mindspore.ops
For more information about dynamic shape support status, please refer to Dynamic Shape Support Status of ops Interface .
Compared with the previous version, the added, deleted and supported platforms change information of mindspore.ops operators in MindSpore, please refer to the link mindspore.ops API Interface Change.
Neural Network Layer Functions
Neural Network
API Name 
Description 
Supported Platforms 
Warning 
Applies a 1D adaptive average pooling over an input Tensor which can be regarded as a composition of 1D input planes. 

None 

Performs 2D adaptive average pooling on a multiplane input signal. 

This is an experimental API that is subject to change or deletion. 

Performs 3D adaptive average pooling on a multiplane input signal. 

None 

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

None 

This operator applies a 2D adaptive max pooling to an input signal composed of multiple input planes. 

None 

Implements the add_layer_norm algorithm. 

This is an experimental API that is subject to change or deletion. 

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

kernel_size is in the range [1, 255]. stride is in the range [1, 63]. 

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

kernel_size is in the range [1, 255]. stride is in the range [1, 63]. 

Applies a 3D average pooling over an input Tensor which can be regarded as a composition of 3D input planes. 

kernel_size is in the range [1, 255]. stride is in the range [1, 63]. 

Batch Normalization for input data and updated parameters. 

For Atlas 200/300/500 inference product, the result accuracy fails to reach 1‰ due to the square root instruction. 

Returns the sum of the input_x and the bias Tensor. 

None 

Applies bilinear dense connected layer for input1 and input2. 

This is an experimental API that is subject to change or deletion. 

Performs greedy decoding on the logits given in inputs. 

None 

Applies a 1D convolution over an input tensor. 

None 

Applies a 2D convolution over an input tensor. 

None 

Applies a 3D convolution over an input tensor. 

None 

Given 4D tensor inputs x, weight and offsets, compute a 2D deformable convolution. 

This is an experimental API that is subject to change or deletion. 

Applies the dense connected operation to the input. 

This is an experimental API that is subject to change or deletion. 

During training, randomly zeroes some of the elements of the input tensor with probability p from a Bernoulli distribution. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

During training, randomly zeroes some channels of the input tensor with probability p from a Bernoulli distribution(For a 3dimensional tensor with a shape of \(NCL\), the channel feature map refers to a 1dimensional feature map with the shape of \(L\)). 

None 

During training, randomly zeroes some channels of the input tensor with probability p from a Bernoulli distribution(For a 4dimensional tensor with a shape of \(NCHW\), the channel feature map refers to a 2dimensional feature map with the shape of \(HW\)). 

None 

During training, randomly zeroes some channels of the input tensor with probability p from a Bernoulli distribution(For a 5dimensional tensor with a shape of \(NCDHW\), the channel feature map refers to a 3dimensional feature map with a shape of \(DHW\)). 

None 

Retrieve the word embeddings in weight using indices specified in input. 

On Ascend, the behavior is unpredictable when the value of input is invalid. 

Flatten a tensor along dimensions from start_dim to start_dim. 

None 

Combines an array of sliding local blocks into a large containing tensor. 

The input must be a 3dimensional Tensor with shape \((N, C \times \prod(\text{kernel_size}), L)\) . The output must be a 4dimensional Tensor with shape \((N, C, output\_size[0], output\_size[1], ...)\) . 

Applies the 3D FractionalMaxPool operation over input. 

This is an experimental API that is subject to change or deletion. 

Group Normalization over a minibatch of inputs. 

None 

Applies the Layer Normalization on the minibatch input. 

None 

Applying 1D LPPooling operation on an input Tensor can be regarded as forming a 1D input plane. 

None 

Applying 2D LPPooling operation on an input Tensor can be regarded as forming a 2D input plane. 

None 

Local Response Normalization. 

lrn is deprecated on Ascend due to potential accuracy problem. It's recommended to use other normalization methods, e.g. 

Performs a 2D max pooling on the input Tensor. 

None 

Performs a 3D max pooling on the input Tensor. 

None 

Computes the inverse of max_pool1d. 

None 

Computes the inverse of max_pool2d. 

None 

Computes the inverse of 

None 

The RmsNorm(Root Mean Square Layer Normalization) operator is a normalization operation. 

This is an experimental API that is subject to change or deletion. This API is only supported in Atlas A2 training series for now. 

Implements the Rotary Position Embedding algorithm. 

This is an experimental API that is subject to change or deletion. 

Extracts sliding local blocks from a batched input tensor. 

The output is a 3dimensional Tensor whose shape is \((N, C \times \prod(\text{kernel_size}), L)\) . This is an experimental API that is subject to change or deletion. 
Loss Functions
API Name 
Description 
Supported Platforms 
Warning 
Computes the binary cross entropy(Measure the difference information between two probability distributions) between predictive value logits and target value labels. 

The value of logits must range from 0 to l. 

Adds sigmoid activation function to input input as logits, and uses the given logits to compute binary cross entropy between the input and the target. 

None 

CosineEmbeddingLoss creates a criterion to measure the similarity between two tensors using cosine distance. 

None 

The cross entropy loss between input and target. 

None 

Calculates the CTC (Connectionist Temporal Classification) loss and the gradient. 

None 

Gaussian negative log likelihood loss. 

None 

Measures Hinge Embedding Loss given an input Tensor intputs and a labels Tensor targets (containing 1 or 1). 

None 

Calculates the error between the predicted value and the target value, which has the best of both the loss of 

None 

Computes the KullbackLeibler divergence between the logits and the labels. 

None 

Calculate the mean absolute error between the input value and the target value. 

None 

MarginRankingLoss creates a criterion that measures the loss. 

None 

Calculates the mean squared error between the predicted value and the label value. 

None 

Hinge loss for optimizing a multiclass classification. 

None 

Hinge loss for optimizing a multilabel classification. 

None 

Calculates the MultiLabelSoftMarginLoss. 

None 

Gets the negative log likelihood loss between inputs and target. 

None 

Computes smooth L1 loss, a robust L1 loss. 

None 

Calculate the soft margin loss of input and target. 

This is an experimental API that is subject to change or deletion. 

TripletMarginLoss operation. 

None 
Activation Functions
API Name 
Description 
Supported Platforms 
Warning 
celu activation function, computes celu (Continuously differentiable exponential linear units) of input tensors elementwise. 

This is an experimental API that is subject to change or deletion. 

Exponential Linear Unit activation function. 

None 

Fast Gaussian Error Linear Units activation function. 

None 

Gaussian Error Linear Units activation function. 

None 

Computes GLU (Gated Linear Unit activation function) of input tensors. 

None 

Returns the samples from the GumbelSoftmax distribution and optionally discretizes. 

None 

Hard Shrink activation function. 

None 

Hard Sigmoid activation function. 

None 

Hard Swish activation function. 

None 

Applies the hardtanh activation function elementwise. 

None 

leaky_relu activation function. 

None 

Applies the Log Softmax function to the input tensor on the specified axis. 

None 

Applies logsigmoid activation elementwise. 

None 

Computes MISH(A Self Regularized NonMonotonic Neural Activation Function) of input tensors elementwise. 

None 

Parametric Rectified Linear Unit activation function. 

None 

Computes ReLU (Rectified Linear Unit activation function) of input tensors elementwise. 

None 

Computes ReLU (Rectified Linear Unit) upper bounded by 6 of input tensors elementwise. 

None 

Randomized Leaky ReLU activation function. 

None 

Activation function SeLU (Scaled exponential Linear Unit). 

None 

Computes Sigmoid of input elementwise. 

None 

Computes Sigmoid Linear Unit of input elementwise. 

None 

Applies the Softmax operation to the input tensor on the specified axis. 

None 

Applies the Softmin operation to the input tensor on the specified axis. 

None 

Soft Shrink activation function. 

None 

SoftSign activation function. 

None 

Computes hyperbolic tangent of input elementwise. 

None 

Returns each element of input after thresholding by thr as a Tensor. 

None 
Distance Functions
API Name 
Description 
Supported Platforms 
Warning 
Computes pnorm distance between each pair of row vectors of two input Tensors. 

None 

Computes batched the \(p\)norm distance between each pair of the two collections of row vectors. 

None 

Calculates the distance between every pair of row vectors in the input using the pnorm. 

None 
Sampling Functions
API Name 
Description 
Supported Platforms 
Warning 
Generates a random sample as index tensor with a mask tensor from a given tensor. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Generates random samples from a given categorical distribution tensor. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Generates random labels with a loguniform distribution for sampled_candidates. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Uniform candidate sampler. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. The Ascend backend does not support dynamic shape scenarios currently. 
Image Functions
API Name 
Description 
Supported Platforms 
Warning 
Returns a 2D or 3D flow field (sampling grid) based on theta, a batch of affine matrices. 

None 

Decode the bounding box locations, calculate the offset, and convert the offset into a Bbox, which is used to mark the target in the subsequent images, etc. 

None 

Encode the bounding box locations, calculate the offset between the predicted bounding boxes and the real bounding boxes, and the offset will be used as a variable for the loss. 

None 

Combines an array of sliding local blocks into a large containing tensor. 

None 

Checks whether the bounding box is in the image. 

The bounding box specified by bboxes and the image information specified by img_metas need to be valid, i.e.: \(x0 <= x1\) , \(y0 <= y1\) , and \((height, width, ratio)\) are all positive. 

Extracts crops from the input image Tensor and resizes them. 

None 

Given an input and a flowfield grid, computes the output using input values and pixel locations from grid. 

None 

Samples the input Tensor to the given size or scale_factor by using one of the interpolate algorithms. 

None 

Calculates intersection over union for boxes. 

In Ascend, only computation of float16 data is supported. To avoid overflow, the input length and width are scaled by 0.2 internally. 

Pads the input tensor according to the padding. 

None 

Extends the last dimension of the input tensor from 1 to pad_dim_size, by filling with 0. 

None 

Applies the PixelShuffle operation over input input which implements subpixel convolutions with stride \(1/r\) . 

None 

Applies the PixelUnshuffle operation over input input which is the inverse of PixelShuffle. 

None 

Alias for 

None 
Mathematical Functions
Elementwise Operations
API Name 
Description 
Supported Platforms 
Warning 
Returns absolute value of a tensor elementwise. 

None 

Alias for 

None 

Computes accumulation of all input tensors elementwise. 

None 

Computes arccosine of input tensors elementwise. 

None 

Alias for 

None 

Computes inverse hyperbolic cosine of the inputs elementwise. 

None 

Adds other value to input Tensor. 

None 

Performs the elementwise division of tensor tensor1 by tensor tensor2, multiply the result by the scalar value and add it to input data. 

None 

Performs the elementwise product of tensor tensor1 and tensor tensor2, multiply the result by the scalar value and add it to input data. 

None 

Multiplies matrix mat and vector vec. 

None 

Computes addition of all input tensors elementwise. 

None 

Returns the elementwise argument of a complex tensor. 
Ascend`` 
None 

Alias for 

None 

Alias for 

None 

Alias for 

None 

Alias for 

None 

Alias for 

None 

Alias for 

None 

Computes arcsine of input tensors elementwise. 

None 

Computes inverse hyperbolic sine of the input elementwise. 

None 

Computes the trigonometric inverse tangent of the input elementwise. 

None 

Returns arctangent of input/other elementwise. 

None 

Computes inverse hyperbolic tangent of the input elementwise. 

None 

Reshapes Tensor in inputs, every Tensor has at least one dimension after this operation. 

None 

Reshapes Tensor in inputs, every Tensor has at least 2 dimension after this operation. 

None 

Reshapes Tensor in inputs, every Tensor has at least 3 dimension after this operation. 

None 

Computes modified Bessel function of the first kind, order 0 elementwise. 

None 

Computes exponential scaled modified Bessel function of the first kind, order 0 elementwise. 

None 

Computes modified Bessel function of the first kind, order 1 elementwise. 

None 

Computes exponential scaled modified Bessel function of the first kind, order 1 elementwise. 

None 

Computes Bessel function of the first kind, order 0 elementwise. 

None 

Computes Bessel function of the first kind, order 1 elementwise. 

None 

Computes modified Bessel function of the second kind, order 0 elementwise. 

None 

Computes exponential scaled modified Bessel function of the second kind, order 0 elementwise. 

None 

Computes modified Bessel function of the second kind, order 1 elementwise. 

None 

Computes exponential scaled modified Bessel function of the second kind, order 1 elementwise. 

None 

Computes Bessel function of the second kind, order 0 elementwise. 

None 

Computes Bessel function of the second kind, order 1 elementwise. 

None 

Returns bitwise and of two tensors elementwise. 

None 

Perform a left bitwise shift operation on the input elementwise, where the number of bits to shift is specified by other. 

None 

Returns bitwise or of two tensors elementwise. 

None 

Perform a right bitwise shift operation on the input elementwise, where the number of bits to shift is specified by other. 

None 

Returns bitwise xor of two tensors elementwise. 

None 

Rounds a tensor up to the closest integer elementwise. 

None 

Clamps tensor values between the specified minimum value and maximum value. 

None 

Alias for 

None 

Returns all rlength subsequences of input Tensor. 

None 

Create a new floatingpoint tensor with the magnitude of x and the sign of other, elementwise. 

None 

Computes cosine of input elementwise. 

Using float64 may cause a problem of missing precision. 

Computes hyperbolic cosine of input elementwise. 

None 

Calculate cosine similarity between x1 and x2 along the axis, dim. 

None 

Given the input and weights, returns the covariance matrix (the square matrix of the covariance of each pair of variables) of input, where the input row is the variable and the column is the observation value. 

The values of fweights and aweights cannot be negative, and the negative weight scene result is undefined. 

Creates a tensor with diagonals filled by input. 

None 

Computes the nth discrete difference along a specified axis of a given input x. 

None 

Converts angles in degrees to angles in radians elementwise. 

None 

Computes the derivative of the lgamma function on input. 

This is an experimental API that is subject to change or deletion. 

Divides the first input tensor by the second input tensor in floatingpoint type elementwise. 

None 

Alias for 

None 

Computes the Gauss error function of input elementwise. 

None 

Computes the complementary error function of input elementwise. 

None 

Returns the result of the inverse error function with input. 

None 

Returns exponential of a tensor elementwise. 

None 

Computes base two exponential of Tensor input elementwise. 

None 

Returns exponential then minus 1 of a tensor elementwise. 
Ascend`` 
None 

Rounds a tensor down to the closest integer elementwise. 

None 

Alias for 

None 

Divides the first input tensor by the second input tensor elementwise and round down to the closest integer. 

This is an experimental API that is subject to change or deletion. 

Computes the remainder of division elementwise. 

Data of input y should not be 0, or the maximum value of its dtype will be returned. When the elements of input exceeds 2048 , the accuracy of operator cannot guarantee the requirement of double thousandths in the mini form. Due to different architectures, the calculation results of this operator on NPU and CPU may be inconsistent. If shape is expressed as \((D1, D2 ..., Dn)\), then D1*D2... *DN<=1000000,n<=8. 

Computes input to the power of the exponent. 

None 

Computes the floatingpoint remainder of the division operation input/other. 

None 

Calculates the fractional part of each element in the input 

None 

Computes greatest common divisor of input tensors elementwise. 

This is an experimental API that is subject to change or deletion. 

Computes hypotenuse of input tensors elementwise as legs of a right triangle. 

None 

Calculates lower regularized incomplete Gamma function. 

This is an experimental API that is subject to change or deletion. 

Calculates upper regularized incomplete Gamma function. 

This is an experimental API that is subject to change or deletion. 

Returns a new tensor containing imaginary value of the input. 

None 

Alias for 

None 

Computes Reciprocal of input tensor elementwise. 

None 

Flips all bits of input tensor elementwise. 

None 

Computes least common multiplier of input tensors elementwise. 

None 

Multiplies input Tensor by \(2^{other}\) elementwise. 

None 

Does a linear interpolation of two tensors input and end based on a float or tensor weight. 

None 

Returns the natural logarithm of a tensor elementwise. 

If the input value of operator Log is within the range (0, 0.01] or [0.95, 1.05], the output accuracy may be affacted. 

Returns a new Tensor by taking the base 2 logarithm of the elements in the input Tensor. 

If the input value of operator log2 is within the range (0, 0.01] or [0.95, 1.05], the output accuracy may be affected. 

Returns a new Tensor by taking the base 10 logarithm of the elements in the input Tensor. 

If the input value of operator log10 is within the range (0, 0.01] or [0.95, 1.05], the output accuracy may be affected. 

Returns the natural logarithm of one plus the input tensor elementwise. 

None 

Computes the logarithm of the sum of exponentiations of the inputs. 

None 

Computes the logarithm of the sum of exponentiations in base of 2 of the inputs. 

None 

Computes the "logical AND" of two tensors elementwise. 

None 

Computes the "logical NOT" of a tensor elementwise. 

None 

Computes the "logical OR" of two tensors elementwise. 

None 

Computes the "logical XOR" of two tensors elementwise. 

None 

Calculate the logit of a tensor elementwise. 

None 

Multiplies two tensors elementwise. 

None 

Alias for 

None 

Returns the results of the multivariate loggamma function with dimension p elementwise. 

None 

Returns a tensor with negative values of the input tensor elementwise. 

None 

Alias for 

None 

Returns the next representable floatingpoint value after input towards other elementwise. 

None 

Converts polar coordinates to Cartesian coordinates. 

None 

Computes the \(n\)th derivative of the polygamma function on input. 

None 

Return self Tensor. 

None 

Calculates the exponent power of each element in input. 

None 

Converts angles in radians to angles in degrees elementwise. 

None 

Expand the multidimensional Tensor into 1D along the 0 axis direction. 

None 

Returns a Tensor that is the real part of the input. 

None 

Returns reciprocal of a tensor elementwise. 

None 

Computes the remainder of dividing the first input tensor by the second input tensor elementwise. 

When the elements of input exceed 2048, there might be accuracy problems. The calculation results of this operator on Ascend and CPU might be inconsistent. If shape is expressed as (D1,D2... ,Dn), then D1*D2... *DN<=1000000,n<=8. 

Rotate a nD tensor by 90 degrees in the plane specified by dims axis. 

None 

Returns half to even of a tensor elementwise. 

None 

Computes reciprocal of square root of input tensor elementwise. 

None 

Extension of 

None 

Returns an elementwise indication of the sign of a number. 

None 

Determine the symbol of each element. 

None 

Computes sine of the input elementwise. 

None 

Computes the normalized sinc of input. 

None 

Computes hyperbolic sine of the input elementwise. 

None 

Returns sqrt of a tensor elementwise. 

None 

Returns square of a tensor elementwise. 

None 

Subtracts the second input tensor from the first input tensor elementwise. 

None 

Performs the elementwise subtract of input tensors. 

None 

Transposes a 2D Tensor. 

None 

Computes tangent of input elementwise. 

None 

Tanhshrink Activation, \(Tanhshrink(x)=xTanh(x)\) , where \(x\) corresponds to input . 

None 

Integrates y(x) along given dim using trapezoidal rule. 

None 

Calculates the indices of the lower triangular elements in a row * col matrix and returns them as a 2byN Tensor. 

None 

Calculates the indices of the upper triangular elements in a row * col matrix and returns them as a 2byN Tensor. 

None 

Alias for 

None 

Returns a new tensor with the truncated integer values of the elements of the input tensor. 

None 

Divides the first input tensor by the second input tensor elementwise and rounds the results of division towards zero. 

None 

Returns the remainder of division elementwise. 

The input data does not support 0. When the elements of input exceed 2048 , the accuracy of operator cannot guarantee the requirement of double thousandths in the mini form. Due to different architectures, the calculation results of this operator on NPU and CPU may be inconsistent. If shape is expressed as (D1,D2... ,Dn), then D1*D2... *DN<=1000000,n<=8. 

Divides the first input tensor by the second input tensor elementwise. 

None 

Computes the first input tensor multiplied by the logarithm of second input tensor elementwise. 

On Ascend, the data type of input and other must be float16 or float32. 

Elementalwise compute the Hurwitz zeta output. 

This is an experimental API that is subject to change or deletion. 
Reduction Functions
API Name 
Description 
Supported Platforms 
Warning 
Reduces a dimension of input by the "logical AND" of all elements in the dimension, by default. 

None 

Reduces all dimensions of a tensor by returning the maximum value in input, by default. 

None 

Reduces all dimensions of a tensor by returning the minimum value in input, by default. 

None 

It returns the minimum and maximum value along the given axis of input tensor. 

None 

Reduces a dimension of input by the "logical OR" of all elements in the dimension, by default. 

None 

Return the indices of the maximum values of a tensor across a dimension. 

None 

Returns the indices of the minimum value of a tensor across the axis. 

None 

Returns a tuple (values,indices) where 'values' is the cumulative maximum value of input Tensor input along the dimension axis, and indices is the index location of each maximum value. 

None 

Returns a tuple (values,indices) where 'values' is the cumulative minimum value of input Tensor input along the dimension axis, and indices is the index location of each minimum value. 

None 

Computes the cumulative product of the input tensor along dimension dim. 

None 

Computes the cumulative sum of input Tensor along axis. 

None 

Computes the maximum of input tensors elementwise. 

None 

Computes the histogram of a tensor. 

None 

Compute the cumulative logsumexp of the input tensor input along axis . 

This is an experimental API that is subject to change or deletion. 

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

None 

Calculates the maximum value along with the given axis for the input tensor. 

If there are multiple maximum values, the index of the first maximum value is used. The value range of "axis" is [dims, dims  1]. "dims" is the dimension length of "input". 

Reduces all dimension of a tensor by averaging all elements in the dimension, by default. 

None 

Computes the median and indices of input tensor. 

indices does not necessarily contain the first occurrence of each median value found in the input, unless it is unique. The specific implementation of this API is devicespecific. The results may be different on CPU and GPU. 

Calculates the minimum value along with the given axis for the input tensor. 

If there are multiple minimum values, the index of the first minimum value is used. The value range of "axis" is [dims, dims  1]. "dims" is the dimension length of "x". 

Returns the matrix norm or vector norm of a given tensor. 

None 

Reduces a dimension of a tensor by multiplying all elements in the dimension, by default. 

None 

Returns the standarddeviation of each row of the input Tensor by default, or it can calculate them in specified dimension axis. 

None 

Returns the standarddeviation and mean of each row of the input Tensor by default, or it can calculate them in specified dimension axis. 

None 

Returns the variance of each row of the input Tensor by default, or it can calculate them in specified dimension axis. 

None 

Returns the variance and mean of each row of the input Tensor by default, or it can calculate them in specified dimension axis. 

None 
Comparison Functions
API Name 
Description 
Supported Platforms 
Warning 
Sorts the input tensor along the given dimension in specified order and return the sorted indices. 

None 

Returns 

None 

Bucketizes input based on boundaries. 

None 

Computes the equivalence between two tensors elementwise. 

None 

Computes the equivalence between two tensors elementwise. 

None 

Computes the boolean value of \(input >= other\) elementwise. 

None 

Compare the value of the input parameters \(input > other\) elementwise, and the output result is a bool value. 

None 

Given two Tensors, compares them elementwise to check if each element in the first Tensor is greater than or equal to the corresponding element in the second Tensor. 

None 

Compare the value of the input parameters \(input,other\) elementwise, and the output result is a bool value. 

None 

Determines whether the targets are in the top k predictions. 

None 

Returns a new Tensor with boolean elements representing if each element of input is “close” to the corresponding element of other. 

None 

Determine which elements are finite for each position. 

None 

Determines which elements are inf or inf for each position. 

None 

Determines which elements are NaN for each position. 

None 

Tests elementwise for negative infinity. 

None 

Tests elementwise for positive infinity. 

None 

Tests elementwise for real number. 

None 

Return True if the data type of the tensor is complex, otherwise return False. 

None 

Judge whether the data type of input is a floating point data type i.e., one of mindspore.float64, mindspore.float32, mindspore.float16. 

None 

Computes the boolean value of \(input <= other\) elementwise. 

None 

Computes the boolean value of \(input < other\) elementwise. 

None 

Computes the boolean value of \(input <= other\) elementwise. 

None 

Alias for 

None 

Computes the maximum of input tensors elementwise. 

If all inputs are scalar of integers. In GRAPH mode, the output will be Tensor of int32, while in PYNATIVE mode, the output will be Tensor of int64. 

Computes the minimum of input tensors elementwise. 

None 

Sorts the elements in Tensor in ascending order of value along its first dimension. 

None 

Computes the nonequivalence of two tensors elementwise. 

None 

Alias for 

None 

Return the position indices such that after inserting the values into the sorted_sequence, the order of innermost dimension of the sorted_sequence remains unchanged. 

None 

Finds values and indices of the k largest or smallest entries along a given dimension. 

If sorted is set to False, it will use the aicpu operator, the performance may be reduced. In addition, due to different memory layout and traversal methods on different platforms, the display order of calculation results may be inconsistent when sorted is False. 
Linear Algebraic Functions
API Name 
Description 
Supported Platforms 
Warning 
Applies batch matrix multiplication to batch1 and batch2, with a reduced add step and add input to the result. 

None 

Multiplies matrix mat1 and matrix mat2. 

None 

Computes the outer product of two vector vec1 and vec2, and adds the resulting matrix to x. 

None 

Calculates the conjugation of Tensor element by element, and transposes the last two dimensions. 

None 

The result is the sum of the input and a batch matrixmatrix product of matrices in batch1 and batch2. 

None 

Computation of batch dot product between samples in two tensors containing batch dims, i.e. x1 or x2 's first dimension is batch size. 

None 

Computes matrix multiplication between two tensors by batch. 

None 

Returns the Cholesky decomposition of zero or more batch dimensions consisting of symmetric positivedefinite matrices. 

None 

Computes the solution of a set of linear equations with a positive definite matrix, according to its Cholesky decomposition factor input2 . 

This is an experimental API that is subject to change or deletion. 

Returns the matrix norm or vector norm of a given tensor. 

None 

Computation a dot product between samples in two tensors. 

None 

Computes the eigenvalues of a square matrix(batch square matrices). 

This is an experimental API that is subject to change or deletion. 

Decomposes a matrix into the product of an orthogonal matrix Q and an upper triangular matrix R. 

This is an experimental API that is subject to change or deletion. 

Ger product of input and vec2. 

None 

Returns the inner product of two tensors. 

None 

Compute the inverse of the input matrix. 

None 

Computes the Kronecker product \(input ⊗ other\), denoted by ⊗, of input and other. 

None 

Calculates log determinant of one or a batch of square matrices. 

None 

Computes the solution y to the system of linear equations \(Ay = b\) , given LU decomposition \(A\) and column vector \(b\). 

This is an experimental API that is subject to change or deletion. 

Converts LU_data and LU_pivots back into P, L and U matrices, where P is a permutation matrix, L is a lower triangular matrix, and U is an upper triangular matrix. 

None 

Returns the matrix product of two tensors. 

None 

Copy a tensor setting everything outside a central band in each innermost matrix to zero. 

This is an experimental API that is subject to change or deletion. 

Solves systems of linear equations. 

On GPU, if the matrix is irreversible, an error may be reported or an unknown result may be returned. 

Copy a tensor setting everything outside a central band in each innermost matrix to zero. 

This is an experimental API that is subject to change or deletion. 

Returns a Tensor with the contents in x as k[0]th to k[1]th diagonals of a matrix, with everything else padded with padding_value. 

None 

Returns the diagonal part of input tensor. 

None 

Returns a batched matrix tensor with new batched diagonal values. 

None 

Returns the matrix product of two arrays. 

None 

Multiplies matrix mat and vector vec. 

None 

Return outer product of input and vec2. 

None 

Calculates two matrices multiplication of a product of a general matrix with Householder matrices. 

None 

Calculates the explicit representation of the orthogonal matrix \(Q\) returned by 

None 

Calculates two matrices multiplication of a product of a general matrix with Householder matrices. 

None 

Computes the (MoorePenrose) pseudoinverse of a matrix. 

None 

Computes the singular value decompositions of one or more matrices. 

None 

Computes the sign and the log of the absolute value of the determinant of one or more square matrices. 

None 

Returns a new tensor that is the sum of the input main trace. 

None 

Computation of Tensor contraction on arbitrary axes between tensors a and b. 

None 

Generates a Vandermonde matrix. 

None 

Calculates the dot product of two batches of vectors across the specified dimension. 

This is an experimental API that is subject to change or deletion. 
Spectral Functions
API Name 
Description 
Supported Platforms 
Warning 
Bartlett window function is a triangularshaped weighting function used for smoothing or frequency analysis of signals in digital signal processing. 

None 

Blackman window function, usually used to extract finite signal segment for FFT. 

None 

Returns the Hamming window. 

None 

Generates a Hann Window. 

None 

Generates a Kaiser window, which is also known as the KaiserBessel window. 

None 
Tensor Operation Functions
Tensor Creation
API Name 
Description 
Supported Platforms 
Warning 
Create a Tensor with the same data type and shape as input, and the element value is the minimum value that the corresponding data type can express. 

None 

Creates a tensor with ones on the diagonal and zeros in the rest. 

None 

Create a Tensor of the specified shape and fill it with the specified value. 

None 

Create a Tensor of the specified shape and fill it with the specified value. 

None 

Return a Tensor of the same shape as input and filled with fill_value. 

None 

Returns a Tensor whose value is steps evenly spaced in the interval start and end (including start and end), and the length of the output Tensor is steps. 

None 

Returns a 1D Tensor with size steps whose value is from \(base^{start}\) to \(base^{end}\), and use base as the base number. 

None 

Computes a onehot tensor. 

None 

Creates a tensor filled with value ones, whose shape and type are described by the first argument size and second argument dtype respectively. 

For argument shape, Tensor type input will be deprecated in the future version. 

Returns a Tensor with a value of 1 and its shape is the same as the input. 

None 

Creates a sequence of numbers that begins at start and extends by increments of step up to but not including end. 

None 

Creates a sequence of numbers that begins at start and extends by increments of step up to but not including end. 

None 

Creates a tensor filled with value zeros, whose shape and type are described by the first argument size and second argument dtype respectively. 

For argument size, Tensor type input will be deprecated in the future version. 

Creates a tensor filled with 0, with the same size as input, and the given dtype. 

None 

Computes the Heaviside step function for each element in input. 

None 
Randomly Generating Functions
API Name 
Description 
Supported Platforms 
Warning 
Randomly set the elements of output to 0 or 1 with the probability of p which follows the Bernoulli distribution. 

None 

Generates random numbers according to the Gamma random number distribution. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Generates random numbers according to the Laplace random number distribution. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Returns a tensor sampled from the multinomial probability distribution located in the corresponding row of the input tensor. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Returns a tensor where each row contains numsamples indices sampled from the multinomial distribution with replacement. 

None 

Returns a new tensor that fills numbers from the uniform distribution over an interval \([0, 1)\) based on the given shape and dtype. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Returns a new tensor that fills numbers from the uniform distribution over an interval \([0, 1)\) based on the given shape and dtype. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Returns a Tensor whose elements are random integers in the range of [ low , high ) . 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Returns a tensor with the same shape as Tensor input whose elements are random integers in the range of [ low , high ) . 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Returns a new Tensor with given shape and dtype, filled with a sample (or samples) from the standard normal distribution. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Returns a new Tensor with given shape and dtype, filled with a sample (or samples) from the standard normal distribution. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Outputs random values from the Gamma distribution(s) described by alpha. 

None 

Generates random number Tensor with shape shape according to a Poisson distribution with mean rate. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Generates random permutation of integers from 0 to n1. 

This is an experimental API that is subject to change or deletion. The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Generates random numbers according to the Laplace random number distribution (mean=0, lambda=1). 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Generates random numbers according to the standard Normal (or Gaussian) random number distribution. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Generates random numbers according to the Uniform random number distribution. 

None 
Array Operation
API Name 
Description 
Supported Platforms 
Warning 
Return a Tensor of the positions of all nonzero values. 

None 

Divides batch dimension with blocks and interleaves these blocks back into spatial dimensions. 

None 

Counts the number of occurrences of each value in input. 

None 

Creates a block diagonal matrix from the provided Tensor. 

None 

Broadcasts input tensor to a given shape. 

None 

Connect input tensors along with the given axis. 

None 

Divide the channels in a tensor of shape \((*, C, H, W)\) into \(g\) groups and rearrange them as \((*, \frac{C}{g}, g, H*W)\), while keeping the original tensor shapes. 

None 

Cut the input Tensor into chunks subtensors along the specified axis. 

None 

Stacks 1D tensors as columns into a 2D tensor. 

None 

Alias for 

None 

Returns a tensor of complex numbers that are the complex conjugate of each element in input. 

None 

Count number of nonzero elements across axis of input tensor. 

None 

Returns a deepcopy of input tensor. 

None 

Constructs a diagonal tensor with a given diagonal values. 

This is an experimental API that is subject to change or deletion. 

Create a 2D Tensor which diagonal is the flattened input . 

None 

Returns specified diagonals of input. 

None 

dim1 and dim2 specify the two dimensions of input, the elements in these two dimensions will be treated as elements of a matrix, and src is embedded on the diagonal of the matrix. 

None 

Returns the shape of the input tensor. 

None 

Splits a tensor into multiple subtensors along the 3rd axis. 

None 

Stacks tensors along the third axis. 

None 

According to the Einstein summation Convention (Einsum), the product of the input tensor elements is summed along the specified dimension. 

None 

Adds an additional dimension to input_x at the given axis, the dimension of input_x should be greater than or equal to 1. 

None 

Reverses the order of elements in a tensor along the given axis. 

None 

Flips the elements of each row in the left/right direction, while preserving the columns of the input tensor. 

None 

Flips the elements of each column in the up/down direction, while preserving the rows of the input tensor. 

None 

Returns the slice of the input tensor corresponding to the elements of input_indices on the specified axis. 

None 

Gathers elements along an axis specified by dim. 

None 

Gathers elements along an axis specified by dim. 

On Ascend, the behavior is unpredictable in the following cases: the value of index is not in the range [input.shape[dim], input.shape[dim]) in forward; the value of index is not in the range [0, input.shape[dim]) in backward. 

Gathers slices from a tensor by indices. 

None 

Stacks tensors in sequence horizontally. 

None 

Splits a tensor into multiple subtensors horizontally. 

None 

Adds tensor y to specified axis and indices of Parameter x. 

None 

Fills the elements under the axis dimension of the input Tensor x with the input value by selecting the indices in the order given in index. 

None 

Generates a new Tensor that accesses the values of input along the specified axis dimension using the indices specified in index. 

None 

Adds v into specified rows of x. 

None 

Adds Tensor updates to specified axis and indices of Tensor var elementwise. 

None 

Subtracts v into specified rows of x. 

None 

Updates specified values in x to v according to indices. 

This is an experimental API that is subject to change or deletion. 

Determine whether the input Tensor contains 0 or False. 

None 

Fills elements of Tensor with value where mask is True. 

None 

Returns a new 1D Tensor which indexes the input tensor according to the boolean mask. 

None 

Generates coordinate matrices from given coordinate tensors. 

None 

Returns a narrowed tensor from input tensor, and the dimension axis is input from start to start + length. 

None 

Alias for ops.movedim. 

None 

Moves axis of an array from source to destination. 

None 

Replace the NaN, positive infinity and negative infinity values in input with the specified values in nan, posinf and neginf respectively. 

For Ascend, it is only supported on Atlas A2 Training Series Products. This is an experimental API that is subject to change or deletion. 

Computes the mean of input in specified dimension, ignoring NaN. 

None 

Computes the median and indices of input in specified dimension, ignoring NaN. 

indices does not necessarily contain the first occurrence of each median value found in the input, unless it is unique. 

Computes sum of input over a given dimension, treating NaNs as zero. 

None 

Generates random numbers according to the Normal (or Gaussian) random number distribution. 

The Ascend backend does not support the reproducibility of random numbers, so the seed parameter has no effect. 

Return the positions of all nonzero values. 

None 

Returns a Scalar of type int that represents the total number of elements in the Tensor. 

None 

Permutes the dimensions of the input tensor according to input axis . 

None 

Computes elementwise population count(a.k.a bitsum, bitcount). 

None 

Returns the rank of a tensor. 

None 

Repeat elements of a tensor along an axis, like numpy.repeat . 

None 

Repeat elements of a tensor along an axis, like numpy.repeat. 

None 

Rearranges the input Tensor based on the given shape. 

None 



None 

Reverses variable length slices. 

None 

Rolls the elements of a tensor along an axis. 

None 

Alias for 

None 

Update the value in src to input according to the specified index. 

None 

Scatters a tensor into a new tensor depending on the specified indices. 

None 

The conditional tensor determines whether the corresponding element in the output must be selected from input (if True) or other (if False) based on the value of each element. 

None 

On the specified dimension axis of input , src is scattered into input on the specified index of input . 

None 

Returns a mask tensor representing the first N positions of each cell. 

None 

Returns the shape of the input tensor. 

None 

Randomly shuffles a Tensor along its first dimension. 

None 

Returns a Scalar of type int that represents the size of the input Tensor and the total number of elements in the Tensor. 

None 

Slices a tensor in the specified shape. 

None 

Slice the input Tensor in the specified dimension and overlay the slice results with the source Tensor. 

None 

Sorts the elements of the input tensor along the given dimension in the specified order. 

Currently, the data types of float16, uint8, int8, int16, int32, int64 are well supported. If use float32, it may cause loss of accuracy. 

Divides a tensor's spatial dimensions into blocks and combines the block sizes with the original batch. 

None 

Computes a Tensor such that \(output_i = \frac{\sum_j x_{indices[j]}}{N}\) where mean is over \(j\) such that \(segment\_ids[j] == i\) and \(N\) is the total number of values summed. 

None 

Splits the Tensor into chunks along the given axis. 

None 

Return the Tensor after deleting the dimension of size 1 in the specified axis. 

None 

Stacks a list of tensors in specified axis. 

None 

Extracts a strided slice of a Tensor based on begin/end index and strides. 

None 

Calculate sum of Tensor elements over a given dim. 

None 

Interchange two axes of a tensor. 

None 

Interchange two dims of a tensor. 

None 

Creates a new tensor by adding the values from the positions in input_x indicated by indices, with values from updates. 

None 

Creates a new tensor by dividing the values from the positions in input_x indicated by indices, with values from updates. 

None 

By comparing the value at the position indicated by indices in input_x with the value in the updates, the value at the index will eventually be equal to the largest one to create a new tensor. 

None 

By comparing the value at the position indicated by indices in input_x with the value in the updates, the value at the index will eventually be equal to the smallest one to create a new tensor. 

None 

Creates a new tensor by multiplying the values from the positions in input_x indicated by indices, with values from updates. 

None 

Creates a new tensor by subtracting the values from the positions in input_x indicated by indices, with values from updates. 

None 

Write all elements in updates to the index specified by indices in input_x according to the reduction operation specified by reduction. 

The order in which updates are applied is nondeterministic, meaning that if there are multiple index vectors in indices that correspond to the same position, the value of that position in the output will be nondeterministic. On Ascend, the reduction only support set to "none" for now. On Ascend, the data type of input_x must be float16 or float32. This is an experimental API that is subject to change or deletion. 

Splits a tensor into multiple subtensors along the given axis. 

None 

Creates a new tensor by replicating input dims times. 

None 

Returns the lower triangle part of input (elements that contain the diagonal and below), and set the other elements to zeros. 

None 

Returns the upper triangle part of input (elements that contain the diagonal and below), and set the other elements to zeros. 

This is an experimental API that is subject to change or deletion. 

Permutes the dimensions of the input tensor according to input permutation. 

None 

Removes a tensor dimension in specified axis. 

None 

Extracts sliding local blocks from a batched input tensor. 

The output is a 3dimensional Tensor whose shape is \((N, C \times \prod(\text{kernel_size}), L)\) . This is an experimental API that is subject to change or deletion. 

Returns the unique elements of input tensor and also return a tensor containing the index of each value of input tensor corresponding to the output unique tensor. 

This is an experimental API that is subject to change or deletion. 

Returns the elements that are unique in each consecutive group of equivalent elements in the input tensor. 

None 

Returns unique elements and relative indexes in 1D tensor, filled with padding num. 
Deprecated 


Computes the maximum along segments of a tensor. 

None 

Computes the minimum of a tensor along segments. 

None 

Computes the product of a tensor along segments. 

None 

Computes the sum of a tensor along segments. 

None 

Adds an additional dimension to input at the given dim. 

None 

Unstacks tensor in specified axis, this is the opposite of 

None 

View a complex Tensor as a real Tensor. 

None 

Splits input with two or more dimensions, into multiple subtensors vertically according to indices_or_sections. 

None 

Stacks tensors in sequence vertically. 

None 

Selects elements from input or other based on condition and returns a tensor. 

None 

Computes the cross product of input and other in dimension dim. 

None 

Renormalizes the subtensors along dimension axis, and each subtensor's pnorm should not exceed the maxnorm. 

None 
Type Cast
API Name 
Description 
Supported Platforms 
Warning 
Returns a tensor with the new specified data type. 

None 

Check whether the input object is a 

None 

The interface is deprecated from version 2.3 and will be removed in a future version, please use int(x) or float(x) instead. 
Deprecated 
None 

Converts a scalar to a Tensor, and converts the data type to the specified type. 

None 

Converts a tuple to a tensor. 

None 
Gradient Clipping
API Name 
Description 
Supported Platforms 
Warning 
Clips tensor values by the ratio of the sum of their norms. 

None 

Clips tensor values to a specified min and max. 

None 

Clip norm of a set of input Tensors. 

None 
Parameter Operation Functions
API Name 
Description 
Supported Platforms 
Warning 
Assigns Parameter with a value. 

None 

Updates a Parameter by adding a value to it. 

None 

Updates a Parameter by subtracting a value from it. 

None 

Using given values to update tensor value through the add operation, along with the input indices. 

None 

Using given values to update tensor value through the div operation, along with the input indices. 

None 

Using given values to update tensor value through the min operation, along with the input indices. 

None 

Using given values to update tensor value through the max operation, along with the input indices. 

None 

Using given values to update tensor value through the mul operation, along with the input indices. 

None 

Applies sparse addition to individual values or slices in a tensor. 

None 

Applying sparse division to individual values or slices in a tensor. 

None 

Applying sparse maximum to individual values or slices in a tensor. 

None 

Applying sparse minimum to individual values or slices in a tensor. 

None 

Applies sparse multiplication to individual values or slices in a tensor. 

None 

Applies sparse subtraction to individual values or slices in a tensor. 

None 

Updates tensor values by using input indices and value. 

None 
Differential Functions
API Name 
Description 
Supported Platforms 
Warning 
This function is designed to calculate the higher order differentiation of given composite function. 

None 

This function is designed to calculate the higher order differentiation of given composite function. 

None 

StopGradient is used for eliminating the effect of a value on the gradient, such as truncating the gradient propagation from an output of a function. 

None 
Debugging Functions
API Name 
Description 
Supported Platforms 
Warning 
Outputs the inputs to stdout. 

None 

Save Tensor in numpy's npy format. 

None 
Sparse Functions
Warning
These are experimental APIs that are subject to change or deletion.
API Name 
Description 
Supported Platforms 
Warning 
Convert a Tensor to COOTensor. 

None 

Convert a Tensor to CSRTensor. 

None 

Converts a CSRTensor to COOTensor. 

None 
COO Functions
Warning
These are experimental APIs that are subject to change or deletion.
API Name 
Description 
Supported Platforms 
Warning 
Returns coo_absolute value of a COOTensor elementwise. 

None 

Computes arccosine of input coo_tensors elementwise. 

None 

Computes inverse hyperbolic cosine of the inputs elementwise. 

Given an input COOTensor x, the function computes inverse hyperbolic cosine of every element. Input range is [1, inf]. 

Computes the sum of x1(COOTensor) and x2(COOTensor), and return a new COOTensor based on the computed result and thresh. 

None 

Computes arcsine of input coo_tensors elementwise. 

None 

Computes inverse hyperbolic sine of the input elementwise. 

None 

Computes the trigonometric inverse tangent of the input elementwise. 

None 

Computes inverse hyperbolic tangent of the input elementwise. 

This is an experimental API that is subject to change or deletion. 

Rounds a COOTensor up to the closest integer elementwise. 

None 

concatenates the input SparseTensor(COO format) along the specified dimension. 

This is an experimental API that is subjected to change or deletion. Only supported on CPU now. 

Computes cosine of input elementwise. 

If use float64, there may be a problem of missing precision. 

Computes hyperbolic cosine of input elementwise. 

None 

Returns the elementwise exponential of a COOTensor. 

None 

Returns exponential then minus 1 of a COOTensor elementwise. 

None 

Rounds a COOTensor down to the closest integer elementwise. 

None 

Computes Reciprocal of input COOTensor elementwise. 

None 

Determines which elements are finite for each position. 

None 

Determines which elements are inf or inf for each position. 

None 

Determines which elements are NaN for each position. 

None 

Returns the natural logarithm of a COOTensor elementwise. 

If the input value of operator Log is within the range (0, 0.01] or [0.95, 1.05], the output accuracy may be affacted. 

Returns the natural logarithm of one plus the input COOTensor elementwise. 

None 

Returns a COOTensor with coo_negative values of the input COOTensor elementwise. 

None 

Computes ReLU (Rectified Linear Unit activation function) of input coo_tensors elementwise. 

None 

Computes ReLU (Rectified Linear Unit) upper bounded by 6 of input coo_tensors elementwise. 

None 

Returns half to even of a COOTensor elementwise. 

None 

Sigmoid activation function. 

None 

Computes sine of the input elementwise. 

None 

Computes hyperbolic sine of the input elementwise. 

None 

Softsign activation function. 

None 

Returns sqrt of a COOTensor elementwise. 

None 

Returns square of a COOTensor elementwise. 

None 

Computes tangent of x elementwise. 

None 

Computes hyperbolic tangent of input elementwise. 

None 
CSR Functions
Warning
These are experimental APIs that are subject to change or deletion.
API Name 
Description 
Supported Platforms 
Warning 
Returns csr_absolute value of a CSRTensor elementwise. 

None 

Computes arccosine of input csr_tensors elementwise. 

None 

Computes inverse hyperbolic cosine of the inputs elementwise. 

None 

Computes the linear combination of two input CSRTensors a and b. 

None 

Computes arcsine of input csr_tensors elementwise. 

None 

Computes inverse hyperbolic sine of the input elementwise. 

None 

Computes the trigonometric inverse tangent of the input elementwise. 

None 

Computes inverse hyperbolic tangent of the input elementwise. 

This is an experimental API that is subject to change or deletion. 

Rounds a CSRTensor up to the closest integer elementwise. 

None 

Computes cosine of input elementwise. 

Currently support data types float16 and float32. If use float64, there may be a problem of missing precision. 

Computes hyperbolic cosine of input elementwise. 

None 

Returns csr_exponential of a CSRTensor elementwise. 

None 

Returns exponential then minus 1 of a CSRTensor elementwise. 

None 

Rounds a CSRTensor down to the closest integer elementwise. 

None 

Computes Reciprocal of input CSRTensor elementwise. 

None 

Determines which elements are finite for each position. 

None 

Determines which elements are inf or inf for each position. 

None 

Determines which elements are NaN for each position. 

None 

Returns the natural logarithm of a CSRTensor elementwise. 

If the input value of operator Log is within the range (0, 0.01] or [0.95, 1.05], the output accuracy may be affacted. 

Returns the natural logarithm of one plus the input CSRTensor elementwise. 

None 

Return the matrix multiplication result of the rightmultiply matrix (dense or CSRTensor) of the CSRTensor. 

None 

Returns a CSRTensor with csr_negative values of the input CSRTensor elementwise. 

None 

Computes ReLU (Rectified Linear Unit activation function) of input csr_tensors elementwise. 

None 

Computes ReLU (Rectified Linear Unit) upper bounded by 6 of input csr_tensors elementwise. 

None 

Returns half to even of a CSRTensor elementwise. 

None 

Sigmoid activation function. 

None 

Computes sine of the input elementwise. 

None 

Computes hyperbolic sine of the input elementwise. 

None 

Calculates the softmax of a CSRTensorMatrix. 

None 

Softsign activation function. 

None 

Returns sqrt of a CSRTensor elementwise. 

None 

Returns square of a CSRTensor elementwise. 

None 

Computes tangent of x elementwise. 

None 

Computes hyperbolic tangent of input elementwise. 

None 