MobileNetV3

class torch_ecg.models.MobileNetV3(in_channels: int, **config: CFG)[source]

Bases: Sequential, SizeMixin, CitationMixin

MobileNet V3.

MobileNet V3 [1] is an incremental improvement of MobileNet series. MobileNet V3 uses neural architecture search instead of hand-designed architectures to find the optimal network structure. MobileNet V3 has implementions in Torchvision [2], which serves as a reference for this implementation for ECG tasks.

Parameters:

  • in_channels (int) – Number of channels in the input signal tensor.

  • config (dict) –

    Other hyper-parameters of the Module, ref. corr. config file. Keyword arguments that must be set:

    • groups: int, number of groups in the convolutional layer(s) other than depthwise convolutions.

    • norm: bool or str or Module, normalization layer.

    • bias: bool, whether to use bias in the convolutional layer(s).

    • width_multiplier: float, multiplier of the number of output channels of the pointwise convolution.

    • stem: CFG, config of the stem block, with the following keys:

      • num_filters: int or Sequence[int], number of filters in the first convolutional layer(s).

      • filter_lengths: int or Sequence[int], filter lengths (kernel sizes) in the first convolutional layer(s).

      • subsample_lengths: int or Sequence[int], subsample lengths (strides) in the first convolutional layer(s).

    • inv_res: CFG, config of the inverted residual blocks, with the following keys:

      • in_channels: Sequence[int], number of input channels.

      • n_blocks: Sequence[int], number of inverted residual blocks.

      • expansions: sequence of floats or sequence of sequence of floats, expansion ratios of the inverted residual blocks.

      • filter_lengths: sequence of ints or sequence of sequence of ints, filter length of the depthwise convolution in the inverted residual blocks.

      • stride: sequence of ints or sequence of sequence of ints, optional, stride of the depthwise convolution in the inverted residual blocks, defaults to [2] + [1] * (n_blocks - 1).

      • groups: int, default 1, number of groups in the expansion and pointwise convolution in the inverted residual blocks.

      • dilation: sequence of ints or sequence of sequence of ints, optional, dilation of the depthwise convolution in the inverted residual blocks.

      • batch_norm: bool or str or nn.Module, default True, normalization layer to use, defaults to batch normalization.

      • activation: str or nn.Module or sequence of str or torch.nn.Module, activation function to use.

      • width_multiplier: float or sequence of floats, default 1.0, width multiplier of the inverted residual blocks.

      • out_channels: sequence of ints or sequence of Sequence[int], optional, number of output channels of the inverted residual blocks, defaults to 2 * in_channels.

      • attn: sequence of CFG or sequence of sequence of CFG, optional, config of attention layer to use, defaults to None.

    • exit_flow: CFG,

      config of the exit flow blocks, with the following keys:

      • num_filters: int or Sequence[int], number of filters in the final convolutional layer(s).

      • filter_lengths: int or Sequence[int], filter lengths (kernel sizes) in the final convolutional layer(s).

      • subsample_lengths: int or Sequence[int], subsample lengths (strides) in the final convolutional layer(s).

References

compute_output_shape(seq_len: int | None = None, batch_size: int | None = None) → Sequence[int | None][source]

Compute the output shape of the model.

Parameters:

  • seq_len (int, optional) – Length of the input tensors.

  • batch_size (int, optional) – Batch size of the input tensors.

Returns:

output_shape – The output shape of the module.

Return type:

sequence