def __init__()

in src/hyperconv.py [0:0]

• 29 lines of code
• 2 McCabe index (conditional complexity)

    def __init__(self, input_size, ch_in, ch_out, kernel_size, dilation=1):
        '''
        HyperConv implements a temporal convolution that has different convolution weights for each time step.
        :param input_size: (int) dimension of the weight generating input variable
        :param ch_in: (int) number of input channels of the temporal convolution
        :param ch_out: (int) number of output channels of the temporal convolution
        :param kernel_size: (int) kernel size of the temporal convolution
        :param dilation: (int) dilation of the temporal convolution
        '''
        super().__init__()
        weight_regressor_hidden_size = 32
        self.ch_in = ch_in
        self.ch_out = ch_out
        self.kernel_size = kernel_size
        self.dilation = dilation
        self.weight_model = nn.Sequential(
            nn.Conv1d(input_size, weight_regressor_hidden_size, kernel_size=1),
            nn.ReLU(),
            nn.Conv1d(weight_regressor_hidden_size, ch_in * ch_out * kernel_size, kernel_size=1)
        )
        self.bias_model = nn.Sequential(
            nn.Conv1d(input_size, weight_regressor_hidden_size, kernel_size=1),
            nn.ReLU(),
            nn.Conv1d(weight_regressor_hidden_size, ch_out, kernel_size=1)
        )
        # initialize weights such that regressed weights are distributed in a suitable way for sine activations
        self.weight_model[0].weight.data.zero_()
        self.weight_model[0].bias.data.zero_()
        self.weight_model[-1].bias.data.uniform_(-np.sqrt(6.0/(self.ch_in*self.kernel_size)),
                                                 np.sqrt(6.0/(self.ch_in*self.kernel_size)))