Relay Core Tensor Operators

This page contains the list of core tensor operator primitives pre-defined in tvm.relay. The core tensor operator primitives cover typical workloads in deep learning. They can represent workloads in front-end frameworks and provide basic building blocks for optimization. Since deep learning is a fast evolving field, it is possible to have operators that are not in here.

Note

This document will directly list the function signature of these operators in the python frontend.

Overview of Operators

Level 1: Basic Operators

This level enables fully connected multi-layer perceptron.

tvm.relay.log Compute elementwise log of data.
tvm.relay.sqrt Compute elementwise sqrt of data.
tvm.relay.exp Compute elementwise exp of data.
tvm.relay.sigmoid Compute elementwise sigmoid of data.
tvm.relay.add Addition with numpy-style broadcasting.
tvm.relay.subtract Subtraction with numpy-style broadcasting.
tvm.relay.multiply Multiplication with numpy-style broadcasting.
tvm.relay.divide Division with numpy-style broadcasting.
tvm.relay.mod Mod with numpy-style broadcasting.
tvm.relay.tanh Compute element-wise tanh of data.
tvm.relay.concatenate Concatenate the input tensors along the given axis.
tvm.relay.expand_dims Insert num_newaxis axises at the position given by axis.
tvm.relay.nn.softmax Computes softmax.
tvm.relay.nn.log_softmax Computes log softmax.
tvm.relay.nn.relu Rectified linear unit.
tvm.relay.nn.dropout Applies the dropout operation to the input array.
tvm.relay.nn.batch_norm Batch normalization layer (Ioffe and Szegedy, 2014).
tvm.relay.nn.bias_add add_bias operator.

Level 2: Convolutions

This level enables typical convnet models.

tvm.relay.nn.conv2d 2D convolution.
tvm.relay.nn.conv2d_transpose Two dimensional transposed convolution operator.
tvm.relay.nn.dense Dense operator.
tvm.relay.nn.max_pool2d 2D maximum pooling operator.
tvm.relay.nn.avg_pool2d 2D average pooling operator.
tvm.relay.nn.global_max_pool2d 2D global maximum pooling operator.
tvm.relay.nn.global_avg_pool2d 2D global average pooling operator.
tvm.relay.nn.upsampling Upsampling.
tvm.relay.nn.batch_flatten BatchFlatten.
tvm.relay.nn.pad Padding
tvm.relay.nn.lrn This operator takes data as input and does local response normalization.
tvm.relay.nn.l2_normalize Perform L2 normalization on the input data
tvm.relay.nn.contrib_conv2d_winograd_without_weight_transform 2D convolution with winograd algorithm.
tvm.relay.nn.contrib_conv2d_winograd_weight_transform Weight Transformation part for 2D convolution with winograd algorithm.

Level 3: Additional Math And Transform Operators

This level enables additional math and transform operators.

tvm.relay.nn.leaky_relu This operator takes data as input and does Leaky version of a Rectified Linear Unit.
tvm.relay.nn.prelu This operator takes data as input and does Leaky version of a Rectified Linear Unit.
tvm.relay.reshape Reshapes the input array.
tvm.relay.reshape_like Reshapes the input array by the size of another array.
tvm.relay.copy Copy a tensor.
tvm.relay.transpose Permutes the dimensions of an array.
tvm.relay.squeeze Squeeze axes in the array.
tvm.relay.floor Compute element-wise floor of data.
tvm.relay.ceil Compute element-wise ceil of data.
tvm.relay.trunc Compute element-wise trunc of data.
tvm.relay.clip Clip the elements in a between a_min and a_max.
tvm.relay.round Compute element-wise round of data.
tvm.relay.abs Compute element-wise absolute of data.
tvm.relay.negative Compute element-wise negative of data.
tvm.relay.take Take elements from an array along an axis.
tvm.relay.zeros Fill array with zeros.
tvm.relay.zeros_like Returns an array of zeros, with same type and shape as the input.
tvm.relay.ones Fill array with ones.
tvm.relay.ones_like Returns an array of ones, with same type and shape as the input.
tvm.relay.full Fill array with scalar value.
tvm.relay.full_like Return a scalar value array with the same shape and type as the input array.
tvm.relay.cast Cast input tensor to data type.
tvm.relay.split Split input tensor along axis by sections or indices.

Level 4: Broadcast and Reductions

tvm.relay.right_shift Right shift with numpy-style broadcasting.
tvm.relay.left_shift Left shift with numpy-style broadcasting.
tvm.relay.equal Broadcasted elementwise test for (lhs == rhs).
tvm.relay.not_equal Broadcasted elementwise test for (lhs != rhs).
tvm.relay.greater Broadcasted elementwise test for (lhs > rhs).
tvm.relay.greater_equal Broadcasted elementwise test for (lhs >= rhs).
tvm.relay.less Broadcasted elementwise test for (lhs < rhs).
tvm.relay.less_equal Broadcasted elementwise test for (lhs <= rhs).
tvm.relay.maximum Maximum with numpy-style broadcasting.
tvm.relay.minimum Minimum with numpy-style broadcasting.
tvm.relay.power Power with numpy-style broadcasting.
tvm.relay.where Selecting elements from either x or y depending on the value of the condition.
tvm.relay.argmax Returns the indices of the maximum values along an axis.
tvm.relay.argmin Returns the indices of the minimum values along an axis.
tvm.relay.sum Computes the sum of array elements over given axes.
tvm.relay.max Computes the max of array elements over given axes.
tvm.relay.min Computes the min of array elements over given axes.
tvm.relay.mean Computes the mean of array elements over given axes.
tvm.relay.prod Computes the products of array elements over given axes.
tvm.relay.strided_slice Strided slice of an array.
tvm.relay.broadcast_to Return a scalar value array with the same type, broadcast to the provided shape.

Level 5: Vision/Image Operators

tvm.relay.image.resize Image resize operator.
tvm.relay.vision.multibox_prior Generate prior(anchor) boxes from data, sizes and ratios.
tvm.relay.vision.multibox_transform_loc Location transformation for multibox detection
tvm.relay.vision.nms Non-maximum suppression operator for object detection.

Level 10: Temporary Operators

This level support backpropagation of broadcast operators. It is temporary.

tvm.relay.broadcast_to_like Return a scalar value array with the same shape and type as the input array.
tvm.relay.collapse_sum_like Return a scalar value array with the same shape and type as the input array.
tvm.relay.slice_like Slice the first input with respect to the second input.
tvm.relay.layout_transform Transform the layout of a tensor
tvm.relay.device_copy Copy data from the source device to the destination device.
tvm.relay.annotation.on_device Annotate an expression with a certain device type.
tvm.relay.reverse_reshape Reshapes the input array where the special values are inferred from right to left.

Level 1 Definitions

tvm.relay.log(data)

Compute elementwise log of data.

Parameters:data (relay.Expr) – The input data
Returns:result – The computed result.
Return type:relay.Expr
tvm.relay.sqrt(data)

Compute elementwise sqrt of data.

Parameters:data (relay.Expr) – The input data
Returns:result – The computed result.
Return type:relay.Expr
tvm.relay.exp(data)

Compute elementwise exp of data.

Parameters:data (relay.Expr) – The input data
Returns:result – The computed result.
Return type:relay.Expr
tvm.relay.sigmoid(data)

Compute elementwise sigmoid of data.

Parameters:data (relay.Expr) – The input data
Returns:result – The computed result.
Return type:relay.Expr
tvm.relay.add(lhs, rhs)

Addition with numpy-style broadcasting.

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

Examples

x = relay.Var("a") # shape is [2, 3]
y = relay.Var("b") # shape is [2, 1]
z = relay.add(x, y)  # result shape is [2, 3]
tvm.relay.subtract(lhs, rhs)

Subtraction with numpy-style broadcasting.

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.multiply(lhs, rhs)

Multiplication with numpy-style broadcasting.

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.divide(lhs, rhs)

Division with numpy-style broadcasting.

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.mod(lhs, rhs)

Mod with numpy-style broadcasting.

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.tanh(data)

Compute element-wise tanh of data.

Parameters:data (relay.Expr) – The input data
Returns:result – The computed result.
Return type:relay.Expr
tvm.relay.concatenate(data, axis)

Concatenate the input tensors along the given axis.

Parameters:
  • data (Union(List[relay.Expr], Tuple[relay.Expr])) – A list of tensors.
  • axis (int) – The axis along which the tensors are concatenated.
Returns:

result – The concatenated tensor.

Return type:

relay.Expr

tvm.relay.expand_dims(data, axis, num_newaxis=1)

Insert num_newaxis axises at the position given by axis.

Parameters:
  • data (relay.Expr) – The input data to the operator.
  • axis (int) – The axis at which the input array is expanded. Should lie in range [-data.ndim - 1, data.ndim]. If axis < 0, it is the first axis inserted; If axis >= 0, it is the last axis inserted in Python’s negative indexing.
  • num_newaxis (int) – Number of axes to be inserted. Should be >= 0.
Returns:

result – The reshaped result.

Return type:

relay.Expr

tvm.relay.nn.softmax(data, axis=-1)

Computes softmax.

\[\text{softmax}(x)_i = \frac{exp(x_i)}{\sum_j exp(x_j)}\]

Note

This operator can be optimized away for inference.

Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • axis (int, optional) – The axis to sum over when computing softmax
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.log_softmax(data, axis=-1)

Computes log softmax.

\[\text{log_softmax}(x)_i = \log \frac{exp(x_i)}{\sum_j exp(x_j)}\]

Note

This operator can be optimized away for inference.

Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • axis (int) – The axis to sum over when computing softmax
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.relu(data)

Rectified linear unit.

\[out = max(x, 0)\]
Parameters:data (tvm.relay.Expr) – The input data
Returns:result – The computed result.
Return type:tvm.relay.Expr
tvm.relay.nn.dropout(data, rate=0.5)

Applies the dropout operation to the input array.

During training, each element of the input is set to zero with probability p. The whole array is rescaled by 1/(1-p) to keep the expected sum of the input unchanged.

Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • rate (float, optional (default=0.5)) – The probability for an element to be reset to 0.
Returns:

result – The result of dropout

Return type:

tvm.relay.Expr

tvm.relay.nn.batch_norm(data, gamma, beta, moving_mean, moving_var, axis=1, epsilon=1e-05, center=True, scale=True)

Batch normalization layer (Ioffe and Szegedy, 2014). Normalizes the input at each batch, i.e. applies a transformation that maintains the mean activation close to 0 and the activation standard deviation close to 1.

\[\begin{split}data\_mean[i] = mean(data[:,i,:,...]) \\ data\_var[i] = var(data[:,i,:,...])\end{split}\]

Then compute the normalized output, which has the same shape as input, as following:

\[out[:,i,:,...] = \frac{data[:,i,:,...] - data\_mean[i]}{\sqrt{data\_var[i]+\epsilon}} * gamma[i] + beta[i]\]

Both mean and var returns a scalar by treating the input as a vector.

Assume the input has size k on axis 1, then both gamma and beta have shape (k,).

Besides the inputs and the outputs, this operator accepts two auxiliary states, moving_mean and moving_var, which are k-length vectors. They are global statistics for the whole dataset, which are updated by:

moving_mean = moving_mean * momentum + data_mean * (1 - momentum) moving_var = moving_var * momentum + data_var * (1 - momentum)

The parameter axis specifies which axis of the input shape denotes the ‘channel’ (separately normalized groups). The default is 1. Specifying -1 sets the channel axis to be the last item in the input shape.

Note

This operator can be optimized away for inference.

Parameters:
  • data (tvm.relay.Expr) – Input to which batch_norm will be applied.
  • gamma (tvm.relay.Expr) – The gamma scale factor.
  • beta (tvm.relay.Expr) – The beta offset factor.
  • moving_mean (tvm.relay.Expr) – Running mean of input,
  • moving_var (tvm.relay.Expr) – Running variance of input.
  • axis (int, optional, default=1) – Specify along which shape axis the channel is specified.
  • epsilon (double, optional, default=1e-5) – Small float added to variance to avoid diving by zero.
  • center (boolean, optional, default=True) – If True, add offset of beta to normalized tensor, If False, beta is ignored.
  • scale (boolean, optional, default=True) – If true, multiply by gamma. If False, gamma is not used. When the next layer is piecewise linear (also e.g. nn.relu), this can be disabled since the scaling will be done by the next layer.
Returns:

result – Tuple of normed data (same shape as input), new running mean (k-length vector), and new running variance (k-length vector)

Return type:

relay.Tuple([tvm.relay.Expr, tvm.relay.Expr, tvm.relay.Expr])

tvm.relay.nn.bias_add(data, bias, axis=1)

add_bias operator.

Add 1D bias to the axis of data. This function is a special case of add which allows inference of shape of the bias from data.

Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • bias (tvm.relay.Expr) – The bias to be added.
  • axis (int, optional) – The axis to add the bias.
Returns:

result – The final result.

Return type:

tvm.relay.Expr

Level 2 Definitions

tvm.relay.nn.conv2d(data, weight, strides=(1, 1), padding=(0, 0), dilation=(1, 1), groups=1, channels=None, kernel_size=None, data_layout='NCHW', kernel_layout='OIHW', out_layout='', out_dtype='')

2D convolution.

This operator takes the weight as the convolution kernel and convolves it with data to produce an output.

In the default case, where the data_layout is NCHW and kernel_layout is OIHW, conv2d takes in a data Tensor with shape (batch_size, in_channels, height, width), and a weight Tensor with shape (channels, in_channels, kernel_size[0], kernel_size[1]) to produce an output Tensor with the following rule:

\[\mbox{out}[b, c, y, x] = \sum_{dy, dx, k} \mbox{data}[b, k, \mbox{strides}[0] * y + dy, \mbox{strides}[1] * x + dx] * \mbox{weight}[c, k, dy, dx]\]

Padding and dilation are applied to data and weight respectively before the computation. This operator accepts data layout specification. Semantically, the operator will convert the layout to the canonical layout (NCHW for data and OIHW for weight), perform the computation, then convert to the out_layout.

Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • weight (tvm.relay.Expr) – The weight expressions.
  • strides (tuple of int, optional) – The strides of convoltution.
  • padding (tuple of int, optional) – The padding of convolution on both sides of inputs before convolution.
  • dilation (tuple of int, optional) – Specifies the dilation rate to be used for dilated convolution.
  • groups (int, optional) – Number of groups for grouped convolution.
  • channels (int, optional) – Number of output channels of this convolution.
  • kernel_size (tuple of int, optional) – The spatial of the convolution kernel.
  • data_layout (str, optional) – Layout of the input.
  • kernel_layout (str, optional) – Layout of the weight.
  • out_layout (str, optional) – Layout of the output, by default, out_layout is the same as data_layout
  • out_dtype (str, optional) – Specifies the output data type for mixed precision conv2d.
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.conv2d_transpose(data, weight, strides=(1, 1), padding=(0, 0), dilation=(1, 1), groups=1, channels=None, kernel_size=None, data_layout='NCHW', kernel_layout='OIHW', output_padding=(0, 0), out_dtype='')

Two dimensional transposed convolution operator.

Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • weight (tvm.relay.Expr) – The weight expressions.
  • strides (Tuple[int], optional) – The strides of convoltution.
  • padding (Tuple[int], optional) – The padding of convolution on both sides of inputs.
  • dilation (Tuple[int], optional) – Specifies the dilation rate to be used for dilated convolution.
  • groups (int, optional) – Number of groups for grouped convolution.
  • data_layout (str, optional) – Layout of the input.
  • kernel_layout (str, optional) – Layout of the weight.
  • output_padding (Tuple[int], optional) – Additional zero-padding to be added to one side of the output.
  • out_dtype (str, optional) – Specifies the output data type for mixed precision conv2d.
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.dense(data, weight, units=None)

Dense operator. Applies a linear transformation

\[\]

Y = X * W

Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • weight (tvm.relay.Expr) – The weight expressions.
  • units (int, optional) – Number of hidden units of the dense transformation.
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.max_pool2d(data, pool_size=(1, 1), strides=(1, 1), padding=(0, 0), layout='NCHW', ceil_mode=False)

2D maximum pooling operator.

This operator takes data as input and does 2D max value calculation with in pool_size sized window by striding defined by stride

In the default case, where the data_layout is NCHW a data Tensor with shape (batch_size, in_channels, height, width), to produce an output Tensor with the following rule:

with data of shape (b, c, h, w) and pool_size (kh, kw)

\[\mbox{out}(b, c, y, x) = \max_{m=0, \ldots, kh-1} \max_{n=0, \ldots, kw-1} \mbox{data}(b, c, \mbox{stride}[0] * y + m, \mbox{stride}[1] * x + n)\]

Padding is applied to data before the computation. ceil_mode is used to take ceil or floor while computing out shape. This operator accepts data layout specification.

Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • strides (tuple of int, optional) – The strides of pooling.
  • padding (tuple of int, optional) – The padding for pooling.
  • layout (str, optional) – Layout of the input.
  • ceil_mode (bool, optional) – To enable or disable ceil while pooling.
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.avg_pool2d(data, pool_size=(1, 1), strides=(1, 1), padding=(0, 0), layout='NCHW', ceil_mode=False, count_include_pad=False)

2D average pooling operator.

This operator takes data as input and does 2D average value calculation with in pool_size sized window by striding defined by stride

In the default case, where the data_layout is NCHW a data Tensor with shape (batch_size, in_channels, height, width), to produce an output Tensor with the following rule:

with data of shape (b, c, h, w), pool_size (kh, kw)

\[\mbox{out}(b, c, y, x) = \frac{1}{kh * kw} \sum_{m=0}^{kh-1} \sum_{n=0}^{kw-1} \mbox{data}(b, c, \mbox{stride}[0] * y + m, \mbox{stride}[1] * x + n)\]

Padding is applied to data before the computation. ceil_mode is used to take ceil or floor while computing out shape. count_include_pad indicates including or excluding padded input values in computation. This operator accepts data layout specification.

Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • strides (tuple of int, optional) – The strides of pooling.
  • padding (tuple of int, optional) – The padding for pooling.
  • layout (str, optional) – Layout of the input.
  • ceil_mode (bool, optional) – To enable or disable ceil while pooling.
  • count_include_pad (bool, optional) – To include padding to compute the average.
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.global_max_pool2d(data, layout='NCHW')

2D global maximum pooling operator.

This operator takes data as input and does 2D max value calculation across each window represented by WxH.

In the default case, where the data_layout is NCHW a data Tensor with shape (batch_size, in_channels, height, width), to produce an output Tensor with the following rule:

with data of shape (b, c, h, w)

\[\mbox{out}(b, c, 1, 1) = \max_{m=0, \ldots, h} \max_{n=0, \ldots, w} \mbox{data}(b, c, m, n)\]
Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • layout (str, optional) – Layout of the input.
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.global_avg_pool2d(data, layout='NCHW')

2D global average pooling operator.

This operator takes data as input and does 2D average value calculation across each window represented by WxH.

In the default case, where the data_layout is NCHW a data Tensor with shape (batch_size, in_channels, height, width), to produce an output Tensor with the following rule:

with data of shape (b, c, h, w)

\[\mbox{out}(b, c, 1, 1) = \frac{1}{h * w} \sum_{m=0}^{h-1} \sum_{n=0}^{w-1} \mbox{data}(b, c, m, n)\]
Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • layout (str, optional) – Layout of the input.
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.upsampling(data, scale=1, layout='NCHW', method='NEAREST_NEIGHBOR')

Upsampling.

This operator takes data as input and does 2D scaling to the given scale factor. In the default case, where the data_layout is NCHW with data of shape (n, c, h, w) out will have a shape (n, c, h*scale, w*scale)

method indicates the algorithm to be used while calculating ghe out value and method can be one of (“BILINEAR”, “NEAREST_NEIGHBOR”)

Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • scale (tvm.relay.Expr) – The scale factor for upsampling.
  • layout (str, optional) – Layout of the input.
  • method (str, optional) – Scale method to used [NEAREST_NEIGHBOR, BILINEAR].
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.batch_flatten(data)

BatchFlatten.

This operator flattens all the dimensions except for the batch dimension. which results a 2D output.

For data with shape (d1, d2, ..., dk) batch_flatten(data) returns reshaped output of shape (d1, d2*...*dk).

Parameters:data (tvm.relay.Expr) – The input data to the operator.
Returns:result – The Flattened result.
Return type:tvm.relay.Expr
tvm.relay.nn.pad(data, pad_width, pad_value=0.0)

Padding

This operator takes in a tensor and pads each axis by the specified widths using the specified value.

Parameters:
  • data (tvm.relay.Expr) – The input data to the operator
  • pad_width (tuple of <tuple of <int>>, required) – Number of values padded to the edges of each axis, in the format of ((before_1, after_1), …, (before_N, after_N))
  • pad_value (float, optional, default=0.0) – The value used for padding
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.lrn(data, size=5, axis=1, bias=2, alpha=1e-05, beta=0.75)

This operator takes data as input and does local response normalization.

Normalize the input in a local region across or within feature maps. Each input value is divided by (data / (bias + (alpha * sum_data ^2 /size))^beta) where n is the size of each local region, and the sum is taken over the region centered at that value (zero padding is added where necessary).

\[(data / (bias + (alpha * sum_data ^2 /size))^beta)\]
Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • size (int, optional) – The size of the local region to be considered for normalization.
  • axis (int, optional) – Input data layout channel axis. Default value is 1 for NCHW format
  • bias (float, optional) – The offset parameter to avoid dividing by 0.
  • alpha (float, optional) – The scaling parameter.
  • beta (float, optional) – The exponent parameter.
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.l2_normalize(data, eps, axis=None)

Perform L2 normalization on the input data

\[y(i, j) = x(i, j) / sqrt(max(sum(x^2), eps))\]
Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • eps (float) – epsilon value
  • axis (list of int, optional) – axis over the normalization applied
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.contrib_conv2d_winograd_without_weight_transform(data, weight, tile_size, strides=(1, 1), padding=(0, 0), dilation=(1, 1), groups=1, channels=None, kernel_size=None, data_layout='NCHW', kernel_layout='OIHW', out_layout='', out_dtype='')

2D convolution with winograd algorithm.

The basic parameters are the same as the ones in vanilla conv2d. It assumes the weight is pre-transformed by nn.contrib_conv2d_winograd_weight_transform

Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • weight (tvm.relay.Expr) – The weight expressions.
  • tile_size (int) – The Tile size of winograd. E.g. 2 for F(2x2, 3x3) and 4 for F(4x4, 3x3)
  • strides (tuple of int, optional) – The strides of convoltution.
  • padding (tuple of int, optional) – The padding of convolution on both sides of inputs before convolution.
  • dilation (tuple of int, optional) – Specifies the dilation rate to be used for dilated convolution.
  • groups (int, optional) – Number of groups for grouped convolution.
  • channels (int, optional) – Number of output channels of this convolution.
  • kernel_size (tuple of int, optional) – The spatial of the convolution kernel.
  • data_layout (str, optional) – Layout of the input.
  • kernel_layout (str, optional) – Layout of the weight.
  • out_layout (str, optional) – Layout of the output, by default, out_layout is the same as data_layout
  • out_dtype (str, optional) – Specifies the output data type for mixed precision conv2d.
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.contrib_conv2d_winograd_weight_transform(weight, tile_size)

Weight Transformation part for 2D convolution with winograd algorithm.

We separate this as a single op to enable pre-compute for inference. Use this together with nn.contrib_conv2d_winograd_without_weight_transform

Parameters:
  • weight (tvm.relay.Expr) – The weight expressions.
  • tile_size (int) – The Tile size of winograd. E.g. 2 for F(2x2, 3x3) and 4 for F(4x4, 3x3)
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

Level 3 Definitions

tvm.relay.nn.leaky_relu(data, alpha)

This operator takes data as input and does Leaky version of a Rectified Linear Unit.

\[`y = x > 0 ? x : alpha * x`\]
Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • alpha (float) – Slope coefficient for the negative half axis.
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.nn.prelu(data, alpha, axis=1)

This operator takes data as input and does Leaky version of a Rectified Linear Unit.

\[`y = x > 0 ? x : alpha * x`\]
Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • alpha (tvm.relay.Expr) – Slope coefficient for the negative half axis.
  • axis (int, optional) – Specify which shape axis the channel is specified.
Returns:

result – The computed result.

Return type:

tvm.relay.Expr

tvm.relay.reshape(data, newshape)

Reshapes the input array.

Example:

To give user more convenience in without doing manual shape inference, some dimensions of the shape can take special values from the set {0, -1, -2, -3, -4}. The significance of each is explained below:

  • 0 copy this dimension from the input to the output shape.

Example:

- data.shape = (2,3,4), newshape = (4,0,2), result.shape = (4,3,2)
- data.shape = (2,3,4), newshape = (2,0,0), result.shape = (2,3,4)
  • -1 infers the dimension of the output shape by using the remainder of the input dimensions

keeping the size of the new array same as that of the input array. At most one dimension of shape can be -1.

Example:

- data.shape = (2,3,4), newshape = (6,1,-1), result.shape = (6,1,4)
- data.shape = (2,3,4), newshape = (3,-1,8), result.shape = (3,1,8)
- data.shape = (2,3,4), newshape = (-1,), result.shape = (24,)
  • -2 copy all/remainder of the input dimensions to the output shape.

Example:

- data.shape = (2,3,4), newshape = (-2,), result.shape = (2,3,4)
- data.shape = (2,3,4), newshape = (2,-2), result.shape = (2,3,4)
- data.shape = (2,3,4), newshape = (-2,1,1), result.shape = (2,3,4,1,1)
  • -3 use the product of two consecutive dimensions of the input shape

as the output dimension.

Example:

- data.shape = (2,3,4), newshape = (-3,4), result.shape = (6,4)
- data.shape = (2,3,4,5), newshape = (-3,-3), result.shape = (6,20)
- data.shape = (2,3,4), newshape = (0,-3), result.shape = (2,12)
- data.shape = (2,3,4), newshape = (-3,-2), result.shape = (6,4)
  • -4 split one dimension of the input into two dimensions passed subsequent

to -4 in shape (can contain -1).

Example:

- data.shape = (2,3,4), newshape = (-4,1,2,-2), result.shape = (1,2,3,4)
- data.shape = (2,3,4), newshape = (2,-4,-1,3,-2), result.shape = (2,1,3,4)
Parameters:
  • data (relay.Expr) – The input data to the operator.
  • newshape (Union[int, Tuple[int], List[int]]) – The new shape. Should be compatible with the original shape.
Returns:

result – The reshaped result.

Return type:

relay.Expr

tvm.relay.reshape_like(data, shape_like)

Reshapes the input array by the size of another array. For an input array with shape (d1, d2, ..., dk), reshape_like operation reshapes the input array into an output array with the same shape as the second input array. .. note:: Sizes for both array should be compatible.

Parameters:
  • data (relay.Expr) – The input data to the operator.
  • shape_like (tuple of int) – The new shape. Should be compatible with the original shape.
Returns:

ret – The computed result.

Return type:

relay.Expr

tvm.relay.copy(data)

Copy a tensor.

Parameters:data (relay.Expr) – The tensor to be copied.
Returns:result – The copied result.
Return type:relay.Expr
tvm.relay.transpose(data, axes=None)

Permutes the dimensions of an array.

Parameters:
  • data (relay.Expr) – The input data to the operator.
  • axes (None or List[int]) – The target axes order, reverse order if not specified.
Returns:

result – The transposed result.

Return type:

relay.Expr

tvm.relay.squeeze(data, axis=None)

Squeeze axes in the array.

Parameters:
  • data (tvm.relay.Expr) – The input data to the operator.
  • axis (None or List[int]) – The set of axes to remove. If axis = None, remove all axis of dimensions 1. If any specified axis has dimension that does not equal 1, it is an error.
Returns:

result – The squeezed result.

Return type:

tvm.relay.Expr

tvm.relay.floor(data)

Compute element-wise floor of data.

Parameters:data (relay.Expr) – The input data
Returns:result – The computed result.
Return type:relay.Expr
tvm.relay.ceil(data)

Compute element-wise ceil of data.

Parameters:data (relay.Expr) – The input data
Returns:result – The computed result.
Return type:relay.Expr
tvm.relay.trunc(data)

Compute element-wise trunc of data.

Parameters:data (relay.Expr) – The input data
Returns:result – The computed result.
Return type:relay.Expr
tvm.relay.clip(a, a_min, a_max)

Clip the elements in a between a_min and a_max. a_min and a_max are cast to a’s dtype.

Parameters:
  • a (relay.Expr) – The input tensor.
  • a_min (float) – The clip minimum.
  • a_max (float) – The clip maximum.
Returns:

resulta with elements clipped between a_min and a_max.

Return type:

relay.Expr

Examples

tvm.relay.round(data)

Compute element-wise round of data.

Parameters:data (relay.Expr) – The input data
Returns:result – The computed result.
Return type:relay.Expr
tvm.relay.abs(data)

Compute element-wise absolute of data.

Parameters:data (relay.Expr) – The input data
Returns:result – The computed result.
Return type:relay.Expr
tvm.relay.negative(data)

Compute element-wise negative of data.

Parameters:data (relay.Expr) – The input data
Returns:result – The computed result.
Return type:relay.Expr
tvm.relay.take(data, indices, axis=None)

Take elements from an array along an axis.

Parameters:
  • data (relay.Expr) – The source array.
  • indices (rely.Expr) – The indices of the values to extract.
  • axis (int, optional) – The axis over which to select values. By default, the flattened input array is used.
Returns:

ret – The computed result.

Return type:

relay.Expr

tvm.relay.zeros(shape, dtype)

Fill array with zeros.

Parameters:
  • shape (tuple of int) – The shape of the target.
  • dtype (data type) – The data type of the target.
Returns:

result – The resulting tensor.

Return type:

relay.Expr

tvm.relay.zeros_like(data)

Returns an array of zeros, with same type and shape as the input.

Parameters:data (relay.Expr) – The input data
Returns:result – The computed result.
Return type:relay.Expr
tvm.relay.ones(shape, dtype)

Fill array with ones.

Parameters:
  • shape (tuple of int) – The shape of the target.
  • dtype (data type) – The data type of the target.
Returns:

result – The resulting tensor.

Return type:

relay.Expr

tvm.relay.ones_like(data)

Returns an array of ones, with same type and shape as the input.

Parameters:data (relay.Expr) – The input data
Returns:result – The computed result.
Return type:relay.Expr
tvm.relay.full(fill_value, shape=(), dtype='')

Fill array with scalar value.

Parameters:
  • fill_value (relay.Expr) – The value to fill. Must be a scalar.
  • shape (tuple of int) – The shape of the target.
  • dtype (data type, optional (defaults to data type of the fill value)) – The data type of the target.
Returns:

result – The resulting tensor.

Return type:

relay.Expr

tvm.relay.full_like(data, fill_value)

Return a scalar value array with the same shape and type as the input array.

Parameters:
  • data (relay.Expr) – The input tensor.
  • fill_value (relay.Expr) – The scalar value to fill.
Returns:

result – The resulting tensor.

Return type:

relay.Expr

tvm.relay.cast(data, dtype)

Cast input tensor to data type.

Parameters:
  • data (relay.Expr) – The input data to the operator.
  • dtype (str) – The target data type
Returns:

result – The casted result.

Return type:

relay.Expr

tvm.relay.split(data, indices_or_sections, axis=0)

Split input tensor along axis by sections or indices.

If indices_or_sections is an integer, the input will be divided equally along given axis. If such a split is not possible, an error is raised.

If indices_or_sections is a tuple of sorted integers, the entries indicate where along axis the array is split.

Parameters:
  • data (relay.Expr) – The source array.
  • indices_or_sections (int or tuple of int) – Indices or sections to split into. Accepts an int or a tuple
  • axis (int, optional) – The axis over which to split.
Returns:

ret – The computed result.

Return type:

relay.Tuple([relay.Expr, relay.Expr])

Level 4 Definitions

tvm.relay.right_shift(lhs, rhs)

Right shift with numpy-style broadcasting.

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.left_shift(lhs, rhs)

Left shift with numpy-style broadcasting.

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.equal(lhs, rhs)

Broadcasted elementwise test for (lhs == rhs).

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.not_equal(lhs, rhs)

Broadcasted elementwise test for (lhs != rhs).

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.greater(lhs, rhs)

Broadcasted elementwise test for (lhs > rhs).

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.greater_equal(lhs, rhs)

Broadcasted elementwise test for (lhs >= rhs).

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.less(lhs, rhs)

Broadcasted elementwise test for (lhs < rhs).

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.less_equal(lhs, rhs)

Broadcasted elementwise test for (lhs <= rhs).

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.maximum(lhs, rhs)

Maximum with numpy-style broadcasting.

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.minimum(lhs, rhs)

Minimum with numpy-style broadcasting.

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.power(lhs, rhs)

Power with numpy-style broadcasting.

Parameters:
  • lhs (relay.Expr) – The left hand side input data
  • rhs (relay.Expr) – The right hand side input data
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.where(condition, x, y)

Selecting elements from either x or y depending on the value of the condition.

Parameters:
  • condition (relay.Expr) – The condition array. The n-th element in y is selected when the n-th value in the condition array is zero. Otherwise, the corresponding element from x will be picked.
  • x (relay.Expr) – The first array to be selected.
  • y (relay.Expr) – The second array to be selected.
Returns:

result – The selected array.

Return type:

relay.Expr

Examples

x = [[1, 2], [3, 4]]
y = [[5, 6], [7, 8]]
condition = [[0, 1], [-1, 0]]
relay.where(conditon, x, y) = [[5, 2], [3, 8]]

condition = [1, 0]
relay.where(conditon, x, y) = [[1, 2], [7, 8]]

Note that the shape of condition, x, and y needs to be the same.

tvm.relay.argmax(data, axis=None, keepdims=False, exclude=False)

Returns the indices of the maximum values along an axis.

Parameters:
  • data (relay.Expr) – The input data
  • axis (None or int or tuple of int) – Axis or axes along which a argmax operation is performed. The default, axis=None, will find the indices of the maximum element of the elements of the input array. If axis is negative it counts from the last to the first axis.
  • keepdims (bool) – If this is set to True, the axes which are reduced are left in the result as dimensions with size one. With this option, the result will broadcast correctly against the input array.
  • exclude (bool) – If exclude is true, reduction will be performed on the axes that are NOT in axis instead.
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.argmin(data, axis=None, keepdims=False, exclude=False)

Returns the indices of the minimum values along an axis.

Parameters:
  • data (relay.Expr) – The input data
  • axis (None or int or tuple of int) – Axis or axes along which a argmin operation is performed. The default, axis=None, will find the indices of minimum element all of the elements of the input array. If axis is negative it counts from the last to the first axis.
  • keepdims (bool) – If this is set to True, the axes which are reduced are left in the result as dimensions with size one. With this option, the result will broadcast correctly against the input array.
  • exclude (bool) – If exclude is true, reduction will be performed on the axes that are NOT in axis instead.
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.sum(data, axis=None, keepdims=False, exclude=False)

Computes the sum of array elements over given axes.

Parameters:
  • data (relay.Expr) – The input data
  • axis (None or int or tuple of int) – Axis or axes along which a sum is performed. The default, axis=None, will sum all of the elements of the input array. If axis is negative it counts from the last to the first axis.
  • keepdims (bool) – If this is set to True, the axes which are reduced are left in the result as dimensions with size one. With this option, the result will broadcast correctly against the input array.
  • exclude (bool) – If exclude is true, reduction will be performed on the axes that are NOT in axis instead.
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.max(data, axis=None, keepdims=False, exclude=False)

Computes the max of array elements over given axes.

Parameters:
  • data (relay.Expr) – The input data
  • axis (None or int or tuple of int) – Axis or axes along which the max operation is performed. The default, axis=None, will find the max element from all of the elements of the input array. If axis is negative it counts from the last to the first axis.
  • keepdims (bool) – If this is set to True, the axes which are reduced are left in the result as dimensions with size one. With this option, the result will broadcast correctly against the input array.
  • exclude (bool) – If exclude is true, reduction will be performed on the axes that are NOT in axis instead.
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.min(data, axis=None, keepdims=False, exclude=False)

Computes the min of array elements over given axes.

Parameters:
  • data (relay.Expr) – The input data
  • axis (None or int or tuple of int) – Axis or axes along which a minimum operation is performed. The default, axis=None, will find the minimum element from all of the elements of the input array. If axis is negative it counts from the last to the first axis.
  • keepdims (bool) – If this is set to True, the axes which are reduced are left in the result as dimensions with size one. With this option, the result will broadcast correctly against the input array.
  • exclude (bool) – If exclude is true, reduction will be performed on the axes that are NOT in axis instead.
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.mean(data, axis=None, keepdims=False, exclude=False)

Computes the mean of array elements over given axes.

Parameters:
  • data (relay.Expr) – The input data
  • axis (None or int or tuple of int) – Axis or axes along which a mean operation is performed. The default, axis=None, will find the indices of minimum element all of the elements of the input array. If axis is negative it counts from the last to the first axis.
  • keepdims (bool) – If this is set to True, the axes which are reduced are left in the result as dimensions with size one. With this option, the result will broadcast correctly against the input array.
  • exclude (bool) – If exclude is true, reduction will be performed on the axes that are NOT in axis instead.
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.prod(data, axis=None, keepdims=False, exclude=False)

Computes the products of array elements over given axes.

Parameters:
  • data (relay.Expr) – The input data
  • axis (None or int or tuple of int) – Axis or axes along which a product is performed. The default, axis=None, will find the indices of minimum element all of the elements of the input array. If axis is negative it counts from the last to the first axis.
  • keepdims (bool) – If this is set to True, the axes which are reduced are left in the result as dimensions with size one. With this option, the result will broadcast correctly against the input array.
  • exclude (bool) – If exclude is true, reduction will be performed on the axes that are NOT in axis instead.
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.strided_slice(data, begin, end, strides=None)

Strided slice of an array.

Parameters:
  • data (relay.Expr) – The source array to be sliced.
  • begin (list of int) – The indices to begin with in the slicing.
  • end (list of int) – Indicies indicating end of the slice.
  • strides (list of int, optional) – Specifies the stride values, it can be negative in that case, the input tensor will be reversed in that particular axis.
Returns:

ret – The computed result.

Return type:

relay.Expr

tvm.relay.broadcast_to(data, shape)

Return a scalar value array with the same type, broadcast to the provided shape.

Parameters:
  • data (relay.Expr) – The input tensor.
  • shape (shape) – Provide the shape to broadcast to.
Returns:

result – The resulting tensor.

Return type:

relay.Expr

Level 5 Definitions

tvm.relay.image.resize(data, size, layout='NCHW', method='BILINEAR', align_corners=False)

Image resize operator.

This operator takes data as input and does 2D scaling to the given scale factor. In the default case, where the data_layout is NCHW with data of shape (n, c, h, w) out will have a shape (n, c, size[0], size[1])

method indicates the algorithm to be used while calculating ghe out value and method can be one of (“BILINEAR”, “NEAREST_NEIGHBOR”)

Parameters:
  • data (relay.Expr) – The input data to the operator.
  • size (Tuple of Expr) – The out size to which the image will be resized.
  • layout (str, optional) – Layout of the input.
  • method (str, optional) – Scale method to used [NEAREST_NEIGHBOR, BILINEAR].
  • align_corners (int, optional) – Should be true to preserve the values at the corner pixels
Returns:

result – The resized result.

Return type:

relay.Expr

tvm.relay.vision.multibox_prior(data, sizes=(1.0, ), ratios=(1.0, ), steps=(-1.0, -1.0), offsets=(0.5, 0.5), clip=False)

Generate prior(anchor) boxes from data, sizes and ratios.

Parameters:
  • data (relay.Expr) – The input data tensor.
  • sizes (tuple of float, optional) – Tuple of sizes for anchor boxes.
  • ratios (tuple of float, optional) – Tuple of ratios for anchor boxes.
  • steps (Tuple of float, optional) – Priorbox step across y and x, -1 for auto calculation.
  • offsets (tuple of int, optional) – Priorbox center offsets, y and x respectively.
  • clip (boolean, optional) – Whether to clip out-of-boundary boxes.
Returns:

out – 3-D tensor with shape [1, h_in * w_in * (num_sizes + num_ratios - 1), 4]

Return type:

relay.Expr

tvm.relay.vision.multibox_transform_loc(cls_prob, loc_pred, anchor, clip=True, threshold=0.01, variances=(0.1, 0.1, 0.2, 0.2))

Location transformation for multibox detection

Parameters:
  • cls_prob (tvm.relay.Expr) – Class probabilities.
  • loc_pred (tvm.relay.Expr) – Location regression predictions.
  • anchor (tvm.relay.Expr) – Prior anchor boxes.
  • clip (boolean, optional) – Whether to clip out-of-boundary boxes.
  • threshold (double, optional) – Threshold to be a positive prediction.
  • variances (Tuple of float, optional) – variances to be decoded from box regression output.
Returns:

ret

Return type:

tuple of tvm.relay.Expr

tvm.relay.vision.nms(data, valid_count, overlap_threshold=0.5, force_suppress=False, topk=-1)

Non-maximum suppression operator for object detection.

Parameters:
  • data (relay.Expr) – 3-D tensor with shape [batch_size, num_anchors, 6]. The last dimension should be in format of [class_id, score, box_left, box_top, box_right, box_bottom].
  • valid_count (relay.Expr) – 1-D tensor for valid number of boxes.
  • overlap_threshold (float, optional) – Non-maximum suppression threshold.
  • force_suppress (bool, optional) – Suppress all detections regardless of class_id.
  • topk (int, optional) – Keep maximum top k detections before nms, -1 for no limit.
Returns:

out – 3-D tensor with shape [batch_size, num_anchors, 6].

Return type:

relay.Expr

Level 10 Definitions

tvm.relay.broadcast_to_like(data, broadcast_type)

Return a scalar value array with the same shape and type as the input array.

Parameters:
  • data (relay.Expr) – The input tensor.
  • broadcast_type (relay.Expr) – Provide the type to broadcast to.
Returns:

result – The resulting tensor.

Return type:

relay.Expr

tvm.relay.collapse_sum_like(data, collapse_type)

Return a scalar value array with the same shape and type as the input array.

Parameters:
  • data (relay.Expr) – The input tensor.
  • collapse_type (relay.Expr) – Provide the type to collapse to.
Returns:

result – The resulting tensor.

Return type:

relay.Expr

tvm.relay.slice_like(data, shape_like, axes=None)

Slice the first input with respect to the second input.

For an input array with shape (d1, d2, ..., dk), slice_like operation slices the the input array corresponding size of second array. By default will slice on all axes.

Parameters:
  • data (tvm.relay.Expr) – The source array.
  • shape_like (tvm.relay.Expr) – The new shape.
  • axes (Optional[Tuple[int]]) – List of axes on which input data will be sliced according to the corresponding size of the second input. By default will slice on all axes. Negative axes mean counting in reverse.
Returns:

result – The computed result.

Return type:

relay.Expr

tvm.relay.layout_transform(data, src_layout, dst_layout)

Transform the layout of a tensor

Parameters:
  • data (relay.Expr) – The source tensor to be transformed
  • src_layout (str) – The source layout. (e.g NCHW)
  • dst_layout (str) – The destination layout. (e.g. NCHW16c)
Returns:

ret – The transformed tensor.

Return type:

relay.Expr

tvm.relay.device_copy(data, src_dev, dst_dev)

Copy data from the source device to the destination device. This operator helps data transferring between difference contexts for heterogeneous execution.

Parameters:
  • data (tvm.relay.Expr) – The tensor to be copied.
  • src_dev (Union[TVMContext, str]) – The source device where the data is copied from.
  • dst_dev (Union[TVMContext, str]) – The destination device where the data is copied to.
Returns:

result – The copied result.

Return type:

tvm.relay.Expr

tvm.relay.annotation.on_device(data, device)

Annotate an expression with a certain device type.

Parameters:
  • data (tvm.relay.Expr) – The expression to be annotated.
  • device (Union[TVMContext, str]) – The device type to annotate.
Returns:

result – The annotated expression.

Return type:

tvm.relay.Expr

tvm.relay.reverse_reshape(data, newshape)

Reshapes the input array where the special values are inferred from right to left.

Example:

The special values have the same semantics as tvm.relay.reshape. The difference is that special values are inferred from right to left. It can be explained in the example below:

- data.shape = (10,5,4), newshape = (-1,0), reshape results in (40,5)
- data.shape = (10,5,4), newshape = (-1,0), reverse_reshape results in (40,5)
Parameters:
  • data (relay.Expr) – The input data to the operator.
  • newshape (Union[int, Tuple[int], List[int]]) – The new shape. Should be compatible with the original shape.
Returns:

result – The reshaped result.

Return type:

relay.Expr