Open dwf opened 8 years ago
One note: we use tied_biases
in Convolutional
to indicate biases tied across spatial locations. We should either change this to be more specific or else choose another name for channel-tied biases. I'm using one_bias_per_channel
in my private code for such things. one_bias_per_output
also works.
This is a bit tricky, but can probably be accomplished by turning off the Linear brick's bias and applying a separate Bias brick on reshaped input.