Table Of Contents
Table Of Contents

gluonts.block.cnn module

class gluonts.block.cnn.CausalConv1D(channels: int, kernel_size: int, dilation: int = 1, activation: Optional[str] = 'relu', **kwargs)[source]

Bases: mxnet.gluon.block.HybridBlock

1D causal temporal convolution, where the term causal means that output[t] does not depend on input[t+1:]. Notice that Conv1D is not implemented in Gluon.

This is the basic structure used in Wavenet [ODZ+16] and Temporal Convolution Network [BKK18].

The output has the same shape as the input, while we always left-pad zeros.

Parameters:
  • channels – The dimensionality of the output space, i.e. the number of output channels (filters) in the convolution.
  • kernel_size – Specifies the dimensions of the convolution window.
  • dilation – Specifies the dilation rate to use for dilated convolution.
  • activation – Activation function to use. See Activation(). If you don’t specify anything, no activation is applied (ie. “linear” activation: a(x) = x).
hybrid_forward(F, data: Union[mxnet.ndarray.ndarray.NDArray, mxnet.symbol.symbol.Symbol]) → Union[mxnet.ndarray.ndarray.NDArray, mxnet.symbol.symbol.Symbol][source]

In Gluon’s conv1D implementation, input has dimension NCW where N is batch_size, C is channel, and W is time (sequence_length).

Parameters:data – Shape (batch_size, num_features, sequence_length)
Returns:causal conv1d output. Shape (batch_size, num_features, sequence_length)
Return type:Tensor
class gluonts.block.cnn.DilatedCausalGated(inner_channels: int, out_channels: int, kernel_size: Union[int, Tuple[int], List[int]], dilation: Union[int, Tuple[int], List[int]], **kwargs)[source]

Bases: mxnet.gluon.block.HybridBlock

1D convolution with Gated mechanism, see the Wavenet papers described above.

Parameters:
  • inner_channels – The dimensionality of the intermediate space
  • out_channels – The dimensionality of the output space
  • kernel_size – Specifies the dimensions of the convolution window.
  • dilation – Specifies the dilation rate to use for dilated convolution.
hybrid_forward(F, x: Union[mxnet.ndarray.ndarray.NDArray, mxnet.symbol.symbol.Symbol]) → Union[mxnet.ndarray.ndarray.NDArray, mxnet.symbol.symbol.Symbol][source]

Compute the 1D convolution with Gated mechanism.

Parameters:x – input features, shape (batch_size, num_features, sequence_length)
Returns:output, shape (batch_size, num_features, sequence_length)
Return type:Tensor
class gluonts.block.cnn.ResidualSequential(**kwargs)[source]

Bases: mxnet.gluon.nn.basic_layers.HybridSequential

Adding residual connection to each layer of the hybrid sequential blocks

hybrid_forward(F, x: Union[mxnet.ndarray.ndarray.NDArray, mxnet.symbol.symbol.Symbol]) → Union[mxnet.ndarray.ndarray.NDArray, mxnet.symbol.symbol.Symbol][source]
Parameters:
  • F – A module that can either refer to the Symbol API or the NDArray API in MXNet.
  • x – input tensor
Returns:

output of the ResidualSequential

Return type:

Tensor