DenseNet

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

Bases: Sequential, SizeMixin, CitationMixin

The core part of the SOTA model (framework) of CPSC2020.

DenseNet is originally proposed in [1], [2] (journal version). The original implementation is available at [3] and [4]. [5] is an unofficial implementation of DenseNet in PyTorch. Torchvision also provides an implementation of DenseNet [6].

DenseNet is not only successful in image classification, but also in various ECG-related tasks, and is the core part of the SOTA model (framework) of CPSC2020.

Parameters:

  • in_channels (int) – Number of features (channels) of the input.

  • config (dict) –

    Other hyper-parameters of the Module, ref. corresponding config file. Keyword arguments that must be set are as follows:

    • num_layers: sequence of int, number of building block layers of each dense (macro) block

    • init_num_filters: sequence of int, number of filters of the first convolutional layer

    • init_filter_length: sequence of int, filter length (kernel size) of the first convolutional layer

    • init_conv_stride: int, stride of the first convolutional layer

    • init_pool_size: int, pooling kernel size of the first pooling layer

    • init_pool_stride: int, pooling stride of the first pooling layer

    • growth_rates: int or sequence of int or sequence of sequences of int, growth rates of the building blocks, with granularity to the whole network, or to each dense (macro) block, or to each building block

    • filter_lengths: int or sequence of int or sequence of sequences of int, filter length(s) (kernel size(s)) of the convolutions, with granularity to the whole network, or to each macro block, or to each building block

    • subsample_lengths: int or sequence of int, subsampling length(s) (ratio(s)) of the transition blocks

    • compression: float, compression factor of the transition blocks

    • bn_size: int, bottleneck base width, used only when building block is DenseBottleNeck

    • dropouts: float or dict, dropout ratio of each building block

    • groups: int, connection pattern (of channels) of the inputs and outputs

    • block: dict, other parameters that can be set for the building blocks

    For a full list of configurable parameters, ref. corr. config file

Note

The difference of forward output of [5] from others, however [5] doesnot support dropout.

TODO

  1. For groups > 1, the concatenated output should be re-organized in the channel dimension?

  2. memory-efficient mode, i.e. storing the new_features in a shared memory instead of stacking in newly created Tensor after each mini-block.

References

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

Compute the output shape of the network.

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