Convenience function for accessing the global UnitTest implementation object.

get a onnx node from singa operator, prepare its type, inputs and outputs Args: op: a given operator Args:

Check if the devices of the input tensor are the same. Keep the device where each tensors is located the same as the first tensor. C

Accessors for the implementation object.

Sets the model in evaluation mode.

Natural logarithm. This function is exact to rounding for all rounding modes. See also:** Documentation for [std::log](https://en.cppreference.com/w/

Returns: the number of dimensions of the tensor.

Move the tensor data onto a given device. Args: device: a swig Device converted from CudaGPU or CppCPU or OpenclGPU

Makes a successful assertion result.

TODO(wangwei) return Block to the memory manager

Performs backward propagation from the loss and parameter update. From the loss, it performs backward propagation to get the gradients and do

Check for NaN. See also:** Documentation for [std::isnan](https://en.cppreference.com/w/cpp/numeric/math/isnan). \param arg number to check \retval tr

Sum of tensor elements over given axis Args: t: Singa.tensor The array_like tensor to be sumed axis: None or int or t

Returns: the number of elements of the tensor.

Makes a failed assertion result.

Performs a single optimization step. Args: param_name(String): the name of the param param_value(Tensor): par

Close the server.

Register a list of sublayers. Can only be called once in each subclass. Args: *args: a list of sublayers or a dictionar

Performs a single optimization step. Args: param_name(String): the name of the param param_value(Tensor): par

Copy constructor.

Multiply row 'v' with each row of matrix M; write results into 'out'

normal transpose without axes

Performs a single optimization step. Args: param(Tensor): param values to be update grad(Tensor): param gradi

check which axes the x has been broadcasted Args: y_shape: the shape of result x_shape: the shape of x Return: a

for a brodcasted tensor, restore its shape of x from y_shape to x_shape Args: y_shape: the shape of result x_shape: the shape

Power function. This function may be 1 ULP off the correctly rounded exact result for any rounding mode in ~0.00025% of inputs. See also:** Documenta

Select value or signaling NaN. \param x preferred half-precision value \param y ignored half-precision value except for signaling NaN \return \a y if

Generate a value for each element following a uniform distribution. Args: low (float): the lower bound high (float):

Returns: the number of dimensions of the tensor.

Raise pole error and return infinity. \param sign half-precision value with sign bit only \return half-precision infinity with sign of \a sign \except

Set the model in evaluation mode. Args: mode(bool): when mode is True, this model will enter training mode

Routine for aborting the program which is safe to call from an exec-style death test child process, in which case the error message is propagated back

Flush is to "commit to DB"

Get the default host device which is a CppCPU device

Fixed point binary logarithm. This uses the BKM algorithm in L-mode. \param m mantissa in [1,2) as Q1.30 \param n number of iterations (at most 32) \r

Returns true iff the unit test passed (i.e. all test cases passed).

Args: ut_channels: int, the channel of output, also is the number of filters bias: bool

Args: ut_channels: int, the channel of output, also is the number of filters bias: bool

check if all input have same data type. Args: *inputs: input args consisting of only PyTensors