In the LPD implementation the bias terms of the convolution layers are turned off (bias=False in nn.Conv2d). However, in the original LPD paper they say that the convolution layers do have biasses. (Section IV.A In the equation after the line "where the k:th component is given by"). I also checked the original code, and it does not specify the bias, but the Tensorflow default is True.
Was it a deliberate choice to not follow the paper? Otherwise I think it's better to change the code so that bias=True for all convolution layers.
In the LPD implementation the bias terms of the convolution layers are turned off (bias=False in nn.Conv2d). However, in the original LPD paper they say that the convolution layers do have biasses. (Section IV.A In the equation after the line "where the k:th component is given by"). I also checked the original code, and it does not specify the bias, but the Tensorflow default is True.
Was it a deliberate choice to not follow the paper? Otherwise I think it's better to change the code so that bias=True for all convolution layers.