ResNet1D

class selfeeg.models.zoo.ResNet1D(nb_classes: int, Chans: int, Samples: int, block: ~torch.nn.modules.module.Module = <class 'selfeeg.models.encoders.BasicBlock1'>, Layers: list of 4 int = [2, 2, 2, 2], inplane: int = 16, kernLength: int = 7, addConnection: bool = False, preBlock: ~torch.nn.modules.module.Module = None, postBlock: ~torch.nn.modules.module.Module = None, classifier: ~torch.nn.modules.module.Module = None, return_logits: bool = True, seed: int = None)[source]

Pytorch implementation of the Resnet model

This model adopts temporal convolutional layers, so conv2d layers with horizontal kernel. Implemented using as reference the following paper [res1] .

The expected input is a 3D tensor with size (Batch x Channels x Samples).

Parameters:

  • nb_classes (int) – The number of classes. If less than 2, a binary classification problem is considered (output dimensions will be [batch, 1] in this case).

  • Chans (int) – The number of EEG channels.

  • Samples (int) – The sample length. It will be used to calculate the embedding size (for head initialization).

  • block (nn.Module, optional) –

    An nn.Module defining the resnet block.

    Default: selfeeg.models.BasicBlock1

  • Layers (list of 4 int, optional) –

    A list of integers indicating the number of times the resnet block is repeated at each stage. It must be a list of length 4 with positive integers. Shorter lists are padded to 1 on the right. Only the first four elements of longer lists are considered. Zeros are changed to 1.

    Default = [2, 2, 2, 2]

  • inplane (int, optional) –

    The number of output filters.

    Default = 16

  • kernLength (int, optional) –

    The length of the temporal convolutional layer.

    Default = 7

  • addConnection (bool, optional) –

    Whether to add a connection from the start of the resnet part to the network head. If set to True the output of the following conv2d will be concatenate to the postblock output:

    1. nn.Conv2d(inplane, 2, kernel_size=(Chans, kernLength), stride=(1, int(kernLength//2)), padding=(0, 0), bias=False)

    Default = None

  • preBlock (nn.Module, optional) –

    A custom nn.Module to pass before entering the sequence of resnet blocks. If none is left, the following sequence is used:

    1. nn.conv2d(1, self.inplane, kernel_size=(1, kernLength), stride=(1, 2), padding=(0, kernLength//2), bias=False)

    2. nn.BatchNorm2d()

    3. nn.ReLU()

    Default = None

  • postBlock (nn.Module, optional) –

    A custom nn.Module to pass after the sequence of resnet blocks and before the network head. If none is left, the following sequence is used:

    1. nn.conv2d(1, self.inplane, kernel_size=(1, kernLength), bias=False)

    2. nn.BatchNorm2d()

    3. nn.ReLU()

    Default = None

  • classifier (nn.Module, optional) –

    A custom nn.Module defining the network head. If none is left, a single dense layer is used.

    Default = None

  • return_logits (bool, optional) –

    Whether to return the output as logit or probability. It is suggested to not use False as the pytorch crossentropy applies the softmax internally.

    Default = True

  • seed (int, optional) –

    A custom seed for model initialization. It must be a nonnegative number. If None is passed, no custom seed will be set

    Default = None

Note

The compatibility between each custom nn.Module given as argument has not been carefully checked. Errors may arise.

References

[res1]

Zheng et al., Task-oriented Self-supervised Learning for Anomaly Detection in Electroencephalography

Example

>>> import selfeeg.models
>>> import torch
>>> x = torch.randn(4,8,512)
>>> mdl = models.ResNet1D(4,8,512)
>>> out = mdl(x)
>>> print(out.shape) # shoud return torch.Size([4, 4])
>>> print(torch.isnan(out).sum()) # shoud return 0