Custom connections in neural network layers

[jamesproud]

Here, also, is my implementation:

Gradient masking in register_backward_hook for custom connectivity - efficient implementation

@jonathanEphrath Here is my implementation which was inspired by @jukiewiczm’s code. class MaskedLinear(nn.Module): def __init__(self, in_dim, out_dim, indices_mask): """ :param in_features: number of input features :param out_features: number of output features :param indices_mask: list of two lists containing indices for dimensions 0 and 1, used to create the mask """ super(MaskedLinear, self).__init__() def backward_hook(grad): …

@uchida-takumi I would be interested in your thoughts on comparisons between the approaches.