tvm.relay.nn#
Neural network related operators.
Classes:
|
A constant expression in Relay. |
|
Functions:
|
1D adaptive average pooling operator. |
|
2D adaptive average pooling operator. |
|
3D adaptive avg pooling operator. |
|
1D adaptive max pooling operator. |
|
2D adaptive max pooling operator. |
|
3D adaptive max pooling operator. |
|
1D average pooling operator. |
|
2D average pooling operator. |
|
Gradient of 2D average pooling operator. |
|
3D average pooling operator. |
|
BatchFlatten. |
|
Compute batch matrix multiplication of tensor_a and tensor_b. |
|
Batch normalization layer (Ioffe and Szegedy, 2014). |
|
Reshape the batch dimension into spatial dimensions. |
|
add_bias operator. |
|
Tensor packing for bitserial operations. |
|
2D convolution using bitserial computation. |
|
Bitserial Dense operator. |
|
Create a constant value. |
Weight Transformation part for 2D convolution with gemm algorithm. |
|
2D convolution with gemm algorithm. |
|
|
Variant of 2D convolution. |
Weight Transformation part for 2D convolution with winograd algorithm. |
|
Weight Transformation part for 2D convolution with winograd algorithm. |
|
2D convolution with winograd algorithm. |
|
Weight Transformation part for 3D convolution with winograd algorithm. |
|
3D convolution with winograd algorithm. |
|
|
Dense operator. |
|
Variant of 2D depthwise convolution. |
|
1D convolution. |
|
One dimensional transposed convolution operator. |
|
2D convolution. |
|
The gradient of conv2d with respect to weight. |
|
Two dimensional transposed convolution operator. |
|
3D convolution. |
|
3D transpose convolution. |
|
Applies correlation to inputs. |
|
CrossEntropy without logits. |
|
CrossEntropy with logits. |
|
Deformable 2d convolution. |
|
Dense operator. |
|
Convert channels into spatial blocks. |
|
Dilate data with given dilation value (0 by default). |
|
Applies the dropout operation to the input array. |
|
Applies the dropout operation to the input array. |
|
Computes softmax. |
|
FIFO buffer to enable computation reuse in CNNs with sliding indow input |
|
Common code to get the 1 dimensional pad option Parameters ---------- padding : Union[int, Tuple[int, ...]] Padding size Returns ------- pad_left : int Padding size on left pad_right : int Padding size on right. |
|
Common code to get the pad option Parameters ---------- padding : Union[int, Tuple[int, ...]] Padding size Returns ------- pad_top : int Padding size on top pad_left : int Padding size on left pad_down : int Padding size on down. pad_right : int Padding size on right. |
|
Common code to get the pad option Parameters ---------- padding : Union[int, Tuple[int, ...]] Padding size Returns ------- pad_front : int Padding size on front pad_top : int Padding size on top pad_left : int Padding size on left pad_back : int Padding size on back pad_down : int Padding size on down. pad_right : int Padding size on right. |
|
1D global average pooling operator. |
|
2D global average pooling operator. |
|
3D global average pooling operator. |
|
1D global maximum pooling operator. |
|
2D global maximum pooling operator. |
|
3D global maximum pooling operator. |
|
Group normalization normalizes over group of channels for each training examples. |
|
Instance Normalization (Ulyanov and et al., 2016) Applies instance normalization to the n-dimensional input array. |
|
Perform L2 normalization on the input data |
|
Layer normalization (Lei Ba and et al., 2016). |
|
This operator takes data as input and does Leaky version of a Rectified Linear Unit. |
|
Computes log softmax. |
|
This operator takes data as input and does local response normalization. |
|
Matmul operator. |
|
1D maximum pooling operator. |
|
2D maximum pooling operator. |
|
Gradient of 2D maximum pooling operator. |
|
3D maximum pooling operator. |
|
MirrorPadding |
|
Negative log likelihood loss. |
|
Padding |
|
This operator takes data as input and does Leaky version of a Rectified Linear Unit. |
|
Rectified linear unit. |
|
Computes softmax. |
|
Divide spatial dimensions of the data into a grid of blocks and interleave them into batch dim. |
|
Convert spatial blocks into channels. |
|
Computes the matrix addition of dense_mat and sparse_mat, where dense_mat is a dense matrix and sparse_mat is a sparse (CSR) namedtuple with fields data, indices, and indptr. |
|
Computes the matrix multiplication of dense_mat and sparse_mat, where dense_mat is a dense matrix and sparse_mat is a sparse (either BSR or CSR) namedtuple with fields data, indices, and indptr. |
Computes the fast matrix transpose of x, where x is a sparse tensor in CSR format (represented as a namedtuple with fields data, indices, and indptr). |
|
|
Upsampling. |
|
3D Upsampling. |
- class tvm.relay.nn.Constant(data, span=None)#
A constant expression in Relay.
Parameters#
- datatvm.nd.NDArray
The data content of the constant expression.
- span: Optional[tvm.relay.Span]
Span that points to original source code.
- 参数:
span (Span | None)
- tvm.relay.nn.Expr#
RelayExpr
的别名 Attributes:checked_type
Get the checked type of tvm.relay.Expr.
struct_info
Get the struct info field
- tvm.relay.nn.adaptive_avg_pool1d(data, output_size=None, layout='NCW', out_layout='')#
1D adaptive average pooling operator. This operator is experimental.
This operator takes data as input and does 1D average value calculation across each window represented by W.
In the default case, where the data_layout is NCW a data Tensor with shape (batch_size, in_channels, width), to produce an output Tensor with shape (batch_size, in_channels, output_width).
The pooling kernel and stride sizes are automatically chosen for desired output sizes.
- For output_size:
If this argument is not provided, input height and width will be used as output width.
If a single integer is provided for output_size, the output size is (N x C x output_size) for any input (NCW).
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- output_sizetuple of int. optional
Output height and width.
- layoutstr, optional
Layout of the input.
- out_layoutstr, optional
Layout of the output.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.adaptive_avg_pool2d(data, output_size=None, layout='NCHW', out_layout='')#
2D adaptive average pooling operator. This operator is experimental.
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 shape (batch_size, in_channels, output_height, output_width).
The pooling kernel and stride sizes are automatically chosen for desired output sizes.
- For output_size:
If this argument is not provided, input height and width will be used as output height and width.
If a single integer is provided for output_size, the output size is (N x C x output_size x output_size) for any input (NCHW).
If a tuple of integers (height, width) are provided for output_size, the output size is (N x C x height x width) for any input (NCHW).
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- output_sizetuple of int. optional
Output height and width.
- layoutstr, optional
Layout of the input.
- out_layoutstr, optional
Layout of the output.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.adaptive_avg_pool3d(data, output_size=None, layout='NCDHW', out_layout='')#
3D adaptive avg pooling operator. This operator is experimental.
This operator takes data as input and does 3D avg value calculation across each window represented by DxWxH.
In the default case, where the data_layout is NCDHW a data Tensor with shape (batch_size, in_channels, depth, height, width), to produce an output Tensor with shape (batch_size, in_channels, output_depth, output_height, output_width).
The pooling kernel and stride sizes are automatically chosen for desired output sizes.
- For output_size:
If this argument is not provided, input depth, height and width will be used as output depth, height and width.
If a single integer is provided for output_size, the output size is (N x C x output_size x output_size x output_size) for any input (NCDHW).
If a tuple of integers (depth, height, width) are provided for output_size, the output size is (N x C x depth x height x width) for any input (NCDHW).
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- output_sizetuple of int. optional
Output height and width.
- layoutstr, optional
Layout of the input.
- out_layoutstr, optional
Layout of the output.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.adaptive_max_pool1d(data, output_size=None, layout='NCW', out_layout='')#
1D adaptive max pooling operator. This operator is experimental.
This operator takes data as input and does 1D max value calculation across each window represented by W.
In the default case, where the data_layout is NCW a data Tensor with shape (batch_size, in_channels, width), to produce an output Tensor with shape (batch_size, in_channels, output_width).
The pooling kernel and stride sizes are automatically chosen for desired output sizes.
- For output_size:
If this argument is not provided, input height and width will be used as output height and width.
If a single integer is provided for output_size, the output size is (N x C x output_size) for any input (NCW).
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- output_sizetuple of int. optional
Output height and width.
- layoutstr, optional
Layout of the input.
- out_layoutstr, optional
Layout of the output.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.adaptive_max_pool2d(data, output_size=None, layout='NCHW', out_layout='')#
2D adaptive max pooling operator. This operator is experimental.
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 shape (batch_size, in_channels, output_height, output_width).
The pooling kernel and stride sizes are automatically chosen for desired output sizes.
- For output_size:
If this argument is not provided, input height and width will be used as output height and width.
If a single integer is provided for output_size, the output size is (N x C x output_size x output_size) for any input (NCHW).
If a tuple of integers (height, width) are provided for output_size, the output size is (N x C x height x width) for any input (NCHW).
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- output_sizetuple of int. optional
Output height and width.
- layoutstr, optional
Layout of the input.
- out_layoutstr, optional
Layout of the output.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.adaptive_max_pool3d(data, output_size=None, layout='NCDHW', out_layout='')#
3D adaptive max pooling operator. This operator is experimental.
This operator takes data as input and does 3D max value calculation across each window represented by DxWxH.
In the default case, where the data_layout is NCDHW a data Tensor with shape (batch_size, in_channels, depth, height, width), to produce an output Tensor with shape (batch_size, in_channels, output_depth, output_height, output_width).
The pooling kernel and stride sizes are automatically chosen for desired output sizes.
- For output_size:
If this argument is not provided, input depth, height and width will be used as output depth, height and width.
If a single integer is provided for output_size, the output size is (N x C x output_size x output_size x output_size) for any input (NCDHW).
If a tuple of integers (depth, height, width) are provided for output_size, the output size is (N x C x depth x height x width) for any input (NCDHW).
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- output_sizetuple of int. optional
Output height and width.
- layoutstr, optional
Layout of the input.
- out_layoutstr, optional
Layout of the output.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.avg_pool1d(data, pool_size=(1,), strides=(1,), dilation=(1,), padding=(0,), layout='NCW', out_layout='', ceil_mode=False, count_include_pad=False)#
1D average pooling operator.
This operator takes data as input and does 1D average value calculation with in pool_size sized window by striding defined by stride
In the default case, where the data_layout is NCW a data Tensor with shape (batch_size, channels, width), to produce an output Tensor.
The 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#
- datatvm.relay.Expr
The input data to the operator.
- pool_sizeint or tuple of int, optional
The size of window for pooling.
- stridesint or tuple of int, optional
The strides of pooling.
- dilationint or tuple of int, optional
The dilation of pooling.
- paddingint or tuple of int, optional
The padding for pooling.
- layoutstr, optional
Layout of the input.
- out_layoutOptional[str]
Layout of the output
- ceil_modebool, optional
To enable or disable ceil while pooling.
- count_include_padbool, optional
To include padding to compute the average.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.avg_pool2d(data, pool_size=(1, 1), strides=(1, 1), dilation=(1, 1), padding=(0, 0), layout='NCHW', out_layout='', 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#
- datatvm.relay.Expr
The input data to the operator.
- pool_sizeint or tuple of int, optional
The size of window for pooling.
- stridestuple of int, optional
The strides of pooling.
- dilationint or tuple of int, optional
The dilation of pooling.
- paddingtuple of int, optional
The padding for pooling.
- layoutstr, optional
Layout of the input.
- out_layoutOptional[str]
Layout of the output
- ceil_modebool, optional
To enable or disable ceil while pooling.
- count_include_padbool, optional
To include padding to compute the average.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.avg_pool2d_grad(out_grad, data, pool_size=(1, 1), strides=(1, 1), padding=(0, 0), layout='NCHW', out_layout='', ceil_mode=False, count_include_pad=False)#
Gradient of 2D average pooling operator.
This operator takes out_grad and data as input and calculates gradient of avg_pool2d.
Parameters#
- out_gradtvm.relay.Expr
The output gradient
- datatvm.relay.Expr
The input data to the operator.
- pool_sizeint or tuple of int, optional
The size of window for pooling.
- stridestuple of int, optional
The strides of pooling.
- paddingtuple of int, optional
The padding for pooling.
- layoutstr, optional
Layout of the input.
- out_layoutOptional[str]
Layout of the output
- ceil_modebool, optional
To enable or disable ceil while pooling.
- count_include_padbool, optional
To include padding to compute the average.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.avg_pool3d(data, pool_size=(1, 1, 1), strides=(1, 1, 1), dilation=(1, 1, 1), padding=(0, 0, 0), layout='NCDHW', out_layout='', ceil_mode=False, count_include_pad=False)#
3D average pooling operator.
This operator takes data as input and does 3D average value calculation with in pool_size sized window by striding defined by stride
In the default case, where the data_layout is NCDHW a data Tensor with shape (batch_size, channels, depth, height, width), to produce an output Tensor.
The 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#
- datatvm.relay.Expr
The input data to the operator.
- pool_sizeint or tuple of int, optional
The size of window for pooling.
- stridestuple of int, optional
The strides of pooling.
- dilationint or tuple of int, optional
The dilation of pooling.
- paddingtuple of int, optional
The padding for pooling.
- layoutstr, optional
Layout of the input.
- out_layoutOptional[str]
Layout of the output
- ceil_modebool, optional
To enable or disable ceil while pooling.
- count_include_padbool, optional
To include padding to compute the average.
Returns#
- resulttvm.relay.Expr
The computed result.
- 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#
- datatvm.relay.Expr
The input data to the operator.
Returns#
- resulttvm.relay.Expr
The Flattened result.
- tvm.relay.nn.batch_matmul(tensor_a, tensor_b, out_dtype='', transpose_a=False, transpose_b=True)#
Compute batch matrix multiplication of tensor_a and tensor_b.
Both tensor_a and tensor_b can be transposed. For legacy reason, we use NT format (transpose_a=False, transpose_b=True) by default.
\[\mbox{batch_matmul}(A, B)[i, :, :] = \mbox{matmul}(A[i, :, :], B[i, :, :])\]Parameters#
- tensor_atvm.relay.Expr
The first input.
- tensor_btvm.relay.Expr
The second input.
- out_dtypeOptional[str]
Specifies the output data type for mixed precision batch matmul.
- transpose_aOptional[bool] = False
Whether the first tensor is in transposed format.
- transpose_bOptional[bool] = True
Whether the second tensor is in transposed format.
Returns#
- result: tvm.relay.Expr
The computed result.
- 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
andbeta
have shape (k,).Besides the inputs and the outputs, this operator accepts two auxiliary states,
moving_mean
andmoving_var
, which are k-length vectors. They are global statistics for the whole dataset, which are updated bymoving_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.备注
This operator can be optimized away for inference.
Parameters#
- datatvm.relay.Expr
Input to which batch_norm will be applied.
- gammatvm.relay.Expr
The gamma scale factor.
- betatvm.relay.Expr
The beta offset factor.
- moving_meantvm.relay.Expr
Running mean of input,
- moving_vartvm.relay.Expr
Running variance of input.
- axisint, optional, default=1
Specify along which shape axis the channel is specified.
- epsilondouble, optional, default=1e-5
Small float added to variance to avoid dividing by zero.
- centerboolean, optional, default=True
If True, add offset of beta to normalized tensor, If False, beta is ignored.
- scaleboolean, 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#
- resultrelay.Tuple([tvm.relay.Expr, tvm.relay.Expr, tvm.relay.Expr])
Tuple of normed data (same shape as input), new running mean (k-length vector), and new running variance (k-length vector)
- tvm.relay.nn.batch_to_space_nd(data, block_shape, crops)#
Reshape the batch dimension into spatial dimensions.
Parameters#
- datatvm.te.Tensor
N-D with shape [batch, spatial_shape, remaining_shape]
- block_shaperelay.Expr
1-D of size [M] where M is number of spatial dims, specifies block size for each spatial dimension.
- cropsrelay.Expr
2-D of shape [M, 2] where M is number of spatial dims, specifies [begin, end] crop size for each spatial dimension.
Returns#
- resultrelay.Expr
N-D Tensor with shape [batch / prod(block_shape), in_shape[1] * block_shape[0] - crops[0,0] - crops[0,1], …, in_shape[M] * block_shape[M-1] - crops[M-1, 0] - crops[M-1, 1], remaining_shape]
- 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#
- datatvm.relay.Expr
The input data to the operator.
- biastvm.relay.Expr
The bias to be added.
- axisint, optional
The axis to add the bias.
Returns#
- resulttvm.relay.Expr
The final result.
- tvm.relay.nn.bitpack(data, bits=1, pack_axis=1, bit_axis=2, pack_type='uint32', name='BitPack')#
Tensor packing for bitserial operations.
The values along the input tensor’s pack_axis are quantized and packed together into the specified pack_type in a new bit axis.
For example, consider bitpacking with data to be a tensor with shape [1, 64, 128, 128], pack_axis=1, bit_axis=4, pack_type=uint8, and bits=2. The output in this case will be of shape [1, 8, 128, 128, 2]. The dimension of axis 1 has been reduced by a factor of 8 since each value is packed into an 8-bit uint8. Axis 4 is now two bitplanes representing the quantized value of the incoming data. The output tensor is now ready to be used in a bitserial operation.
Parameters#
- datatvm.relay.expr
The incoming tensor to be packed.
- bitsint
Number of bits that should be packed.
- pack_axisint
Axis that should be decomposed and packed.
- bit_axisint
New axis containing bitplane.
- pack_typestr
Datatype to pack bits into.
- namestr, optional
Name of the operation.
Returns#
- resulttvm.relay.Expr
The packed tensor.
- tvm.relay.nn.bitserial_conv2d(data, weight, strides=(1, 1), padding=(0, 0), channels=None, kernel_size=(3, 3), activation_bits=1, weight_bits=1, data_layout='NCHW', kernel_layout='OIHW', pack_dtype='uint32', out_dtype='int16', unipolar=True)#
2D convolution using bitserial computation.
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- weighttvm.relay.Expr
The weight expressions.
- stridestuple of int, optional
The strides of convolution.
- paddingtuple of int, optional
The padding of convolution on both sides of inputs before convolution.
- channelsint, optional
Number of output channels of this convolution.
- kernel_sizetuple of int, optional
The spatial of the convolution kernel.
- activation_bitsint
Number of bits to pack for activations.
- weight_bitsint
Number of bits to pack for weights.
- data_layoutstr, optional
Layout of the input.
- kernel_layoutstr, optional
Layout of the kernel
- pack_dtype: str, optional
Datatype to pack bits into.
- out_dtypestr, optional
Specifies the output data type for mixed precision conv2d.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.bitserial_dense(data, weight, units=None, data_bits=1, weight_bits=1, pack_dtype='uint32', out_dtype='int16', unipolar=True)#
Bitserial Dense operator. Applies matrix multiplication of two quantized matrices using a fast bitserial algorithm.
\[\]Y = X * W
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- weighttvm.relay.Expr
The weight expressions.
- unitsint, optional
Number of hidden units of the dense transformation.
- data_bitsint
Number of bits incoming tensor should be packed with.
- weight_bitsint
Number of bits weight tensor should be packed with.
- pack_dtypestr, optional
Datatype to pack individual bits into before computation.
- out_dtypestr, optional
Specifies the output data type for mixed precision dense.
- unipolarbool, optional
Whether to use unipolar or bipolar quantization for inputs.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.const(value, dtype=None, span=None)#
Create a constant value.
Parameters#
- value: Union[bool, int, float, numpy.ndarray, tvm.nd.NDArray]
The constant value.
- dtype: str, optional
The data type of the resulting constant.
- span: Optional[tvm.relay.Span]
Span that points to original source code.
Note#
When dtype is None, we use the following rule:
int maps to “int32”
float maps to “float32”
bool maps to “bool”
other using the same default rule as numpy.
- tvm.relay.nn.contrib_conv2d_gemm_weight_transform(weights, tile_N, tile_K)#
Weight Transformation part for 2D convolution with gemm algorithm.
We separate this as a single op to enable pre-compute for inference. Use this together with nn.contrib_conv2d_gemm_without_weight_transform
Parameters#
- weightstvm.relay.Expr
The weight expressions.
- tile_N: int
Tile size across N axis of the weight transformation for ConvGemm. (N = OC)
- tile_K: int
Tile size across K axis of the weight transformation for ConvGemm. (K = KW * KH * IC)
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.contrib_conv2d_gemm_without_weight_transform(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 with gemm algorithm.
The basic parameters are the same as the ones in vanilla conv2d. It assumes the weight is pre-transformed by nn.contrib_conv2d_gemm_weight_transform
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- weighttvm.relay.Expr
The weight expressions.
- stridestuple of int, optional
The strides of convolution.
- paddingtuple of int, optional
The padding of convolution on both sides of inputs before convolution.
- dilationtuple of int, optional
Specifies the dilation rate to be used for dilated convolution.
- groupsint, optional
Number of groups for grouped convolution.
- channelsint, optional
Number of output channels of this convolution.
- kernel_sizetuple of int, optional
The spatial of the convolution kernel.
- data_layoutstr, optional
Layout of the input.
- kernel_layoutstr, optional
Layout of the weight.
- out_layoutstr, optional
Layout of the output, by default, out_layout is the same as data_layout
- out_dtypestr, optional
Specifies the output data type for mixed precision conv2d.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.contrib_conv2d_nchwc(data, kernel, strides=(1, 1), padding=(0, 0), dilation=(1, 1), groups=1, channels=None, kernel_size=None, data_layout='NCHW8c', kernel_layout='OIHW', out_layout='', out_dtype='')#
Variant of 2D convolution.
This operator takes the weight as the convolution kernel and convolves it with data to produce an output, following a specialized NCHWc data layout.
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- kerneltvm.relay.Expr
The kernel expressions.
- stridestuple of int, optional
The strides of convolution.
- paddingtuple of int, optional
The padding of convolution on both sides of inputs before convolution.
- dilationtuple of int, optional
Specifies the dilation rate to be used for dilated convolution.
- groupsint, optional
Number of groups for grouped convolution.
- channelsint, optional
Number of output channels of this convolution.
- kernel_sizetuple of int, optional
The spatial of the convolution kernel.
- data_layoutstr, optional
Layout of the input.
- kernel_layoutstr, optional
Layout of the weight.
- out_layoutstr, optional
Layout of the output, by default, out_layout is the same as data_layout
- out_dtypestr, optional
Specifies the output data type for mixed precision conv2d.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.contrib_conv2d_winograd_nnpack_weight_transform(weight, convolution_algorithm, out_dtype='')#
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#
- weighttvm.relay.Expr
The weight expressions.
- convolution_algorithmint
The Tile size of winograd. E.g. 2 for F(2x2, 3x3) and 4 for F(4x4, 3x3)
Returns#
- resulttvm.relay.Expr
The computed result.
- 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#
- weighttvm.relay.Expr
The weight expressions.
- tile_sizeint
The Tile size of winograd. E.g. 2 for F(2x2, 3x3) and 4 for F(4x4, 3x3)
Returns#
- resulttvm.relay.Expr
The computed result.
- 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#
- datatvm.relay.Expr
The input data to the operator.
- weighttvm.relay.Expr
The weight expressions.
- tile_sizeint
The Tile size of winograd. E.g. 2 for F(2x2, 3x3) and 4 for F(4x4, 3x3)
- stridestuple of int, optional
The strides of convolution.
- paddingtuple of int, optional
The padding of convolution on both sides of inputs before convolution.
- dilationtuple of int, optional
Specifies the dilation rate to be used for dilated convolution.
- groupsint, optional
Number of groups for grouped convolution.
- channelsint, optional
Number of output channels of this convolution.
- kernel_sizetuple of int, optional
The spatial of the convolution kernel.
- data_layoutstr, optional
Layout of the input.
- kernel_layoutstr, optional
Layout of the weight.
- out_layoutstr, optional
Layout of the output, by default, out_layout is the same as data_layout
- out_dtypestr, optional
Specifies the output data type for mixed precision conv2d.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.contrib_conv3d_winograd_weight_transform(weight, tile_size)#
Weight Transformation part for 3D convolution with winograd algorithm.
We separate this as a single op to enable pre-compute for inference. Use this together with nn.contrib_conv3d_winograd_without_weight_transform
Parameters#
- weighttvm.relay.Expr
The weight expressions.
- tile_sizeint
The Tile size of winograd. E.g. 2 for F(2x2x2, 3x3x3) and 4 for F(4x4x4, 3x3x3)
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.contrib_conv3d_winograd_without_weight_transform(data, weight, tile_size, strides=(1, 1, 1), padding=(0, 0, 0), dilation=(1, 1, 1), groups=1, channels=None, kernel_size=None, data_layout='NCDHW', kernel_layout='OIDHW', out_layout='', out_dtype='')#
3D convolution with winograd algorithm.
The basic parameters are the same as the ones in vanilla conv3d. It assumes the weight is pre-transformed by nn.contrib_conv3d_winograd_weight_transform
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- weighttvm.relay.Expr
The weight expressions.
- tile_sizeint
The Tile size of winograd. E.g. 2 for F(2x2x2, 3x3x3) and 4 for F(4x4x4, 3x3x3)
- stridestuple of int, optional
The strides of convolution.
- paddingtuple of int, optional
The padding of convolution on both sides of inputs before convolution.
- dilationtuple of int, optional
Specifies the dilation rate to be used for dilated convolution.
- groupsint, optional
Number of groups for grouped convolution.
- channelsint, optional
Number of output channels of this convolution.
- kernel_sizetuple of int, optional
The spatial of the convolution kernel.
- data_layoutstr, optional
Layout of the input.
- kernel_layoutstr, optional
Layout of the weight.
- out_layoutstr, optional
Layout of the output, by default, out_layout is the same as data_layout
- out_dtypestr, optional
Specifies the output data type for mixed precision conv2d.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.contrib_dense_pack(data, weight, weight_layout='NC', units=None, out_dtype='')#
Dense operator. Applies a linear transformation with packed weight
\[\]Y = X * W^T
Parameters#
- datatvm.relay.Expr
The input data to the operator, of shape (batch, units_in).
- weighttvm.relay.Expr
The transformed weight expressions, 3-D matrix, of shape (units // pack_weight_tile, units_in, pack_weight_tile).
- weight_layout: str
The layout of weight, such as “NC” or “NC8n”.
- unitsint, optional
Number of hidden units of the dense transformation.
- out_dtypestr, optional
Specifies the output data type for mixed precision dense.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.contrib_depthwise_conv2d_nchwc(data, kernel, strides=(1, 1), padding=(0, 0), dilation=(1, 1), groups=1, channels=None, kernel_size=None, data_layout='NCHW8c', kernel_layout='OIHW', out_layout='', out_dtype='')#
Variant of 2D depthwise convolution.
This operator takes the weight as the depthwise convolution kernel and depthwise convolves it with data to produce an output, following a specialized NCHWc data layout.
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- kerneltvm.relay.Expr
The kernel expressions.
- stridestuple of int, optional
The strides of convolution.
- paddingtuple of int, optional
The padding of convolution on both sides of inputs before convolution.
- dilationtuple of int, optional
Specifies the dilation rate to be used for dilated convolution.
- groupsint, optional
Number of groups for grouped convolution.
- channelsint, optional
Number of output channels of this convolution.
- kernel_sizetuple of int, optional
The spatial of the convolution kernel.
- data_layoutstr, optional
Layout of the input.
- kernel_layoutstr, optional
Layout of the weight.
- out_layoutstr, optional
Layout of the output, by default, out_layout is the same as data_layout
- out_dtypestr, optional
Specifies the output data type for mixed precision conv2d.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.conv1d(data, weight, strides=1, padding=0, dilation=1, groups=1, channels=None, kernel_size=None, data_layout='NCW', kernel_layout='OIW', out_layout='', out_dtype='')#
1D 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 NCW and kernel_layout is OIW, conv1d takes in a data Tensor with shape (batch_size, in_channels, width), and a weight Tensor with shape (channels, in_channels, kernel_size) to produce an output Tensor with the following rule:
\[\mbox{out}[b, c, w] = \sum_{dw, k} \mbox{data}[b, k, \mbox{strides}[0] * w + dw] * \mbox{weight}[c, k, dw]\]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 (NCW for data and OIW for weight), perform the computation, then convert to the out_layout.
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- weighttvm.relay.Expr
The weight expressions.
- stridesOptional[int, Tuple[int]]
The strides of convolution.
- paddingOptional[int, Tuple[int]]
The padding of convolution on both sides of the input before convolution.
- dilationOptional[int, Tuple[int]]
Specifies the dilation rate to be used for dilated convolution.
- groupsOptional[int]
Currently unused for 1D convolution.
- channelsOptional[int]
Number of output channels of this convolution.
- kernel_sizeOptional[int, Tuple[int]]
The spatial dimension of the convolution kernel.
- data_layoutOptional[str]
Layout of the input.
- kernel_layoutOptional[str]
Layout of the weight.
- out_layoutOptional[str]
Layout of the output, by default, out_layout is the same as data_layout
- out_dtypeOptional[str]
Specifies the output data type for mixed precision conv2d.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.conv1d_transpose(data, weight, strides=(1,), padding=(0,), dilation=(1,), groups=1, channels=None, kernel_size=None, data_layout='NCW', kernel_layout='IOW', out_layout='', output_padding=(0,), out_dtype='')#
One dimensional transposed convolution operator.
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- weighttvm.relay.Expr
The weight expressions.
- stridesTuple[int], optional
The strides of convolution.
- paddingTuple[int], optional
The padding of convolution on both sides of inputs.
- dilationTuple[int], optional
Specifies the dilation rate to be used for dilated convolution.
- channelsint, optional
Number of output channels of this convolution.
- kernel_sizetuple of int, optional
The spatial of the convolution kernel.
- groupsint, optional
Number of groups for grouped convolution.
- data_layoutstr, optional
Layout of the input.
- kernel_layoutstr, optional
Layout of the weight.
- out_layoutOptional[str]
Layout of the output, by default, out_layout is the same as data_layout
- output_paddingTuple[int], optional
Used to disambiguate the output shape.
- out_dtypestr, optional
Specifies the output data type for mixed precision conv2d.
Returns#
- resulttvm.relay.Expr
The computed result.
- 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#
- datatvm.relay.Expr
The input data to the operator.
- weighttvm.relay.Expr
The weight expressions.
- stridesOptional[int, Tuple[int]]
The strides of convolution.
- paddingOptional[int, Tuple[int]]
The padding of convolution on both sides of inputs before convolution.
- dilationOptional[int, Tuple[int]]
Specifies the dilation rate to be used for dilated convolution.
- groupsOptional[int]
Number of groups for grouped convolution.
- channelsOptional[int]
Number of output channels of this convolution.
- kernel_sizeOptional[int, Tuple[int]]
The spatial of the convolution kernel.
- data_layoutOptional[str]
Layout of the input.
- kernel_layoutOptional[str]
Layout of the weight.
- out_layoutOptional[str]
Layout of the output, by default, out_layout is the same as data_layout
- out_dtypeOptional[str]
Specifies the output data type for mixed precision conv2d.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.conv2d_backward_weight(grad, data, strides=(1, 1), padding=(0, 0), dilation=(1, 1), groups=1, channels=None, kernel_size=None, grad_layout='NCHW', data_layout='NCHW', kernel_layout='OIHW', out_dtype='')#
The gradient of conv2d with respect to weight.
This operator takes the output gradient grad and convolves it with data as the convolution kernel, to produce the gradient with respect to weight.
Note that the parameter kernel_size is the spatial size of the corresponding forward convolution kernel, not that of data. grad_layout and kernel_layout are the layouts of grad and the weight gradient respectively.
Other parameters are the same as the conv2d op. See its documentation for more details.
- 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='IOHW', out_layout='', output_padding=(0, 0), out_dtype='')#
Two dimensional transposed convolution operator.
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- weighttvm.relay.Expr
The weight expressions.
- stridesTuple[int], optional
The strides of convolution.
- paddingTuple[int], optional
The padding of convolution on both sides of inputs.
- dilationTuple[int], optional
Specifies the dilation rate to be used for dilated convolution.
- channelsint, optional
Number of output channels of this convolution.
- kernel_sizetuple of int, optional
The spatial of the convolution kernel.
- groupsint, optional
Number of groups for grouped convolution.
- data_layoutstr, optional
Layout of the input.
- kernel_layoutstr, optional
Layout of the weight.
- out_layoutOptional[str]
Layout of the output, by default, out_layout is the same as data_layout
- output_paddingTuple[int], optional
Used to disambiguate the output shape.
- out_dtypestr, optional
Specifies the output data type for mixed precision conv2d.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.conv3d(data, weight, strides=(1, 1, 1), padding=(0, 0, 0), dilation=(1, 1, 1), groups=1, channels=None, kernel_size=None, data_layout='NCDHW', kernel_layout='OIDHW', out_layout='', out_dtype='')#
3D 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 NCDHW and kernel_layout is OIDHW, conv3d takes in a data Tensor with shape (batch_size, in_channels, depth, height, width), and a weight Tensor with shape (channels, in_channels, kernel_size[0], kernel_size[1], kernel_size[2]) to produce an output Tensor with the following rule:
\[\mbox{out}[b, c, z, y, x] = \sum_{dz, dy, dx, k} \mbox{data}[b, k, \mbox{strides}[0] * z + dz, \mbox{strides}[1] * y + dy, \mbox{strides}[2] * x + dx] * \mbox{weight}[c, k, dz, dy, dx]\]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 (NCDHW for data and OIDHW for weight), perform the computation, then convert to the out_layout.
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- weighttvm.relay.Expr
The weight expressions.
- stridesOptional[Tuple[int]]
The strides of convolution.
- paddingOptional[int, Tuple[int]]
The padding of convolution on both sides of inputs before convolution.
- dilationOptional[int, Tuple[int]]
Specifies the dilation rate to be used for dilated convolution.
- groupsOptional[int]
Number of groups for grouped convolution.
- channelsOptional[int]
Number of output channels of this convolution.
- kernel_sizeOptional[int, Tuple[int]]
The spatial of the convolution kernel.
- data_layoutOptional[str]
Layout of the input.
- kernel_layoutOptional[str]
Layout of the weight.
- out_layoutOptional[str]
Layout of the output, by default, out_layout is the same as data_layout
- out_dtypeOptional[str]
Specifies the output data type for mixed precision conv2d.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.conv3d_transpose(data, weight, strides=(1, 1, 1), padding=(0, 0, 0), dilation=(1, 1, 1), groups=1, channels=None, kernel_size=None, data_layout='NCDHW', kernel_layout='IODHW', out_layout='', output_padding=(0, 0, 0), out_dtype='')#
3D transpose convolution.
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- weighttvm.relay.Expr
The weight expressions.
- stridesOptional[Tuple[int]]
The strides of convolution.
- paddingOptional[int, Tuple[int]]
The padding of convolution on both sides of inputs before convolution.
- dilationOptional[int, Tuple[int]]
Specifies the dilation rate to be used for dilated convolution.
- groupsOptional[int]
Number of groups for grouped convolution.
- channelsOptional[int]
Number of output channels of this convolution.
- kernel_sizeOptional[int, Tuple[int]]
The spatial of the convolution kernel.
- data_layoutOptional[str]
Layout of the input.
- kernel_layoutOptional[str]
Layout of the weight.
- out_layoutOptional[str]
Layout of the output, by default, out_layout is the same as data_layout
- out_dtypeOptional[str]
Specifies the output data type for mixed precision conv3d.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.correlation(data1, data2, kernel_size, max_displacement, stride1, stride2, padding, is_multiply, layout)#
Applies correlation to inputs.
The correlation layer performs multiplicative patch comparisons between two feature maps. Given two multi-channel feature maps \(f_{1}, f_{2}\), with \(w\), \(h\), and \(c\) being their width, height, and number of channels, the correlation layer lets the network compare each patch from \(f_{1}\) with each patch from \(f_{2}\).
For now we consider only a single comparison of two patches. The ‘correlation’ of two patches centered at \(x_{1}\) in the first map and \(x_{2}\) in the second map is then defined as:
\[c(x_{1}, x_{2}) = \sum_{o \in [-k,k] \times [-k,k]} <f_{1}(x_{1} + o), f_{2}(x_{2} + o)>\]for a square patch of size \(K:=2k+1\).
Note that the equation above is identical to one step of a convolution in neural networks, but instead of convolving data with a filter, it convolves data with other data. For this reason, it has no training weights.
Computing \(c(x_{1}, x_{2})\) involves \(c * K^{2}\) multiplications. Comparing all patch combinations involves \(w^{2}*h^{2}\) such computations.
Given a maximum displacement \(d\), for each location \(x_{1}\) it computes correlations \(c(x_{1}, x_{2})\) only in a neighborhood of size \(D:=2d+1\), by limiting the range of \(x_{2}\). We use strides \(s_{1}, s_{2}\), to quantize \(x_{1}\) globally and to quantize \(x_{2}\) within the neighborhood centered around \(x_{1}\).
The final output is defined by the following expression:
\[out[n, q, i, j] = c(x_{i, j}, x_{q})\]where \(i\) and \(j\) enumerate spatial locations in \(f_{1}\), and \(q\) denotes the \(q^{th}\) neighborhood of \(x_{i,j}\).
Parameters#
- data1tvm.te.Tensor
4-D with shape [batch, channel, height, width]
- data2tvm.te.Tensor
4-D with shape [batch, channel, height, width]
- kernel_size: int
Kernel size for correlation, must be an odd number
- max_displacement: int
Max displacement of Correlation
- stride1: int
Stride for data1
- stride2: int
Stride for data2 within the neightborhood centered around data1
- paddingint or a list/tuple of 2 or 4 ints
Padding size, or [pad_height, pad_width] for 2 ints, or [pad_top, pad_left, pad_bottom, pad_right] for 4 ints
- is_multiply: bool
operation type is either multiplication or substraction
- layout: str
layout of data1, data2 and the output
Returns#
- Outputtvm.te.Tensor
4-D with shape [batch, out_channel, out_height, out_width]
- tvm.relay.nn.cross_entropy(predictions, targets)#
CrossEntropy without logits.
Parameters#
- predictionstvm.relay.Expr
The predictions.
- targetstvm.relay.Expr
The targets.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.cross_entropy_with_logits(predictions, targets)#
CrossEntropy with logits.
Parameters#
- predictionstvm.relay.Expr
The predictions.
- targetstvm.relay.Expr
The targets.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.deformable_conv2d(data, offset, weight, strides=(1, 1), padding=(0, 0), dilation=(1, 1), deformable_groups=1, groups=1, channels=None, kernel_size=None, data_layout='NCHW', kernel_layout='OIHW', out_layout='', out_dtype='')#
Deformable 2d convolution.
The deformable convolution operation is described in https://arxiv.org/abs/1703.06211
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- offsettvm.relay.Expr
The offset expressions.
- weighttvm.relay.Expr
The weight expressions.
- stridestuple of int, optional
The strides of convolution.
- paddingtuple of int, optional
The padding of convolution on both sides of inputs before convolution.
- dilationtuple of int, optional
Specifies the dilation rate to be used for dilated convolution.
- deformable_groupsint, optional
Number of deformable groups.
- groupsint, optional
Number of groups for grouped convolution.
- channelsint, optional
Number of output channels of this convolution.
- kernel_sizetuple of int, optional
The spatial of the convolution kernel.
- data_layoutstr, optional
Layout of the input.
- kernel_layoutstr, optional
Layout of the weight.
- out_layoutstr, optional
Layout of the output, by default, out_layout is the same as data_layout
- out_dtypestr, optional
Specifies the output data type for mixed precision conv2d.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.dense(data, weight, units=None, out_dtype='')#
Dense operator. Applies a linear transformation
\[\]Y = X * W^T
Parameters#
- datatvm.relay.Expr
The input data to the operator, of shape (d_1, d_2, …, d_n, units_in).
- weighttvm.relay.Expr
The weight expressions, 2-D matrix, of shape (units, units_in).
- unitsint, optional
Number of hidden units of the dense transformation.
- out_dtypestr, optional
Specifies the output data type for mixed precision dense, of shape (d_1, d_2, …, d_n, units).
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.depth_to_space(data, block_size, layout='NCHW', mode='DCR')#
Convert channels into spatial blocks.
Parameters#
- datatvm.relay.Expr
Input data with channels divisible by block_size**2
- block_sizeint
Size of blocks to convert channels into.
- layoutstring
One of NCHW or NHWC, indicates channel axis.
- modestring
One of DCR or CDR, indicates which order channels are accessed in.
Returns#
- resulttvm.relay.Expr
- Tensor with shape [in_batch, in_channel / block_size * block_size,
in_height * block_size, in_width * block_size]
- tvm.relay.nn.dilate(data, strides, dilation_value=0.0)#
Dilate data with given dilation value (0 by default).
Parameters#
- datatvm.relay.Expr
n-D, can be any layout.
- stridestuple of <int>
Dilation stride on each dimension, 1 means no dilation.
- dilation_valueint/float, optional
Value used to dilate the input.
Returns#
- Outputtvm.relay.Expr
The computed result
- 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 by1/(1-p)
to keep the expected sum of the input unchanged.Parameters#
- datatvm.relay.Expr
The input data to the operator.
- ratefloat, optional (default=0.5)
The probability for an element to be reset to 0.
Returns#
- resulttvm.relay.Expr
The result of dropout
- tvm.relay.nn.dropout_raw(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 by1/(1-p)
to keep the expected sum of the input unchanged.Parameters#
- datatvm.relay.Expr
The input data to the operator.
- ratefloat, optional (default=0.5)
The probability for an element to be reset to 0.
Returns#
- resulttvm.relay.Expr
The result of dropout
- tvm.relay.nn.fast_softmax(data, axis=-1)#
Computes softmax. Use approximation to compute exponent for faster speed.
\[\text{softmax}(x)_i = \frac{exp(x_i)}{\sum_j exp(x_j)}\]备注
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#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.fifo_buffer(data, buffer, axis)#
FIFO buffer to enable computation reuse in CNNs with sliding indow input
Compute equivalent of
concat(buffer, data, axis=axis) .slice_axis(axis=axis, begin=data.shape[axis], end=data.shape[axis]+buffer.shape[axis])
Useful for
Encoding explicit re-use of computation in convolution ops operated on a sliding window input
Implementing a FIFO queue to cache intermediate results, e.g. as in Fast WaveNet.
Parameters#
- datatvm.relay.Expr
The input data
- buffertvm.relay.Expr
Previous value of the FIFO buffer
- axisint
Specify which axis should be used for buffering
Returns#
- resulttvm.relay.Expr
Updated value for the buffer
- tvm.relay.nn.get_pad_tuple1d(padding)#
Common code to get the 1 dimensional pad option Parameters ———- padding : Union[int, Tuple[int, …]]
Padding size
Returns#
- pad_leftint
Padding size on left
- pad_rightint
Padding size on right.
- tvm.relay.nn.get_pad_tuple2d(padding)#
Common code to get the pad option Parameters ———- padding : Union[int, Tuple[int, …]]
Padding size
Returns#
- pad_topint
Padding size on top
- pad_leftint
Padding size on left
- pad_downint
Padding size on down.
- pad_rightint
Padding size on right.
- tvm.relay.nn.get_pad_tuple3d(padding)#
Common code to get the pad option Parameters ———- padding : Union[int, Tuple[int, …]]
Padding size
Returns#
- pad_frontint
Padding size on front
- pad_topint
Padding size on top
- pad_leftint
Padding size on left
- pad_backint
Padding size on back
- pad_downint
Padding size on down.
- pad_rightint
Padding size on right.
- tvm.relay.nn.global_avg_pool1d(data, layout='NCW', out_layout='')#
1D global average pooling operator.
This operator takes data as input and does 1D average value calculation across each window represented by W.
In the default case, where the data_layout is NCW a data Tensor with shape (batch_size, in_channels, width), to produce an output Tensor with the following rule:
with data of shape (b, c, w)
\[\mbox{out}(b, c, 1) = \frac{1}{w} \sum_{n=0}^{w-1} \mbox{data}(b, c, n)\]Parameters#
- datatvm.relay.Expr
The input data to the operator.
- layoutstr, optional
Layout of the input.
- out_layoutstr, optional
Layout of the output.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.global_avg_pool2d(data, layout='NCHW', out_layout='')#
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#
- datatvm.relay.Expr
The input data to the operator.
- layoutstr, optional
Layout of the input.
- out_layoutOptional[str]
Layout of the output
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.global_avg_pool3d(data, layout='NCDHW', out_layout='')#
3D global average pooling operator.
This operator takes data as input and does 3D average value calculation across each window represented by DxWxH.
In the default case, where the data_layout is NCDHW a data Tensor with shape (batch_size, in_channels, depth, height, width), to produce an output Tensor with the following rule:
with data of shape (b, c, d, h, w)
\[\mbox{out}(b, c, 1, 1, 1) = \frac{1}{d * h * w} \sum_{l=0}^{d-1} \sum_{m=0}^{h-1} \sum_{n=0}^{w-1} \mbox{data}(b, c, l, m, n)\]Parameters#
- datatvm.relay.Expr
The input data to the operator.
- layoutstr, optional
Layout of the input.
- out_layoutstr, optional
Layout of the output.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.global_max_pool1d(data, layout='NCW', out_layout='')#
1D global maximum pooling operator.
This operator takes data as input and does 1D max value calculation across each window represented by W.
In the default case, where the data_layout is NCW a data Tensor with shape (batch_size, in_channels, width), to produce an output Tensor with the following rule:
with data of shape (b, c, w) .. math:
\mbox{out}(b, c, 1) = \max_{n=0, \ldots, w} \mbox{data}(b, c, n)
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- layoutstr, optional
Layout of the input.
- out_layoutstr, optional
Layout of the output.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.global_max_pool2d(data, layout='NCHW', out_layout='')#
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#
- datatvm.relay.Expr
The input data to the operator.
- layoutstr, optional
Layout of the input.
- out_layoutOptional[str]
Layout of the output
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.global_max_pool3d(data, layout='NCDHW', out_layout='')#
3D global maximum pooling operator.
This operator takes data as input and does 3D max value calculation across each window represented by DxWxH.
In the default case, where the data_layout is NCDHW a data Tensor with shape (batch_size, in_channels, depth, height, width), to produce an output Tensor with the following rule:
with data of shape (b, c, d, h, w) .. math:
\mbox{out}(b, c, 1, 1, 1) = \max_{l=0, \ldots, d}, \max_{m=0, \ldots, h}, \max_{n=0, \ldots, w} \mbox{data}(b, c, l, m, n)
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- layoutstr, optional
Layout of the input.
- out_layoutstr, optional
Layout of the output.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.group_norm(data, gamma, beta, num_groups, axis=1, epsilon=1e-05, center=True, scale=True)#
Group normalization normalizes over group of channels for each training examples. We can say that, Group Norm is in between Instance Norm and Layer Norm. When we put all the channels into a single group, group normalization becomes Layer normalization. And, when we put each channel into different groups it becomes Instance normalization
https://arxiv.org/pdf/1803.08494.pdf
Applies group normalization to the n-dimensional input array by seperating the input channels into ‘num_groups’ groups, each containing ‘num_channels / num_groups’ channels. The mean and standard-deviation are calculated separately over the each group. gamma and beta are learnable per-channel affine transform parameter vectors of size num_channels.
\[out = \frac{data - mean(data, axis)}{\sqrt{var(data, axis)+\epsilon}} * gamma + beta\]Unlike batch normalization, the mean and var are computed along a group of channels.
If the input has size k on axis 1, then both gamma and beta have shape (k,).
备注
This operator can be optimized away for inference.
Parameters#
- datatvm.relay.Expr
Input to which group_norm will be applied.
- gammatvm.relay.Expr
The gamma scale factor.
- betatvm.relay.Expr
The beta offset factor.
- num_groupsint
The number of groups to separate the channels into.
- axisint, optional, default=1
The axis of the channels.
- epsilondouble, optional, default=1e-5
Small float added to variance to avoid dividing by zero.
- centerboolean, optional, default=True
If True, add offset of beta to normalized tensor, If False, beta is ignored.
- scaleboolean, optional, default=True
If True, multiply by gamma. If False, gamma is not used.
Returns#
- resulttvm.relay.Expr
The normalized data.
- tvm.relay.nn.instance_norm(data, gamma, beta, axis=1, epsilon=1e-05, center=True, scale=True)#
Instance Normalization (Ulyanov and et al., 2016) Applies instance normalization to the n-dimensional input array.
\[out = \frac{data - mean(data)}{\sqrt{var(data)+\epsilon}} * gamma + beta\]The instance normalization is similar to batch normalization, but unlike batch normalization, the mean and var are calculated per-dimension separately for each object(instance) in a mini-batch, not over a batch. And the same normalization is applied both at test and train time.
Assume the input has size k on axis 1, then both
gamma
andbeta
have shape (k,).The parameter
axis
specifies which axis of the input shape denotes the ‘channel’. The default is 1. Specifying -1 sets the channel axis to be the last item in the input shape.备注
This operator can be optimized away for inference.
Parameters#
- datatvm.relay.Expr
Input to which instance_norm will be applied.
- gammatvm.relay.Expr
The gamma scale factor.
- betatvm.relay.Expr
The beta offset factor.
- axisint, optional, default=1
Specify along which shape axis the channel is specified.
- epsilondouble, optional, default=1e-5
Small float added to variance to avoid dividing by zero.
- centerboolean, optional, default=True
If True, add offset of beta to normalized tensor, If False, beta is ignored.
- scaleboolean, optional, default=True
If True, multiply by gamma. If False, gamma is not used.
Returns#
- resulttvm.relay.Expr
The normalized data.
- 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#
- datatvm.relay.Expr
The input data to the operator.
- epsfloat
epsilon value
- axislist of int, optional
axis over the normalization applied
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.layer_norm(data, gamma, beta, axis=-1, epsilon=1e-05, center=True, scale=True)#
Layer normalization (Lei Ba and et al., 2016). Applies layer normalization to the n-dimensional input array. This operator takes an n-dimensional input array and normalizes the input using the given axis:
\[out = \frac{data - mean(data, axis)}{\sqrt{var(data, axis)+\epsilon}} * gamma + beta\]Unlike batch normalization, the mean and var are computed along the channel dimension.
Assume the input has size k on axis 1, then both gamma and beta have shape (k,).
备注
This operator can be optimized away for inference.
Parameters#
- datatvm.relay.Expr
Input to which layer_norm will be applied.
- gammatvm.relay.Expr
The gamma scale factor.
- betatvm.relay.Expr
The beta offset factor.
- axisint, optional, default=-1
The axis that should be normalized, typically the axis of the channels.
- epsilondouble, optional, default=1e-5
Small float added to variance to avoid dividing by zero.
- centerboolean, optional, default=True
If True, add offset of beta to normalized tensor, If False, beta is ignored.
- scaleboolean, optional, default=True
If True, multiply by gamma. If False, gamma is not used.
Returns#
- resulttvm.relay.Expr
The normalized data.
- tvm.relay.nn.leaky_relu(data, alpha=0.01)#
This operator takes data as input and does Leaky version of a Rectified Linear Unit.
\[`y = x > 0 ? x : alpha * x`\]Parameters#
- datatvm.relay.Expr
The input data to the operator.
- alphafloat
Slope coefficient for the negative half axis.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.log_softmax(data, axis=-1)#
Computes log softmax.
\[\text{log_softmax}(x)_i = \log \frac{exp(x_i)}{\sum_j exp(x_j)}\]备注
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 log softmax
Returns#
- resulttvm.relay.Expr
The computed result.
- 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#
- datatvm.relay.Expr
The input data to the operator.
- sizeint, optional
The size of the local region to be considered for normalization.
- axisint, optional
Input data layout channel axis. Default value is 1 for NCHW format
- biasfloat, optional
The offset parameter to avoid dividing by 0.
- alphafloat, optional
The scaling parameter.
- betafloat, optional
The exponent parameter.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.matmul(tensor_a, tensor_b, units=None, out_dtype='', transpose_a=False, transpose_b=False)#
Matmul operator. Applies a linear transformation. The A & B can be transposed.
\[`C = A * B`\]Parameters#
- datatvm.relay.Expr
The first input of the operator, of shape (d_1, d_2, …, d_n, units_in) or (d_1, d_2, …, units_in, d_n).
- weighttvm.relay.Expr
The second input expressions, 2-D matrix, of shape (units_in, units) or (units, units_in).
- unitsOptional[int]
Number of hidden units of the matmul transformation.
- out_dtypeOptional[str]
Specifies the output data type for mixed precision matmul, of shape (d_1, d_2, …, d_n, units).
- transpose_aOptional[bool] = False
Whether the data tensor is in transposed format.
- transpose_bOptional[bool] = False
Whether the weight tensor is in transposed format.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.max_pool1d(data, pool_size=(1,), strides=(1,), dilation=(1,), padding=(0,), layout='NCW', out_layout='', ceil_mode=False)#
1D maximum pooling operator.
This operator takes data as input and does 1D max value calculation with in pool_size sized window by striding defined by stride.
In the default case, where the data_layout is NCW a data Tensor with shape (batch_size, channels, width), to produce an output Tensor.
The 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#
- datatvm.relay.Expr
The input data to the operator.
- pool_sizeint or tuple of int, optional
The size of window for pooling.
- stridesint or tuple of int, optional
The strides of pooling.
- dilationint or tuple of int, optional
The dilation of pooling.
- paddingint or tuple of int, optional
The padding for pooling.
- layoutstr, optional
Layout of the input.
- out_layoutOptional[str]
Layout of the output
- ceil_modebool, optional
To enable or disable ceil while pooling.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.max_pool2d(data, pool_size=(1, 1), strides=(1, 1), dilation=(1, 1), padding=(0, 0), layout='NCHW', out_layout='', 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#
- datatvm.relay.Expr
The input data to the operator.
- pool_sizeint or tuple of int, optional
The size of window for pooling.
- stridestuple of int, optional
The strides of pooling.
- dilationint or tuple of int, optional
The dilation of pooling.
- paddingtuple of int, optional
The padding for pooling.
- layoutstr, optional
Layout of the input.
- out_layoutOptional[str]
Layout of the output
- ceil_modebool, optional
To enable or disable ceil while pooling.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.max_pool2d_grad(out_grad, data, pool_size=(1, 1), strides=(1, 1), padding=(0, 0), layout='NCHW', out_layout='', ceil_mode=False)#
Gradient of 2D maximum pooling operator.
This operator takes out_grad and data as input and calculates gradient of max_pool2d.
Parameters#
- out_gradtvm.relay.Expr
The output gradient
- datatvm.relay.Expr
The input data to the operator.
- pool_sizeint or tuple of int, optional
The size of window for pooling.
- stridestuple of int, optional
The strides of pooling.
- paddingtuple of int, optional
The padding for pooling.
- layoutstr, optional
Layout of the input.
- out_layoutOptional[str]
Layout of the output
- ceil_modebool, optional
To enable or disable ceil while pooling.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.max_pool3d(data, pool_size=(1, 1, 1), strides=(1, 1, 1), dilation=(1, 1, 1), padding=(0, 0, 0), layout='NCDHW', out_layout='', ceil_mode=False)#
3D maximum pooling operator.
This operator takes data as input and does 3D max value calculation with in pool_size sized window by striding defined by stride.
In the default case, where the data_layout is NCDHW a data Tensor with shape (batch_size, channels, depth, height, width), to produce an output Tensor.
The 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#
- datatvm.relay.Expr
The input data to the operator.
- pool_sizeint or tuple of int, optional
The size of window for pooling.
- stridestuple of int, optional
The strides of pooling.
- dilationint or tuple of int, optional
The dilation of pooling.
- paddingtuple of int, optional
The padding for pooling.
- layoutstr, optional
Layout of the input.
- out_layoutOptional[str]
Layout of the output
- ceil_modebool, optional
To enable or disable ceil while pooling.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.mirror_pad(data, pad_width, mode='SYMMETRIC')#
MirrorPadding
This operator takes in a tensor and pads each axis by the specified widths using mirroring of the border pixels.
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))
- mode: string, optional, default=’SYMMETRIC’
What type of mirroring to use, must be SYMMETRIC or REFLECT.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.nll_loss(predictions, targets, weights, reduction='mean', ignore_index=-100)#
Negative log likelihood loss.
- output{n, i_1, i_2, …, i_k} = -p * w
- where t = target{n, i_1, i_2, …, i_k}
p = predictions{n, t, i_1, i_2, i_k} w = weights{n, i_1, i_2, …, i_k} if t != ignore_index else 0
result = reduction(output)
Parameters#
- predictionstvm.relay.Expr
The predictions.
- targetstvm.relay.Expr
The target value of each prediction.
- weightstvm.relay.Expr
The weight of each target value.
- reductionstring
The reduction method to apply to the output. Possible values are “mean”, “sum” and “none”.
- ignore_indexint
The target value to ignore.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.pad(data, pad_width, pad_value=0, pad_mode='constant')#
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>>, or tvm.relay.Expr, 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, or tvm.relay.Expr, optional, default=0
The value used for padding
- pad_mode: ‘constant’, ‘edge’, ‘reflect’
‘constant’ pads with constant_value pad_value ‘edge’ pads using the edge values of the input array ‘reflect’ pads by reflecting values with respect to the edge
Returns#
- resulttvm.relay.Expr
The computed result.
- 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#
- datatvm.relay.Expr
The input data to the operator.
- alphatvm.relay.Expr
Slope coefficient for the negative half axis.
- axisint, optional
Specify which shape axis the channel is specified.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.relu(data)#
Rectified linear unit.
\[out = max(x, 0)\]Parameters#
- datatvm.relay.Expr
The input data
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.softmax(data, axis=-1)#
Computes softmax.
\[\text{softmax}(x)_i = \frac{exp(x_i)}{\sum_j exp(x_j)}\]备注
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#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.space_to_batch_nd(data, block_shape, paddings, pad_value=0)#
Divide spatial dimensions of the data into a grid of blocks and interleave them into batch dim.
Parameters#
- datatvm.te.Tensor
N-D with shape [batch, spatial_shape, remaining_shape]
- block_shaperelay.Expr
1-D of size [M] where M is number of spatial dims, specifies block size for each spatial dimension.
- paddingsrelay.Expr
2-D of shape [M, 2] where M is number of spatial dims, specifies [before, after] paddings for each spatial dimension.
- pad_valuefloat, or relay.Expr, optional, default=0
The value used for padding.
Returns#
- resultrelay.Expr
N-D Tensor with shape [in_batch * prod(block_shape), padded_data[1] / block_shape[0], …, padded_data[M] / block_shape[M-1], remaining_shape]
- tvm.relay.nn.space_to_depth(data, block_size, layout='NCHW')#
Convert spatial blocks into channels.
Parameters#
- datatvm.relay.Expr
Input data with spatial dimensions divisible by block_size
- block_sizeint
Size of blocks to decompose into channels.
- layoutstring
One of NCHW or NHWC, indicates channel axis.
Returns#
- resulttvm.relay.Expr
- Tensor with shape [in_batch, in_channel * block_size * block_size,
in_height / block_size, in_width / block_size]
- tvm.relay.nn.sparse_add(dense_mat, sparse_mat)#
Computes the matrix addition of dense_mat and sparse_mat, where dense_mat is a dense matrix and sparse_mat is a sparse (CSR) namedtuple with fields data, indices, and indptr.
\[\mbox{sparse_add}(dense_mat, sparse_mat)[m, n] = \mbox{add}(\mbox{as_dense}(S), (D))[m, n]\]where as_dense returns dense equivalent of the given S(sparse matrix) while performing addition with given D(dense matrix).
Parameters#
- dense_mattvm.relay.Expr
The input dense matrix for the matrix addition
- sparse_matUnion[namedtuple, Tuple[ndarray, ndarray, ndarray]].
The input sparse matrix(CSR) for the matrix addition.
Returns#
- result: tvm.relay.Expr
The computed result.
Examples#
dense_data = [[ 3., 4., 4. ] [ 4., 2., 5. ]] sparse_data = [4., 8.] sparse_indices =[0, 2] sparse_indptr =[0, 1, 2] output = relay.sparse_add(dense_data, sparse_data, sparse_indices, sparse_indptr) output = [[ 7., 4., 4. ] [ 4., 2., 13. ]]
- tvm.relay.nn.sparse_dense(dense_mat, sparse_mat, sparse_lhs=False)#
Computes the matrix multiplication of dense_mat and sparse_mat, where dense_mat is a dense matrix and sparse_mat is a sparse (either BSR or CSR) namedtuple with fields data, indices, and indptr.
- if sparse_lhs=False:
- \[\mbox{sparse_dense}(dense_mat, sparse_mat)[m, n] = \mbox{matmul}(D, \mbox{as_dense}(S)^T)[m, n]\]
- if sparse_lhs=True:
- \[\mbox{sparse_dense}(dense_mat, sparse_mat)[m, n] = \mbox{matmul}(\mbox{as_dense}(S), (D)^T)[m, n]\]
where as_dense returns dense equivalent of the given S(sparse matrix) while performing matmul with given D(dense matrix).
See https://docs.scipy.org/doc/scipy/reference/generated/scipy.sparse.csr_matrix.html and https://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.sparse.bsr_matrix.html for more detail on the sparse matrix representation.
Parameters#
- dense_mattvm.relay.Expr
The input dense matrix for the matrix multiplication
- sparse_matUnion[namedtuple, Tuple[ndarray, ndarray, ndarray]].
The input sparse matrix for the matrix multiplication.
- sparse_lhsbool, optional
Indicates whether lhs or rhs matrix is sparse. Default value is False.
Returns#
- result: tvm.relay.Expr
The computed result.
- tvm.relay.nn.sparse_transpose(x)#
Computes the fast matrix transpose of x, where x is a sparse tensor in CSR format (represented as a namedtuple with fields data, indices, and indptr).
** Currently only support Square Matrices **
\[\mbox{sparse_transpose}(x)[n, n] = (x^T)[n, n]\]Please refer to scipy/scipy for the algorithm implemented in this operator.
Parameters#
- xUnion[namedtuple, Tuple[ndarray, ndarray, ndarray]].
The sparse weight matrix for the fast matrix transpose.
Returns#
- resultrelay.Tuple([tvm.relay.Expr, tvm.relay.Expr, tvm.relay.Expr])
Tuple of output sparse tensor (same shape and format as input), i.e. if CSR then output is in ([data, indices, indptr]) form
- tvm.relay.nn.upsampling(data, scale_h=1, scale_w=1, layout='NCHW', method='nearest_neighbor', align_corners=False)#
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_h, w*scale_w)
method indicates the algorithm to be used while calculating the out value and method can be one of (“bilinear”, “nearest_neighbor”, “bicubic”)
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- scale_htvm.relay.Expr or int or float
The scale factor for height upsampling.
- scale_wtvm.relay.Expr or int or float
The scale factor for width upsampling.
- layoutstr, optional
Layout of the input.
- methodstr, optional
Scale method to used [nearest_neighbor, bilinear, bicubic].
- align_cornersbool, optional
Whether to keep corners in proper place.
Returns#
- resulttvm.relay.Expr
The computed result.
- tvm.relay.nn.upsampling3d(data, scale_d=1, scale_h=1, scale_w=1, layout='NCDHW', method='nearest_neighbor', coordinate_transformation_mode='half_pixel')#
3D Upsampling.
This operator takes data as input and does 3D scaling to the given scale factor. In the default case, where the data_layout is NCDHW with data of shape (n, c, d, h, w) out will have a shape (n, c, d*scale_d, h*scale_h, w*scale_w)
method indicates the algorithm to be used while calculating the out value and method can be one of (“trilinear”, “nearest_neighbor”)
Parameters#
- datatvm.relay.Expr
The input data to the operator.
- scale_dtvm.relay.Expr
The scale factor for depth upsampling.
- scale_htvm.relay.Expr
The scale factor for height upsampling.
- scale_wtvm.relay.Expr
The scale factor for width upsampling.
- layoutstr, optional
Layout of the input.
- methodstr, optional
Scale method to used [nearest_neighbor, trilinear].
- coordinate_transformation_mode: string, optional
Describes how to transform the coordinate in the resized tensor to the coordinate in the original tensor. Refer to the ONNX Resize operator specification for details. Available options are “half_pixel”, “align_corners” and “asymmetric”.
Returns#
- resulttvm.relay.Expr
The computed result.