tvm.relay.nn

Neural network related operators.

Neural network operations.

tvm.relay.op.nn.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.op.nn.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.op.nn.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.op.nn.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

tvm.relay.op.nn.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

tvm.relay.op.nn.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.op.nn.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.op.nn.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 trnasposed 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.op.nn.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.op.nn.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.op.nn.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.op.nn.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.op.nn.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.op.nn.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.op.nn.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.op.nn.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.op.nn.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.op.nn.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.op.nn.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.op.nn.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.op.nn.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.op.nn.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