
>>> result = resize_bilinear(tensor, size=(5,5))
>>> print(result.shape)
(1, 1, 5, 5)
class tinyms.layers.MatrixDiag[源代码]

Returns a batched diagonal tensor with a given batched diagonal values.

Assume x has \(k\) dimensions \([I, J, K, ..., N]\), then the output is a tensor of rank \(k+1\) with dimensions \([I, J, K, ..., N, N]\) where: \(output[i, j, k, ..., m, n] = 1\{m=n\} * x[i, j, k, ..., n]\)

Inputs:

  • x (Tensor) - The diagonal values. It can be one of the following data types: float32, float16, int32, int8, and uint8.

Outputs:

Tensor, has the same type as input x. The shape must be x.shape + (x.shape[-1], ).

引发

TypeError – If dtype of x is not one of float32, float16, int32, int8 or uint8.

Supported Platforms:

Ascend

实际案例

>>> x = Tensor(np.array([1, -1]), mindspore.float32)
>>> matrix_diag = nn.MatrixDiag()
>>> output = matrix_diag(x)
>>> print(output)
[[ 1.  0.]
 [ 0. -1.]]
class tinyms.layers.MatrixDiagPart[源代码]

Returns the batched diagonal part of a batched tensor.

Assume x has \(k\) dimensions \([I, J, K, ..., M, N]\), then the output is a tensor of rank \(k-1\) with dimensions \([I, J, K, ..., min(M, N)]\) where: \(output[i, j, k, ..., n] = x[i, j, k, ..., n, n]\)

Inputs:

  • x (Tensor) - The batched tensor. It can be one of the following data types: float32, float16, int32, int8, and uint8.

Outputs:

Tensor, has the same type as input x. The shape must be x.shape[:-2] + [min(x.shape[-2:])].

引发

TypeError – If dtype of x is not one of float32, float16, int32, int8 or uint8.

Supported Platforms:

Ascend

实际案例

>>> x = Tensor([[[-1, 0], [0, 1]], [[-1, 0], [0, 1]], [[-1, 0], [0, 1]]], mindspore.float32)
>>> matrix_diag_part = nn.MatrixDiagPart()
>>> output = matrix_diag_part(x)
>>> print(output)
[[-1.  1.]
 [-1.  1.]
 [-1.  1.]]
class tinyms.layers.MatrixSetDiag[源代码]

Modifies the batched diagonal part of a batched tensor.

Assume x has \(k+1\) dimensions \([I, J, K, ..., M, N]\) and diagonal has \(k\) dimensions \([I, J, K, ..., min(M, N)]\). Then the output is a tensor of rank \(k+1\) with dimensions \([I, J, K, ..., M, N]\) where:

\[output[i, j, k, ..., m, n] = diagnoal[i, j, k, ..., n]\ for\ m == n\]
\[output[i, j, k, ..., m, n] = x[i, j, k, ..., m, n]\ for\ m != n\]
Inputs:

  • x (Tensor) - The batched tensor. Rank k+1, where k >= 1. It can be one of the following data types: float32, float16, int32, int8, and uint8.

  • diagonal (Tensor) - The diagonal values. Must have the same type as input x. Rank k, where k >= 1.

Outputs:

Tensor, has the same type and shape as input x.

引发

  • TypeError – If dtype of x or diagonal is not one of float32, float16, int32, int8 or uint8.

  • ValueError – If length of shape of x is less than 2.

  • ValueError – If x_shape[-2] < x_shape[-1] and x_shape[:-1] != diagonal_shape.

  • ValueError – If x_shape[-2] >= x_shape[-1] and x_shape[:-2] + x_shape[-1:] != diagonal_shape.

Supported Platforms:

Ascend

实际案例

>>> x = Tensor([[[-1, 0], [0, 1]], [[-1, 0], [0, 1]], [[-1, 0], [0, 1]]], mindspore.float32)
>>> diagonal = Tensor([[-1., 2.], [-1., 1.], [-1., 1.]], mindspore.float32)
>>> matrix_set_diag = nn.MatrixSetDiag()
>>> output = matrix_set_diag(x, diagonal)
>>> print(output)
[[[-1.  0.]
  [ 0.  2.]]
 [[-1.  0.]
  [ 0.  1.]]
 [[-1.  0.]
  [ 0.  1.]]]
class tinyms.layers.L1Regularizer(scale)[源代码]

Applies l1 regularization to weights.

l1 regularization makes weights sparsity

\[\text{loss}=\lambda * \text{reduce_sum}(\text{abs}(\omega))\]

注解

scale(regularization factor) should be a number which greater than 0

参数

scale (int, float) – l1 regularization factor which greater than 0.

Inputs:

  • weights (Tensor) - The input tensor

Outputs:

Tensor, which dtype is higher precision data type between mindspore.float32 and weights dtype, and Tensor shape is ()

引发

  • TypeError – If scale is neither an int nor float.

  • ValueError – If scale is not greater than 0.

  • ValueError – If scale is math.inf or math.nan.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> scale = 0.5
>>> net = nn.L1Regularizer(scale)
>>> weights = Tensor(np.array([[1.0, -2.0], [-3.0, 4.0]]).astype(np.float32))
>>> output = net(weights)
>>> print(output.asnumpy())
5.0
class tinyms.layers.Embedding(vocab_size, embedding_size, use_one_hot=False, embedding_table='normal', dtype=mindspore.float32, padding_idx=None)[源代码]

A simple lookup table that stores embeddings of a fixed dictionary and size.

This module is often used to store word embeddings and retrieve them using indices. The input to the module is a list of indices, and the output is the corresponding word embeddings.

注解

When ‘use_one_hot’ is set to True, the type of the input must be mindspore.int32.

参数

  • vocab_size (int) – Size of the dictionary of embeddings.

  • embedding_size (int) – The size of each embedding vector.

  • use_one_hot (bool) – Specifies whether to apply one_hot encoding form. Default: False.

  • embedding_table (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the embedding_table. Refer to class initializer for the values of string when a string is specified. Default: ‘normal’.

  • dtype (mindspore.dtype) – Data type of input. Default: mindspore.float32.

  • padding_idx (int, None) – When the padding_idx encounters index, the output embedding vector of this index will be initialized to zero. Default: None. The feature is inactivated.

Inputs:

  • input (Tensor) - Tensor of shape \((\text{batch_size}, \text{input_length})\). The elements of the Tensor must be integer and not larger than vocab_size. Otherwise the corresponding embedding vector will be zero.

Outputs:

Tensor of shape \((\text{batch_size}, \text{input_length}, \text{embedding_size})\).

引发

  • TypeError – If vocab_size or embedding_size is not an int.

  • TypeError – If use_one_hot is not a bool.

  • ValueError – If padding_idx is an int which not in range [0, vocab_size].

Supported Platforms:

Ascend GPU

实际案例

>>> net = nn.Embedding(20000, 768,  True)
>>> input_data = Tensor(np.ones([8, 128]), mindspore.int32)
>>>
>>> # Maps the input word IDs to word embedding.
>>> output = net(input_data)
>>> result = output.shape
>>> print(result)
(8, 128, 768)
class tinyms.layers.EmbeddingLookup(vocab_size, embedding_size, param_init='normal', target='CPU', slice_mode='batch_slice', manual_shapes=None, max_norm=None, sparse=True, vocab_cache_size=0)[源代码]

Returns a slice of the input tensor based on the specified indices.

注解

When ‘target’ is set to ‘CPU’, this module will use P.EmbeddingLookup().add_prim_attr(‘primitive_target’, ‘CPU’) which specified ‘offset = 0’ to lookup table. When ‘target’ is set to ‘DEVICE’, this module will use P.Gather() which specified ‘axis = 0’ to lookup table. In field slice mode, the manual_shapes must be given. It is a tuple ,where the element is vocab[i], vocab[i] is the row numbers for i-th part.

参数

  • vocab_size (int) – Size of the dictionary of embeddings.

  • embedding_size (int) – The size of each embedding vector.

  • param_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the embedding_table. Refer to class initializer for the values of string when a string is specified. Default: ‘normal’.

  • target (str) – Specifies the target where the op is executed. The value must in [‘DEVICE’, ‘CPU’]. Default: ‘CPU’.

  • slice_mode (str) – The slicing way in semi_auto_parallel/auto_parallel. The value must get through nn.EmbeddingLookup. Default: nn.EmbeddingLookup.BATCH_SLICE.

  • manual_shapes (tuple) – The accompaniment array in field slice mode.

  • max_norm (Union[float, None]) – A maximum clipping value. The data type must be float16, float32 or None. Default: None

  • sparse (bool) – Using sparse mode. When ‘target’ is set to ‘CPU’, ‘sparse’ has to be true. Default: True.

  • vocab_cache_size (int) – Cache size of the dictionary of embeddings. Default: 0. It is valid only in ‘DEVICE’ target. And the moment parameter of corresponding optimizer will also be set to the cache size. In addition, it should be noted that it will cost the ‘DEVICE’ memory, so suggests setting a reasonable value to avoid insufficient memory.

Inputs:

  • input_indices (Tensor) - The shape of tensor is \((y_1, y_2, ..., y_S)\). Specifies the indices of elements of the original Tensor. Values can be out of range of embedding_table, and the exceeding part will be filled with 0 in the output. Values does not support negative and the result is undefined if values are negative. Input_indices must only be a 2d tensor in this interface when run in semi auto parallel/auto parallel mode.

Outputs:

Tensor, the shape of tensor is \((z_1, z_2, ..., z_N)\).

引发

  • TypeError – If vocab_size or embedding_size or vocab_cache_size is not an int.

  • TypeError – If sparse is not a bool or manual_shapes is not a tuple.

  • ValueError – If vocab_size or embedding_size is less than 1.

  • ValueError – If vocab_cache_size is less than 0.

  • ValueError – If target is neither ‘CPU’ nor ‘DEVICE’.

  • ValueError – If slice_mode is not one of ‘batch_slice’ or ‘field_slice’ or ‘table_row_slice’ or ‘table_column_slice’.

  • ValueError – If sparse is False and target is ‘CPU’.

  • ValueError – If slice_mode is ‘field_slice’ and manual_shapes is None.

Supported Platforms:

Ascend CPU

实际案例

>>> input_indices = Tensor(np.array([[1, 0], [3, 2]]), mindspore.int32)
>>> result = nn.EmbeddingLookup(4,2)(input_indices)
>>> print(result.shape)
(2, 2, 2)
class tinyms.layers.MultiFieldEmbeddingLookup(vocab_size, embedding_size, field_size, param_init='normal', target='CPU', slice_mode='batch_slice', feature_num_list=None, max_norm=None, sparse=True, operator='SUM')[源代码]

Returns a slice of input tensor based on the specified indices and the field ids. This operation supports looking up embeddings using multi hot and one hot fields simultaneously.

注解

When ‘target’ is set to ‘CPU’, this module will use P.EmbeddingLookup().add_prim_attr(‘primitive_target’, ‘CPU’) which specified ‘offset = 0’ to lookup table. When ‘target’ is set to ‘DEVICE’, this module will use P.Gather() which specified ‘axis = 0’ to lookup table. The vectors with the same field_ids will be combined by the ‘operator’, such as ‘SUM’, ‘MAX’ and ‘MEAN’. Ensure the input_values of the padded id is zero, so that they can be ignored. The final output will be zeros if the sum of absolute weight of the field is zero. This class only supports [‘table_row_slice’, ‘batch_slice’ and ‘table_column_slice’]. For the operation ‘MAX’ on device Ascend, there is a constrain where batch_size * (seq_length + field_size) < 3500.

参数

  • vocab_size (int) – The size of the dictionary of embeddings.

  • embedding_size (int) – The size of each embedding vector.

  • field_size (int) – The field size of the final outputs.

  • param_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the embedding_table. Refer to class initializer for the values of string when a string is specified. Default: ‘normal’.

  • target (str) – Specifies the target where the op is executed. The value must in [‘DEVICE’, ‘CPU’]. Default: ‘CPU’.

  • slice_mode (str) – The slicing way in semi_auto_parallel/auto_parallel. The value must get through nn.EmbeddingLookup. Default: nn.EmbeddingLookup.BATCH_SLICE.

  • feature_num_list (tuple) – The accompaniment array in field slice mode. This is unused currently.

  • max_norm (Union[float, None]) – A maximum clipping value. The data type must be float16, float32 or None. Default: None

  • sparse (bool) – Using sparse mode. When ‘target’ is set to ‘CPU’, ‘sparse’ has to be true. Default: True.

  • operator (str) – The pooling method for the features in one field. Support ‘SUM, ‘MEAN’ and ‘MAX’

Inputs:

  • input_indices (Tensor) - The shape of tensor is \((batch\_size, seq\_length)\). Specifies the indices of elements of the original Tensor. Input_indices must be a 2d tensor in this interface. Type is Int32, Int64.

  • input_values (Tensor) - The shape of tensor is \((batch\_size, seq\_length)\). Specifies the weights of elements of the input_indices. The lookout vector will multiply with the input_values. Type is Float32.

  • field_ids (Tensor) - The shape of tensor is \((batch\_size, seq\_length)\). Specifies the field id of elements of the input_indices. Type is Int32.

Outputs:

Tensor, the shape of tensor is \((batch\_size, field\_size, embedding\_size)\). Type is Float32.

引发

  • TypeError – If vocab_size or embedding_size or field_size is not an int.

  • TypeError – If sparse is not a bool or feature_num_list is not a tuple.

  • ValueError – If vocab_size or embedding_size or field_size is less than 1.

  • ValueError – If target is neither ‘CPU’ nor ‘DEVICE’.

  • ValueError – If slice_mode is not one of ‘batch_slice’, ‘field_slice’, ‘table_row_slice’, ‘table_column_slice’.

  • ValueError – If sparse is False and target is ‘CPU’.

  • ValueError – If slice_mode is ‘field_slice’ and feature_num_list is None.

  • ValueError – If operator is not one of ‘SUM’, ‘MAX’, ‘MEAN’.

Supported Platforms:

Ascend GPU

实际案例

>>> input_indices = Tensor([[2, 4, 6, 0, 0], [1, 3, 5, 0, 0]], mindspore.int32)
>>> input_values = Tensor([[1, 1, 1, 0, 0], [1, 1, 1, 0, 0]], mindspore.float32)
>>> field_ids = Tensor([[0, 1, 1, 0, 0], [0, 0, 1, 0, 0]], mindspore.int32)
>>> net = nn.MultiFieldEmbeddingLookup(10, 2, field_size=2, operator='SUM', target='DEVICE')
>>> out = net(input_indices, input_values, field_ids)
>>> print(out.shape)
(2, 2, 2)
class tinyms.layers.AvgPool2d(kernel_size=1, stride=1, pad_mode='valid', data_format='NCHW')[源代码]

2D average pooling for temporal data.

Applies a 2D average pooling over an input Tensor which can be regarded as a composition of 2D input planes.

Typically the input is of shape \((N_{in}, C_{in}, H_{in}, W_{in})\), AvgPool2d outputs regional average in the \((H_{in}, W_{in})\)-dimension. Given kernel size \(ks = (h_{ker}, w_{ker})\) and stride \(s = (s_0, s_1)\), the operation is as follows.

\[\text{output}(N_i, C_j, h, w) = \frac{1}{h_{ker} * w_{ker}} \sum_{m=0}^{h_{ker}-1} \sum_{n=0}^{w_{ker}-1} \text{input}(N_i, C_j, s_0 \times h + m, s_1 \times w + n)\]

注解

pad_mode for training only supports “same” and “valid”.

参数

  • kernel_size (Union[int, tuple[int]]) – The size of kernel used to take the average value. The data type of kernel_size must be int and the value represents the height and width, or a tuple of two int numbers that represent height and width respectively. Default: 1.

  • stride (Union[int, tuple[int]]) – The distance of kernel moving, an int number that represents the height and width of movement are both strides, or a tuple of two int numbers that represent height and width of movement respectively. Default: 1.

  • pad_mode (str) –

    The optional value for pad mode, is “same” or “valid”, not case sensitive. Default: “valid”.

    • same: Adopts the way of completion. The height and width of the output will be the same as the input. The total number of padding will be calculated in horizontal and vertical directions and evenly distributed to top and bottom, left and right if possible. Otherwise, the last extra padding will be done from the bottom and the right side.

    • valid: Adopts the way of discarding. The possible largest height and width of output will be returned without padding. Extra pixels will be discarded.

  • data_format (str) – The optional value for data format, is ‘NHWC’ or ‘NCHW’. Default: ‘NCHW’.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in}, H_{in}, W_{in})\).

Outputs:

Tensor of shape \((N, C_{out}, H_{out}, W_{out})\).

引发

  • TypeError – If kernel_size or strides is neither int nor tuple.

  • ValueError – If pad_mode is neither ‘valid’ nor ‘same’ with not case sensitive.

  • ValueError – If data_format is neither ‘NCHW’ nor ‘NHWC’.

  • ValueError – If kernel_size or strides is less than 1.

  • ValueError – If length of shape of input is not equal to 4.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> pool = nn.AvgPool2d(kernel_size=3, stride=1)
>>> x = Tensor(np.random.randint(0, 10, [1, 2, 4, 4]), mindspore.float32)
>>> output = pool(x)
>>> print(output.shape)
(1, 2, 2, 2)
class tinyms.layers.MaxPool2d(kernel_size=1, stride=1, pad_mode='valid', data_format='NCHW')[源代码]

2D max pooling operation for temporal data.

Applies a 2D max pooling over an input Tensor which can be regarded as a composition of 2D planes.

Typically the input is of shape \((N_{in}, C_{in}, H_{in}, W_{in})\), MaxPool2d outputs regional maximum in the \((H_{in}, W_{in})\)-dimension. Given kernel size \(ks = (h_{ker}, w_{ker})\) and stride \(s = (s_0, s_1)\), the operation is as follows.

\[\text{output}(N_i, C_j, h, w) = \max_{m=0, \ldots, h_{ker}-1} \max_{n=0, \ldots, w_{ker}-1} \text{input}(N_i, C_j, s_0 \times h + m, s_1 \times w + n)\]

注解

pad_mode for training only supports “same” and “valid”.

参数

  • kernel_size (Union[int, tuple[int]]) – The size of kernel used to take the max value, is an int number that represents height and width are both kernel_size, or a tuple of two int numbers that represent height and width respectively. Default: 1.

  • stride (Union[int, tuple[int]]) – The distance of kernel moving, an int number that represents the height and width of movement are both strides, or a tuple of two int numbers that represent height and width of movement respectively. Default: 1.

  • pad_mode (str) –

    The optional value for pad mode, is “same” or “valid”, not case sensitive. Default: “valid”.

    • same: Adopts the way of completion. The height and width of the output will be the same as the input. The total number of padding will be calculated in horizontal and vertical directions and evenly distributed to top and bottom, left and right if possible. Otherwise, the last extra padding will be done from the bottom and the right side.

    • valid: Adopts the way of discarding. The possible largest height and width of output will be returned without padding. Extra pixels will be discarded.

  • data_format (str) – The optional value for data format, is ‘NHWC’ or ‘NCHW’. Default: ‘NCHW’.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in}, H_{in}, W_{in})\).

Outputs:

Tensor of shape \((N, C_{out}, H_{out}, W_{out})\).

引发

  • TypeError – If kernel_size or strides is neither int nor tuple.

  • ValueError – If pad_mode is neither ‘valid’ nor ‘same’ with not case sensitive.

  • ValueError – If data_format is neither ‘NCHW’ nor ‘NHWC’.

  • ValueError – If kernel_size or strides is less than 1.

  • ValueError – If length of shape of input is not equal to 4.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> pool = nn.MaxPool2d(kernel_size=3, stride=1)
>>> x = Tensor(np.random.randint(0, 10, [1, 2, 4, 4]), mindspore.float32)
>>> output = pool(x)
>>> print(output.shape)
(1, 2, 2, 2)
class tinyms.layers.AvgPool1d(kernel_size=1, stride=1, pad_mode='valid')[源代码]

1D average pooling for temporal data.

Applies a 1D average pooling over an input Tensor which can be regarded as a composition of 1D input planes.

Typically the input is of shape \((N_{in}, C_{in}, L_{in})\), AvgPool1d outputs regional average in the \((L_{in})\)-dimension. Given kernel size \(ks = l_{ker}\) and stride \(s = s_0\), the operation is as follows.

\[\text{output}(N_i, C_j, l) = \frac{1}{l_{ker}} \sum_{n=0}^{l_{ker}-1} \text{input}(N_i, C_j, s_0 \times l + n)\]

注解

pad_mode for training only supports “same” and “valid”.

参数

  • kernel_size (int) – The size of kernel window used to take the average value, Default: 1.

  • stride (int) – The distance of kernel moving, an int number that represents the width of movement is strides, Default: 1.

  • pad_mode (str) –

    The optional value for pad mode, is “same” or “valid”, not case sensitive. Default: “valid”.

    • same: Adopts the way of completion. The height and width of the output will be the same as the input. The total number of padding will be calculated in horizontal and vertical directions and evenly distributed to top and bottom, left and right if possible. Otherwise, the last extra padding will be done from the bottom and the right side.

    • valid: Adopts the way of discarding. The possible largest height and width of output will be returned without padding. Extra pixels will be discarded.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in}, L_{in})\).

Outputs:

Tensor of shape \((N, C_{out}, L_{out})\).

引发

  • TypeError – If kernel_size or stride is not an int.

  • ValueError – If pad_mode is neither ‘same’ nor ‘valid’ with not case sensitive.

  • ValueError – If kernel_size or strides is less than 1.

  • ValueError – If length of shape of input is not equal to 3.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> pool = nn.AvgPool1d(kernel_size=6, stride=1)
>>> x = Tensor(np.random.randint(0, 10, [1, 3, 6]), mindspore.float32)
>>> output = pool(x)
>>> result = output.shape
>>> print(result)
(1, 3, 1)
class tinyms.layers.MaxPool1d(kernel_size=1, stride=1, pad_mode='valid')[源代码]

1D max pooling operation for temporal data.

Applies a 1D max pooling over an input Tensor which can be regarded as a composition of 1D planes.

Typically the input is of shape \((N_{in}, C_{in}, L_{in})\), MaxPool1d outputs regional maximum in the \((L_{in})\)-dimension. Given kernel size \(ks = (l_{ker})\) and stride \(s = (s_0)\), the operation is as follows.

\[\text{output}(N_i, C_j, l) = \max_{n=0, \ldots, l_{ker}-1} \text{input}(N_i, C_j, s_0 \times l + n)\]

注解

pad_mode for training only supports “same” and “valid”.

参数

  • kernel_size (int) – The size of kernel used to take the max value, Default: 1.

  • stride (int) – The distance of kernel moving, an int number that represents the width of movement is stride, Default: 1.

  • pad_mode (str) –

    The optional value for pad mode, is “same” or “valid”, not case sensitive. Default: “valid”.

    • same: Adopts the way of completion. The total number of padding will be calculated in horizontal and vertical directions and evenly distributed to top and bottom, left and right if possible. Otherwise, the last extra padding will be done from the bottom and the right side.

    • valid: Adopts the way of discarding. The possible largest height and width of output will be returned without padding. Extra pixels will be discarded.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C, L_{in})\).

Outputs:

Tensor of shape \((N, C, L_{out}))\).

引发

  • TypeError – If kernel_size or strides is not an int.

  • ValueError – If pad_mode is neither ‘valid’ nor ‘same’ with not case sensitive.

  • ValueError – If data_format is neither ‘NCHW’ nor ‘NHWC’.

  • ValueError – If kernel_size or strides is less than 1.

  • ValueError – If length of shape of input is not equal to 4.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> max_pool = nn.MaxPool1d(kernel_size=3, stride=1)
>>> x = Tensor(np.random.randint(0, 10, [1, 2, 4]), mindspore.float32)
>>> output = max_pool(x)
>>> result = output.shape
>>> print(result)
(1, 2, 2)
class tinyms.layers.ReduceLogSumExp(axis, keep_dims=False)[源代码]

Reduces a dimension of a tensor by calculating exponential for all elements in the dimension, then calculate logarithm of the sum.

The dtype of the tensor to be reduced is number.

\[ReduceLogSumExp(x) = \log(\sum(e^x))\]
参数

  • axis (Union[int, tuple(int), list(int)]) – (), reduce all dimensions. Only constant value is allowed.

  • keep_dims (bool) – If True, keep these reduced dimensions and the length is 1. If False, don’t keep these dimensions. Default : False.

Inputs:

  • x (Tensor) - The input tensor. With float16 or float32 data type.

Outputs:

Tensor, has the same dtype as the x.

  • If axis is (), and keep_dims is False, the output is a 0-D tensor representing the sum of all elements in the input tensor.

  • If axis is int, set as 2, and keep_dims is False, the shape of output is \((x_1, x_3, ..., x_R)\).

  • If axis is tuple(int), set as (2, 3), and keep_dims is False, the shape of output is \((x_1, x_4, ..., x_R)\).

引发

  • TypeError – If axis is not one of int, list, tuple.

  • TypeError – If keep_dims is not bool.

  • TypeError – If dtype of x is neither float16 nor float32.

Supported Platforms:

Ascend GPU

实际案例

>>> input_x = Tensor(np.random.randn(3, 4, 5, 6).astype(np.float32))
>>> op = nn.ReduceLogSumExp(1, keep_dims=True)
>>> output = op(input_x)
>>> print(output.shape)
(3, 1, 5, 6)
class tinyms.layers.Range(start, limit=None, delta=1)[源代码]

Creates a sequence of numbers in range [start, limit) with step size delta.

The size of output is \(\left \lfloor \frac{limit-start}{delta} \right \rfloor + 1\) and delta is the gap between two values in the tensor.

\[out_{i+1} = out_{i} +delta\]
参数

  • start (Union[int, float]) – If limit is None, the value acts as limit in the range and first entry defaults to 0. Otherwise, it acts as first entry in the range.

  • limit (Union[int, float]) – Acts as upper limit of sequence. If None, defaults to the value of start while set the first entry of the range to 0. It can not be equal to start.

  • delta (Union[int, float]) – Increment of the range. It can not be equal to zero. Default: 1.

Outputs:

Tensor, the dtype is int if the dtype of start, limit and delta all are int. Otherwise, dtype is float.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> net = nn.Range(1, 8, 2)
>>> output = net()
>>> print(output)
[1 3 5 7]
class tinyms.layers.LGamma[源代码]

Calculates LGamma using Lanczos’ approximation referring to “A Precision Approximation of the Gamma Function”. The algorithm is:

\[\begin{split}\begin{array}{ll} \\ lgamma(z + 1) = \frac{(\log(2) + \log(pi))}{2} + (z + 1/2) * log(t(z)) - t(z) + A(z) \\ t(z) = z + kLanczosGamma + 1/2 \\ A(z) = kBaseLanczosCoeff + \sum_{k=1}^n \frac{kLanczosCoefficients[i]}{z + k} \end{array}\end{split}\]

However, if the input is less than 0.5 use Euler’s reflection formula:

\[lgamma(x) = \log(pi) - lgamma(1-x) - \log(abs(sin(pi * x)))\]

And please note that

\[lgamma(+/-inf) = +inf\]

Thus, the behaviour of LGamma follows: when x > 0.5, return log(Gamma(x)) when x < 0.5 and is not an integer, return the real part of Log(Gamma(x)) where Log is the complex logarithm when x is an integer less or equal to 0, return +inf when x = +/- inf, return +inf

Inputs:

  • x (Tensor) - The input tensor. Only float16, float32 are supported.

Outputs:

Tensor, has the same shape and dtype as the x.

引发

TypeError – If dtype of x is neither float16 nor float32.

Supported Platforms:

Ascend GPU

实际案例

>>> input_x = Tensor(np.array([2, 3, 4]).astype(np.float32))
>>> op = nn.LGamma()
>>> output = op(input_x)
>>> print(output)
[3.5762787e-07 6.9314754e-01 1.7917603e+00]
class tinyms.layers.DiGamma[源代码]

Calculates Digamma using Lanczos’ approximation referring to “A Precision Approximation of the Gamma Function”. The algorithm is:

\[\begin{split}\begin{array}{ll} \\ digamma(z + 1) = log(t(z)) + A'(z) / A(z) - kLanczosGamma / t(z) \\ t(z) = z + kLanczosGamma + 1/2 \\ A(z) = kBaseLanczosCoeff + \sum_{k=1}^n \frac{kLanczosCoefficients[i]}{z + k} \\ A'(z) = \sum_{k=1}^n \frac{kLanczosCoefficients[i]}{{z + k}^2} \end{array}\end{split}\]

However, if the input is less than 0.5 use Euler’s reflection formula:

\[digamma(x) = digamma(1 - x) - pi * cot(pi * x)\]
Inputs:

  • x (Tensor[Number]) - The input tensor. Only float16, float32 are supported.

Outputs:

Tensor, has the same shape and dtype as the x.

引发

TypeError – If dtype of x is neither float16 nor float32.

Supported Platforms:

Ascend GPU

实际案例

>>> input_x = Tensor(np.array([2, 3, 4]).astype(np.float32))
>>> op = nn.DiGamma()
>>> output = op(input_x)
>>> print(output)
[0.42278463  0.92278427 1.2561178]
class tinyms.layers.IGamma[源代码]

Calculates lower regularized incomplete Gamma function. The lower regularized incomplete Gamma function is defined as:

\[P(a, x) = gamma(a, x) / Gamma(a) = 1 - Q(a, x)\]

where

\[gamma(a, x) = \int_0^x t^{a-1} \exp^{-t} dt\]

is the lower incomplete Gamma function.

Above \(Q(a, x)\) is the upper regularized complete Gamma function.

Inputs:

  • a (Tensor) - The input tensor. With float32 data type. a should have the same dtype with x.

  • x (Tensor) - The input tensor. With float32 data type. x should have the same dtype with a.

Outputs:

Tensor, has the same dtype as a and x.

引发

TypeError – If dtype of input x and a is not float16 nor float32, or if x has different dtype with a.

Supported Platforms:

Ascend GPU

实际案例

>>> input_a = Tensor(np.array([2.0, 4.0, 6.0, 8.0]).astype(np.float32))
>>> input_x = Tensor(np.array([2.0, 3.0, 4.0, 5.0]).astype(np.float32))
>>> igamma = nn.IGamma()
>>> output = igamma(input_a, input_x)
>>> print (output)
[0.593994  0.35276785  0.21486944  0.13337152]
class tinyms.layers.LBeta[源代码]

This is semantically equal to

\[P(x, y) = lgamma(x) + lgamma(y) - lgamma(x + y).\]

The method is more accurate for arguments above 8. The reason for accuracy loss in the naive computation is catastrophic cancellation between the lgammas. This method avoids the numeric cancellation by explicitly decomposing lgamma into the Stirling approximation and an explicit log_gamma_correction, and cancelling the large terms from the Striling analytically.

Inputs:

  • x (Tensor) - The input tensor. With float16 or float32 data type. x should have the same dtype with y.

  • y (Tensor) - The input tensor. With float16 or float32 data type. y should have the same dtype with x.

Outputs:

Tensor, has the same dtype as x and y.

引发

TypeError – If dtype of x or y is neither float16 nor float32, or if x has different dtype with y.

Supported Platforms:

Ascend GPU

实际案例

>>> input_x = Tensor(np.array([2.0, 4.0, 6.0, 8.0]).astype(np.float32))
>>> input_y = Tensor(np.array([2.0, 3.0, 14.0, 15.0]).astype(np.float32))
>>> lbeta = nn.LBeta()
>>> output = lbeta(input_y, input_x)
>>> print(output)
[-1.7917596  -4.094345  -12.000229  -14.754799]
class tinyms.layers.MatMul(**kwargs)[源代码]

Multiplies matrix x1 by matrix x2.

nn.MatMul will be deprecated in future versions. Please use ops.matmul instead.

  • If both x1 and x2 are 1-dimensional, the dot product is returned.

  • If the dimensions of x1 and x2 are all not greater than 2, the matrix-matrix product will be returned. Note if one of ‘x1’ and ‘x2’ is 1-dimensional, the argument will first be expanded to 2 dimension. After the matrix multiply, the expanded dimension will be removed.

  • If at least one of x1 and x2 is N-dimensional (N>2), the none-matrix dimensions(batch) of inputs will be broadcasted and must be broadcastable. Note if one of ‘x1’ and ‘x2’ is 1-dimensional, the argument will first be expanded to 2 dimension and then the none-matrix dimensions will be broadcasted. After the matrix multiply, the expanded dimension will be removed. For example, if x1 is a \((j \times 1 \times n \times m)\) tensor and x2 is a \((k \times m \times p)\) tensor, the output will be a \((j \times k \times n \times p)\) tensor.

参数

  • transpose_x1 (bool) – If true, a is transposed before multiplication. Default: False.

  • transpose_x2 (bool) – If true, b is transposed before multiplication. Default: False.

Inputs:

  • input_x1 (Tensor) - The first tensor to be multiplied.

  • input_x2 (Tensor) - The second tensor to be multiplied.

Outputs:

Tensor, the shape of the output tensor depends on the dimension of input tensors.

引发

  • TypeError – If transpose_x1 or transpose_x2 is not a bool.

  • ValueError – If the column of matrix dimensions of input_x1 is not equal to the row of matrix dimensions of input_x2.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> net = nn.MatMul()
>>> input_x1 = Tensor(np.ones(shape=[3, 2, 3]), mindspore.float32)
>>> input_x2 = Tensor(np.ones(shape=[3, 4]), mindspore.float32)
>>> output = net(input_x1, input_x2)
>>> print(output.shape)
(3, 2, 4)
class tinyms.layers.Moments(axis=None, keep_dims=None)[源代码]

Calculates the mean and variance of x.

参数

  • axis (Union[int, tuple(int)]) – Calculates the mean and variance along the specified axis. Default: ().

  • keep_dims (bool) – If true, The dimension of mean and variance are identical with input’s. If false, don’t keep these dimensions. Default: False.

Inputs:

  • input_x (Tensor) - The tensor to be calculated. Only float16 and float32 are supported.

Outputs:

  • mean (Tensor) - The mean of input x, with the same date type as input x.

  • variance (Tensor) - The variance of input x, with the same date type as input x.

引发

  • TypeError – If axis is not one of int, tuple, None.

  • TypeError – If keep_dims is neither bool nor None.

  • TypeError – If dtype of input_x is neither float16 nor float32.

Supported Platforms:

Ascend GPU

实际案例

>>> net = nn.Moments(axis=3, keep_dims=True)
>>> input_x = Tensor(np.array([[[[1, 2, 3, 4], [3, 4, 5, 6]]]]), mindspore.float32)
>>> output = net(input_x)
>>> print(output)
(Tensor(shape=[1, 1, 2, 1], dtype=Float32, value=
[[[[ 2.50000000e+00],
   [ 4.50000000e+00]]]]), Tensor(shape=[1, 1, 2, 1], dtype=Float32, value=
[[[[ 1.25000000e+00],
   [ 1.25000000e+00]]]]))
class tinyms.layers.MatInverse[源代码]

Calculates the inverse of Positive-Definite Hermitian matrix using Cholesky decomposition.

Inputs:

  • a (Tensor[Number]) - The input tensor. It must be a positive-definite matrix. With float16 or float32 data type.

Outputs:

Tensor, has the same dtype as the a.

引发

TypeError – If dtype of a is neither float16 nor float32.

Supported Platforms:

GPU

实际案例

>>> input_a = Tensor(np.array([[4, 12, -16], [12, 37, -43], [-16, -43, 98]]).astype(np.float32))
>>> op = nn.MatInverse()
>>> output = op(input_a)
>>> print(output)
[[49.36112  -13.555558  2.1111116]
 [-13.555558  3.7777784  -0.5555557]
 [2.1111116  -0.5555557  0.11111111]]
class tinyms.layers.MatDet[源代码]

Calculates the determinant of Positive-Definite Hermitian matrix using Cholesky decomposition.

Inputs:

  • a (Tensor[Number]) - The input tensor. It must be a positive-definite matrix. With float16 or float32 data type.

Outputs:

Tensor, has the same dtype as the a.

引发

TypeError – If dtype of a is neither float16 nor float32.

Supported Platforms:

GPU

实际案例

>>> input_a = Tensor(np.array([[4, 12, -16], [12, 37, -43], [-16, -43, 98]]).astype(np.float32))
>>> op = nn.MatDet()
>>> output = op(input_a)
>>> print(output)
35.999996
class tinyms.layers.Conv2dBnAct(in_channels, out_channels, kernel_size, stride=1, pad_mode='same', padding=0, dilation=1, group=1, has_bias=False, weight_init='normal', bias_init='zeros', has_bn=False, momentum=0.997, eps=1e-05, activation=None, alpha=0.2, after_fake=True)[源代码]

A combination of convolution, Batchnorm, and activation layer.

This part is a more detailed overview of Conv2d operation.

参数

  • in_channels (int) – The number of input channel \(C_{in}\).

  • out_channels (int) – The number of output channel \(C_{out}\).

  • kernel_size (Union[int, tuple]) – The data type is int or a tuple of 2 integers. Specifies the height and width of the 2D convolution window. Single int means the value is for both height and width of the kernel. A tuple of 2 ints means the first value is for the height and the other is for the width of the kernel.

  • stride (int) – Specifies stride for all spatial dimensions with the same value. The value of stride must be greater than or equal to 1 and lower than any one of the height and width of the input. Default: 1.

  • pad_mode (str) – Specifies padding mode. The optional values are “same”, “valid”, “pad”. Default: “same”.

  • padding (int) – Implicit paddings on both sides of the input. Default: 0.

  • dilation (int) – Specifies the dilation rate to use for dilated convolution. If set to be \(k > 1\), there will be \(k - 1\) pixels skipped for each sampling location. Its value must be greater than or equal to 1 and lower than any one of the height and width of the input. Default: 1.

  • group (int) – Splits filter into groups, in_ channels and out_channels must be divisible by the number of groups. Default: 1.

  • has_bias (bool) – Specifies whether the layer uses a bias vector. Default: False.

  • weight_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the convolution kernel. It can be a Tensor, a string, an Initializer or a number. When a string is specified, values from ‘TruncatedNormal’, ‘Normal’, ‘Uniform’, ‘HeUniform’ and ‘XavierUniform’ distributions as well as constant ‘One’ and ‘Zero’ distributions are possible. Alias ‘xavier_uniform’, ‘he_uniform’, ‘ones’ and ‘zeros’ are acceptable. Uppercase and lowercase are both acceptable. Refer to the values of Initializer for more details. Default: ‘normal’.

  • bias_init (Union[Tensor, str, Initializer, numbers.Number]) – Initializer for the bias vector. Possible Initializer and string are the same as ‘weight_init’. Refer to the values of Initializer for more details. Default: ‘zeros’.

  • has_bn (bool) – Specifies to used batchnorm or not. Default: False.

  • momentum (float) – Momentum for moving average for batchnorm, must be [0, 1]. Default:0.9

  • eps (float) – Term added to the denominator to improve numerical stability for batchnorm, should be greater than 0. Default: 1e-5.

  • activation (Union[str, Cell, Primitive]) – Specifies activation type. The optional values are as following: ‘softmax’, ‘logsoftmax’, ‘relu’, ‘relu6’, ‘tanh’, ‘gelu’, ‘sigmoid’, ‘prelu’, ‘leakyrelu’, ‘hswish’, ‘hsigmoid’. Default: None.

  • alpha (float) – Slope of the activation function at x < 0 for LeakyReLU. Default: 0.2.

  • after_fake (bool) – Determine whether there must be a fake quantization operation after Cond2dBnAct.

Inputs:

  • input (Tensor) - Tensor of shape \((N, C_{in}, H_{in}, W_{in})\).

Outputs:

Tensor of shape \((N, C_{out}, H_{out}, W_{out})\).

引发

  • TypeError – If in_channels, out_channels, stride, padding or dilation is not an int.

  • TypeError – If has_bias is not a bool.

  • ValueError – If in_channels or out_channels stride, padding or dilation is less than 1.

  • ValueError – If pad_mode is not one of ‘same’, ‘valid’, ‘pad’.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> net = nn.Conv2dBnAct(120, 240, 4, has_bn=True, activation='relu')
>>> input = Tensor(np.ones([1, 120, 1024, 640]), mindspore.float32)
>>> result = net(input)
>>> output = result.shape
>>> print(output)
(1, 240, 1024, 640)
class tinyms.layers.DenseBnAct(in_channels, out_channels, weight_init='normal', bias_init='zeros', has_bias=True, has_bn=False, momentum=0.9, eps=1e-05, activation=None, alpha=0.2, after_fake=True)[源代码]

A combination of Dense, Batchnorm, and the activation layer.

This part is a more detailed overview of Dense op.

参数

  • in_channels (int) – The number of channels in the input space.

  • out_channels (int) – The number of channels in the output space.

  • weight_init (Union[Tensor, str, Initializer, numbers.Number]) – The trainable weight_init parameter. The dtype is same as input. The values of str refer to the function initializer. Default: ‘normal’.

  • bias_init (Union[Tensor, str, Initializer, numbers.Number]) – The trainable bias_init parameter. The dtype is same as input. The values of str refer to the function initializer. Default: ‘zeros’.

  • has_bias (bool) – Specifies whether the layer uses a bias vector. Default: True.

  • has_bn (bool) – Specifies to use batchnorm or not. Default: False.

  • momentum (float) – Momentum for moving average for batchnorm, must be [0, 1]. Default:0.9

  • eps (float) – Term added to the denominator to improve numerical stability for batchnorm, should be greater than 0. Default: 1e-5.

  • activation (Union[str, Cell, Primitive]) – Specifies activation type. The optional values are as following: ‘softmax’, ‘logsoftmax’, ‘relu’, ‘relu6’, ‘tanh’, ‘gelu’, ‘sigmoid’, ‘prelu’, ‘leakyrelu’, ‘hswish’, ‘hsigmoid’. Default: None.

  • alpha (float) – Slope of the activation function at x < 0 for LeakyReLU. Default: 0.2.

  • after_fake (bool) – Determine whether there must be a fake quantization operation after DenseBnAct.

Inputs:

  • input (Tensor) - Tensor of shape \((N, in\_channels)\).

Outputs:

Tensor of shape \((N, out\_channels)\).

引发

  • TypeError – If in_channels or out_channels is not an int.

  • TypeError – If has_bias, has_bn or after_fake is not a bool.

  • TypeError – If momentum or eps is not a float.

  • ValueError – If momentum is not in range [0, 1.0].

Supported Platforms:

Ascend GPU

实际案例

>>> net = nn.DenseBnAct(3, 4)
>>> input = Tensor(np.random.randint(0, 255, [2, 3]), mindspore.float32)
>>> result = net(input)
>>> output = result.shape
>>> print(output)
(2, 4)
class tinyms.layers.TimeDistributed(layer, time_axis, reshape_with_axis=None)[源代码]

The time distributed layer.

Time distributed is a wrapper which allows to apply a layer to every temporal slice of an input. And the input should be at least 3D. There are two cases in the implementation. When reshape_with_axis provided, the reshape method will be chosen, which is more efficient; otherwise, the method of dividing the inputs along time axis will be used, which is more general. For example, reshape_with_axis could not be provided when deal with batch normal.

参数

  • layer (Union[Cell, Primitive]) – The Cell or Primitive which will be wrapped.

  • time_axis (int) – The axis of time_step.

  • reshape_with_axis (int) – The axis which will be reshaped with time_axis. Default: None.

Inputs:

  • input (Tensor) - Tensor of shape \((N, T, *)\).

Outputs:

Tensor of shape \((N, T, *)\)

Supported Platforms:

Ascend GPU CPU

引发

TypeError – If layer is not a Cell or Primitive.

实际案例

>>> input = Tensor(np.random.random([32, 10, 3]), mindspore.float32)
>>> dense = nn.Dense(3, 6)
>>> net = nn.TimeDistributed(dense, time_axis=1, reshape_with_axis=0)
>>> output = net(input)
>>> print(output.shape)
(32, 10, 6)
class tinyms.layers.ForwardValueAndGrad(network, weights=None, get_all=False, get_by_list=False, sens_param=False)[源代码]

Network training package class.

Including the network and a gradient function. The resulting Cell is trained with input ‘*inputs’. The backward graph will be created in the gradient function to calculating gradient.

参数

  • network (Cell) – The training network.

  • weights (ParameterTuple) – The parameters of the training network that need to calculate the gradient.

  • get_all (bool) – If True, get all the gradients with respect to inputs. Default: False.

  • get_by_list (bool) – If True, get all the gradients with respect to Parameter variables. If get_all and get_by_list are both False, get the gradient with respect to first input. If get_all and get_by_list are both True, get the gradients with respect to inputs and Parameter variables at the same time in the form of ((gradients with respect to inputs), (gradients with respect to parameters)). Default: False.

  • sens_param (bool) – Whether to append sensitivity (gradient with respect to output) as input. If sens_param is False, a ‘ones_like(outputs)’ sensitivity will be attached automatically. Default: False. If the sens_param is True, a sensitivity (gradient with respect to output) needs to be transferred through the input parameter.

Inputs:

  • (*inputs) (Tuple(Tensor…)) - Tuple of inputs with shape \((N, \ldots)\).

  • (sens) - A sensitivity (gradient with respect to output) as the input of backpropagation.

    If network has single output, the sens is a tensor. If network has multiple outputs, the sens is the tuple(tensor).

Outputs:

  • forward value - The result of network forward running.

  • gradients (tuple(tensor)) - The gradients of network parameters and inputs.

Supported Platforms:

Ascend GPU CPU

实际案例

>>> class Net(nn.Cell):
...    def __init__(self):
...        super(Net, self).__init__()
...        self.weight = Parameter(Tensor(np.ones([2, 2]).astype(np.float32)), name="weight")
...        self.matmul = P.MatMul()
...
...    def construct(self, x):
...        out = self.matmul(x, self.weight)
...        return out
...
>>> net = Net()
>>> criterion = nn.SoftmaxCrossEntropyWithLogits()
>>> net_with_criterion = nn.WithLossCell(net, criterion)
>>> weight = ParameterTuple(net.trainable_params())
>>> train_network = nn.ForwardValueAndGrad(net_with_criterion, weights=weight, get_all=True, get_by_list=True)
>>> inputs = Tensor(np.ones([1, 2]).astype(np.float32))
>>> labels = Tensor(np.zeros([1, 2]).astype(np.float32))
>>> result = train_network(inputs, labels)
>>> print(result)
(Tensor(shape=[1], dtype=Float32, value=[0.00000000e+00]), ((Tensor(shape=[1, 2], dtype=Float32, value=
[[1.00000000e+00, 1.00000000e+00]]), Tensor(shape=[1, 2], dtype=Float32, value=
[[0.00000000e+00, 0.00000000e+00]])), (Tensor(shape=[2, 2], dtype=Float32, value=
[[5.00000000e-01, 5.00000000e-01],
 [5.00000000e-01, 5.00000000e-01]]),)))
class tinyms.layers.TrainOneStepCell(network, optimizer, sens=1.0)[源代码]

Network training package class.

Wraps the network with an optimizer. The resulting Cell is trained with input ‘*inputs’. The backward graph will be created in the construct function to update the parameter. Different parallel modes are available for training.

参数

  • network (Cell) – The training network. The network only supports single output.

  • optimizer (Cell) – Optimizer for updating the weights.

  • sens (Number) – The scaling number to be filled as the input of backpropagation. Default value is 1.0.

Inputs:

  • (*inputs) (Tuple(Tensor)) - Tuple of input tensors with shape \((N, \ldots)\).

Outputs:

Tensor, a scalar Tensor with shape \(()\).

引发

TypeError – If sens is not a number.

Supported Platforms:

Ascend GPU

实际案例

>>> net = Net()
>>> loss_fn = nn.SoftmaxCrossEntropyWithLogits()
>>> optim = nn.Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
>>> #1) Using the WithLossCell existing provide
>>> loss_net = nn.WithLossCell(net, loss_fn)
>>> train_net = nn.TrainOneStepCell(loss_net, optim)
>>>
>>> #2) Using user-defined WithLossCell
>>> class MyWithLossCell(Cell):
...    def __init__(self, backbone, loss_fn):
...        super(MyWithLossCell, self).__init__(auto_prefix=False)
...        self._backbone = backbone
...        self._loss_fn = loss_fn
...
...    def construct(self, x, y, label):
...        out = self._backbone(x, y)
...        return self._loss_fn(out, label)
...
...    @property
...    def backbone_network(self):
...        return self._backbone
...
>>> loss_net = MyWithLossCell(net, loss_fn)
>>> train_net = nn.TrainOneStepCell(loss_net, optim)
class tinyms.layers.WithLossCell(backbone, loss_fn)[源代码]

Cell with loss function.

Wraps the network with loss function. This Cell accepts data and label as inputs and the computed loss will be returned.

参数

  • backbone (Cell) – The target network to wrap.

  • loss_fn (Cell) – The loss function used to compute loss.

Inputs:

  • data (Tensor) - Tensor of shape \((N, \ldots)\).

  • label (Tensor) - Tensor of shape \((N, \ldots)\).

Outputs:

Tensor, a scalar tensor with shape \(()\).

引发

TypeError – If dtype of data or label is neither float16 nor float32.

Supported Platforms:

Ascend GPU

实际案例

>>> net = Net()
>>> loss_fn = nn.SoftmaxCrossEntropyWithLogits(sparse=False)
>>> net_with_criterion = nn.WithLossCell(net, loss_fn)
>>>
>>> batch_size = 2
>>> data = Tensor(np.ones([batch_size, 1, 32, 32]).astype(np.float32) * 0.01)
>>> label = Tensor(np.ones([batch_size, 10]).astype(np.float32))
>>>
>>> output_data = net_with_criterion(data, label)
property backbone_network

Returns the backbone network.

返回

Cell, the backbone network.

class tinyms.layers.WithGradCell(network, loss_fn=None, sens=None)[源代码]

Cell that returns the gradients.

Wraps the network with backward cell to compute gradients. A network with a loss function is necessary as argument. If loss function in None, the network must be a wrapper of network and loss function. This Cell accepts ‘*inputs’ as inputs and returns gradients for each trainable parameter.

注解

Run in PyNative mode.

参数

  • network (Cell) – The target network to wrap. The network only supports single output.

  • loss_fn (Cell) – Primitive loss function used to compute gradients. Default: None.

  • sens (Union[None, Tensor, Scalar, Tuple ...]) – The sensitive for backpropagation, the type and shape must be same as the network output. If None, we will fill one to a same type shape of output value. Default: None.

Inputs:

  • (*inputs) (Tuple(Tensor)) - Tuple of input tensors with shape \((N, \ldots)\).

Outputs:

list, a list of Tensors with identical shapes as trainable weights.

引发

TypeError – If sens is not one of None, Tensor, Scalar or Tuple.

Supported Platforms:

Ascend GPU

实际案例

>>> # For a defined network Net without loss function
>>> net = Net()
>>> loss_fn = nn.SoftmaxCrossEntropyWithLogits()
>>> grad_net = nn.WithGradCell(net, loss_fn)
>>>
>>> # For a network wrapped with loss function
>>> net = Net()
>>> net_with_criterion = nn.WithLossCell(net, loss_fn)
>>> grad_net = nn.WithGradCell(net_with_criterion)
class tinyms.layers.WithEvalCell(network, loss_fn, add_cast_fp32=False)[源代码]

Cell that returns loss, output and label for evaluation.

This Cell accepts a network and loss function as arguments and computes loss for model. It returns loss, output and label to calculate the metrics.

参数

  • network (Cell) – The network Cell.

  • loss_fn (Cell) – The loss Cell.

  • add_cast_fp32 (bool) – Adjust the data type to float32.

Inputs:

  • data (Tensor) - Tensor of shape \((N, \ldots)\).

  • label (Tensor) - Tensor of shape \((N, \ldots)\).

Outputs:

Tuple, containing a scalar loss Tensor, a network output Tensor of shape \((N, \ldots)\) and a label Tensor of shape \((N, \ldots)\).

引发

TypeError – If add_cast_fp32 is not a bool.

Supported Platforms:

Ascend GPU

实际案例

>>> # For a defined network Net without loss function
>>> net = Net()
>>> loss_fn = nn.SoftmaxCrossEntropyWithLogits()
>>> eval_net = nn.WithEvalCell(net, loss_fn)
class tinyms.layers.GetNextSingleOp(dataset_types, dataset_shapes, queue_name)[源代码]

Cell to run for getting the next operation.

参数

  • dataset_types (list[mindspore.dtype]) – The types of dataset.

  • dataset_shapes (list[tuple[int]]) – The shapes of dataset.

  • queue_name (str) – Queue name to fetch the data.

For detailed information, refer to ops.operations.GetNext.

Supported Platforms:

Ascend GPU

实际案例

>>> train_dataset = create_custom_dataset()
>>> dataset_helper = mindspore.DatasetHelper(train_dataset, dataset_sink_mode=True)
>>> dataset = dataset_helper.iter.dataset
>>> dataset_types, dataset_shapes = dataset_helper.types_shapes()
>>> queue_name = dataset.__transfer_dataset__.queue_name
>>> get_next_single_op_net = nn.GetNextSingleOp(dataset_types, dataset_shapes, queue_name)
>>> data, label = get_next_single_op_net()
>>> relu = P.ReLU()
>>> result = relu(data).asnumpy()
>>> print(result.shape)
(32, 1, 32, 32)
class tinyms.layers.TrainOneStepWithLossScaleCell(network, optimizer, scale_sense)[源代码]

Network training with loss scaling.

This is a training step with loss scaling. It takes a network, an optimizer and possibly a scale update Cell as args. The loss scale value can be updated in both host side or device side. The TrainOneStepWithLossScaleCell will be compiled to be graph which takes *inputs as input data. The Tensor type of scale_sense is acting as loss scaling value. If you want to update it on host side, the value must be provided. If the Tensor type of scale_sense is not given, the loss scale update logic must be provied by Cell type of scale_sense.

参数

  • network (Cell) – The training network. The network only supports single output.

  • optimizer (Cell) – Optimizer for updating the weights.

  • scale_sense (Union[Tensor, Cell]) – If this value is Cell type, the loss scaling update logic cell.If this value is Tensor type, Tensor with shape \(()\) or \((1,)\).

Inputs:

  • (*inputs) (Tuple(Tensor)) - Tuple of input tensors with shape \((N, \ldots)\).

Outputs:

Tuple of 3 Tensor, the loss, overflow flag and current loss scaling value.

  • loss (Tensor) - Tensor with shape \(()\).

  • overflow (Tensor) - Tensor with shape \(()\), type is bool.

  • loss scaling value (Tensor) - Tensor with shape \(()\)

引发

  • TypeError – If scale_sense is neither Cell nor Tensor.

  • ValueError – If shape of scale_sense is neither (1,) nor ().

Supported Platforms:

Ascend GPU

实际案例

>>> import numpy as np
>>> from mindspore import Tensor, Parameter, nn
>>> from mindspore.ops import operations as P
>>> from mindspore.nn.wrap.cell_wrapper import WithLossCell
>>> from mindspore.common import dtype as mstype
>>>
>>> class Net(nn.Cell):
...     def __init__(self, in_features, out_features):
...         super(Net, self).__init__()
...         self.weight = Parameter(Tensor(np.ones([in_features, out_features]).astype(np.float32)),
...                                 name='weight')
...         self.matmul = P.MatMul()
...
...     def construct(self, x):
...         output = self.matmul(x, self.weight)
...         return output
...
>>> size, in_features, out_features = 16, 16, 10
>>> #1) when the type of scale_sense is Cell:
>>> net = Net(in_features, out_features)
>>> loss = nn.MSELoss()
>>> optimizer = nn.Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
>>> net_with_loss = WithLossCell(net, loss)
>>> manager = nn.DynamicLossScaleUpdateCell(loss_scale_value=2**12, scale_factor=2, scale_window=1000)
>>> train_network = nn.TrainOneStepWithLossScaleCell(net_with_loss, optimizer, scale_sense=manager)
>>> input = Tensor(np.ones([out_features, in_features]), mindspore.float32)
>>> labels = Tensor(np.ones([out_features,]), mindspore.float32)
>>> output = train_network(input, labels)
>>>
>>> #2) when the type of scale_sense is Tensor:
>>> net = Net(in_features, out_features)
>>> loss = nn.MSELoss()
>>> optimizer = nn.Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
>>> net_with_loss = WithLossCell(net, loss)
>>> inputs = Tensor(np.ones([size, in_features]).astype(np.float32))
>>> label = Tensor(np.zeros([size, out_features]).astype(np.float32))
>>> scaling_sens = Tensor(np.full((1), np.finfo(np.float32).max), dtype=mstype.float32)
>>> train_network = nn.TrainOneStepWithLossScaleCell(net_with_loss, optimizer, scale_sense=scaling_sens)
>>> output = train_network(inputs, label)
get_overflow_status(status, compute_output)[源代码]

Get floating-point overflow status.

Get overflow results after executing the target process for overflow detection.

参数

  • status (object) – A status instance used to detect the overflow.

  • compute_output – Overflow detection should be performed on a certain computation. Set compute_output as the output of the computation, to ensure overflow status is acquired before executing the computation.

返回

bool, whether the overflow occurs or not.

process_loss_scale(overflow)[源代码]

Calculate loss scale according to the overflow.

参数

overflow (bool) – Whether the overflow occurs or not.

返回

bool, overflow value.

set_sense_scale(sens)[源代码]

If the user has set the sens in the training process and wants to reassign the value, he can call this function again to make modification, and sens needs to be of type Tensor.

start_overflow_check(pre_cond, compute_input)[源代码]

Start floating-point overflow detection. Create and clear the overflow detection state.

Specify the argument ‘pre_cond’ and ‘compute_input’ to make sure overflow status is cleared at the right time. Taking this situation as an example, we need to execute state clearing after loss calculation and then detect overflow in the process of gradient calculation. In this case, pre_cond should be the output of the loss function, and compute_input should be the input of gradients-computing function.

参数

  • pre_cond (object) – A precondition for starting overflow detection. It determines the executing order of overflow state clearing and prior processions. It makes sure that the function ‘start_overflow’ clears status after finishing the process of precondition.

  • compute_input (object) – The input of subsequent process. Overflow detection should be performed on a certain computation. Set compute_input as the input of the computation, to ensure overflow status is cleared before executing the computation.

返回

Tuple[object, object], the first value is False for GPU backend, while it is a instance of NPUAllocFloatStatus for other backend. The status is used to detect overflow during overflow detection. The second value is the same as the input of compute_input, but contains some information about the execution order.

class tinyms.layers.DistributedGradReducer(parameters, mean=True, degree=None, fusion_type=1)[源代码]

A distributed optimizer.

Constructs a gradient reducer Cell, which applies communication and average operations on single-process gradient values.

参数

  • parameters (list) – the parameters to be updated.

  • mean (bool) – When mean is true, the mean coefficient (degree) would apply on gradients. Default: False.

  • degree (int) – The mean coefficient. Usually it equals to device number. Default: None.

  • fusion_type (int) – The type of all reduce fusion. Default: 1.

引发

ValueError – If degree is not a int or less than 0.

Supported Platforms:

Ascend, GPU

实际案例

>>> # This example should be run with multiple processes.
>>> # Please refer to the tutorial > Distributed Training on mindspore.cn.
>>> import numpy as np
>>> from mindspore.communication import init
>>> from mindspore.ops import composite as C
>>> from mindspore.ops import operations as P
>>> from mindspore.ops import functional as F
>>> from mindspore import context
>>> from mindspore.context import ParallelMode
>>> from mindspore import Parameter, Tensor
>>> from mindspore import nn
>>> from mindspore.nn.wrap.cell_wrapper import WithLossCell
>>> from mindspore.parallel._utils import (_get_device_num, _get_gradients_mean)
>>>
>>> context.set_context(mode=context.GRAPH_MODE)
>>> init()
>>> context.reset_auto_parallel_context()
>>> context.set_auto_parallel_context(parallel_mode=ParallelMode.DATA_PARALLEL)
>>>
>>> class TrainingWrapper(nn.Cell):
...     def __init__(self, network, optimizer, sens=1.0):
...         super(TrainingWrapper, self).__init__(auto_prefix=False)
...         self.network = network
...         self.network.add_flags(defer_inline=True)
...         self.weights = optimizer.parameters
...         self.optimizer = optimizer
...         self.grad = C.GradOperation(get_by_list=True, sens_param=True)
...         self.sens = sens
...         self.reducer_flag = False
...         self.grad_reducer = None
...         self.parallel_mode = context.get_auto_parallel_context("parallel_mode")
...         if self.parallel_mode in [ParallelMode.DATA_PARALLEL, ParallelMode.HYBRID_PARALLEL]:
...             self.reducer_flag = True
...         if self.reducer_flag:
...             mean = _get_gradients_mean()
...             degree = _get_device_num()
...             self.grad_reducer = nn.DistributedGradReducer(optimizer.parameters, mean, degree)
...
...     def construct(self, *args):
...         weights = self.weights
...         loss = self.network(*args)
...         sens = P.Fill()(P.DType()(loss), P.Shape()(loss), self.sens)
...         grads = self.grad(self.network, weights)(*args, sens)
...         if self.reducer_flag:
...             # apply grad reducer on grads
...             grads = self.grad_reducer(grads)
...         return F.depend(loss, self.optimizer(grads))
>>>
>>> class Net(nn.Cell):
...     def __init__(self, in_features, out_features):
...         super(Net, self).__init__()
...         self.weight = Parameter(Tensor(np.ones([in_features, out_features]).astype(np.float32)),
...                                 name='weight')
...         self.matmul = P.MatMul()
...
...     def construct(self, x):
...         output = self.matmul(x, self.weight)
...         return output
>>>
>>> size, in_features, out_features = 16, 16, 10
>>> network = Net(in_features, out_features)
>>> loss = nn.MSELoss()
>>> net_with_loss = WithLossCell(network, loss)
>>> optimizer = nn.Momentum(net_with_loss.trainable_params(), learning_rate=0.1, momentum=0.9)
>>> train_cell = TrainingWrapper(net_with_loss, optimizer)
>>> inputs = Tensor(np.ones([size, in_features]).astype(np.float32))
>>> label = Tensor(np.zeros([size, out_features]).astype(np.float32))
>>> grads = train_cell(inputs, label)
>>> print(grads)
256.0
construct(grads)[源代码]

Under certain circumstances, the data precision of grads could be mixed with float16 and float32. Thus, the result of AllReduce is unreliable. To solve the problem, grads must be cast to float32 before AllReduce, and cast back after the operation.

参数

grads (Union[Tensor, tuple[Tensor]]) – The gradient tensor or tuple before operation.

返回

new_grads (Union[Tensor, tuple[Tensor]]), the gradient tensor or tuple after operation.

class tinyms.layers.ParameterUpdate(param)[源代码]

Cell that updates parameters.

With this Cell, one can manually update param with the input Tensor.

参数

param (Parameter) – The parameter to be updated manually.

引发

KeyError – If parameter with the specified name does not exist.

Supported Platforms:

Ascend

实际案例

>>> network = nn.Dense(3, 4)
>>> param = network.parameters_dict()['weight']
>>> update = nn.ParameterUpdate(param)
>>> update.phase = "update_param"
>>> weight = Tensor(np.arange(12).reshape((4, 3)), mindspore.float32)
>>> output = update(weight)
class tinyms.layers.DynamicLossScaleUpdateCell(loss_scale_value, scale_factor, scale_window)[源代码]

Dynamic Loss scale update cell.

For loss scaling training, the initial loss scaling value will be set to be loss_scale_value. In each training step, the loss scaling value will be updated by loss scaling value/scale_factor when there is an overflow. And it will be increased by loss scaling value * scale_factor if there is no overflow for a continuous scale_window steps. This cell is used for Graph mode training in which all logic will be executed on device side(Another training mode is normal(non-sink) mode in which some logic will be executed on host).

参数

  • loss_scale_value (float) – Initializes loss scale.

  • scale_factor (int) – Coefficient of increase and decrease.

  • scale_window (int) – Maximum continuous training steps that do not have overflow.

Inputs:

  • inputs (Tensor) - Tensor of shape \((N, \ldots)\).

  • label (Tensor) - Tensor of shape \((N, \ldots)\).

Outputs:

Tensor, a scalar Tensor with shape \(()\).

引发

TypeError – If dtype of inputs or label is neither float16 nor float32.

Supported Platforms:

Ascend GPU

实际案例

>>> import numpy as np
>>> from mindspore import Tensor, Parameter, nn
>>> from mindspore.ops import operations as P
>>> from mindspore.nn.wrap.cell_wrapper import WithLossCell
>>>
>>> class Net(nn.Cell):
...     def __init__(self, in_features, out_features):
...         super(Net, self).__init__()
...         self.weight = Parameter(Tensor(np.ones([in_features, out_features]).astype(np.float32)),
...                                 name='weight')
...         self.matmul = P.MatMul()
...
...     def construct(self, x):
...         output = self.matmul(x, self.weight)
...         return output
...
>>> in_features, out_features = 16, 10
>>> net = Net(in_features, out_features)
>>> loss = nn.MSELoss()
>>> optimizer = nn.Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
>>> net_with_loss = WithLossCell(net, loss)
>>> manager = nn.DynamicLossScaleUpdateCell(loss_scale_value=2**12, scale_factor=2, scale_window=1000)
>>> train_network = nn.TrainOneStepWithLossScaleCell(net_with_loss, optimizer, scale_sense=manager)
>>> input = Tensor(np.ones([out_features, in_features]), mindspore.float32)
>>> labels = Tensor(np.ones([out_features,]), mindspore.float32)
>>> output = train_network(input, labels)
class tinyms.layers.FixedLossScaleUpdateCell(loss_scale_value)[源代码]

Static scale update cell, the loss scaling value will not be updated.

For usage, refer to DynamicLossScaleUpdateCell.

参数

loss_scale_value (float) – Initializes loss scale.

Supported Platforms:

Ascend GPU

实际案例

>>> import numpy as np
>>> from mindspore import Tensor, Parameter, nn
>>> from mindspore.ops import operations as P
>>> from mindspore.nn.wrap.cell_wrapper import WithLossCell
>>>
>>> class Net(nn.Cell):
...     def __init__(self, in_features, out_features):
...         super(Net, self).__init__()
...         self.weight = Parameter(Tensor(np.ones([in_features, out_features]).astype(np.float32)),
...                                 name='weight')
...         self.matmul = P.MatMul()
...
...     def construct(self, x):
...         output = self.matmul(x, self.weight)
...         return output
...
>>> in_features, out_features = 16, 10
>>> net = Net(in_features, out_features)
>>> loss = nn.MSELoss()
>>> optimizer = nn.Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
>>> net_with_loss = WithLossCell(net, loss)
>>> manager = nn.FixedLossScaleUpdateCell(loss_scale_value=2**12)
>>> train_network = nn.TrainOneStepWithLossScaleCell(net_with_loss, optimizer, scale_sense=manager)
>>> input = Tensor(np.ones([out_features, in_features]), mindspore.float32)
>>> labels = Tensor(np.ones([out_features,]), mindspore.float32)
>>> output = train_network(input, labels)
class tinyms.layers.VirtualDatasetCellTriple(backbone)[源代码]

Wrap the network with virtual dataset to convert data parallel layout to model parallel layout.

VirtualDatasetCellTriple is a virtual Primitive, it does not exist in the final executing graph. Inputs and outputs of VirtualDatasetCellTriple are distributed in data parallel pattern, tensor redistribution Primitives is inserted dynamically during the graph compile process.

注解

Only used in semi auto parallel and auto parallel mode. There are three inputs, as contrary to two inputs in _VirtualDatasetCell.

参数

backbone (Cell) – The target network to wrap.

实际案例

>>> net = Net()
>>> net = VirtualDatasetCellTriple(net)