nn.mm requires table argument of matrices multiplied. in case, 1 of matrices output of defined model (e.g. nn.sequential) , other constant matrix. how can inject constant nn's pipeline , should worried optimizer start changing if do?
i'm aware solve injection problem by:
- writing own
nn.module. seems heavy handed. - breaking model 2 parts , manually injecting constant. want model
nn.modulesubclass gets called:forward(input), allows consumers blissfully ignorant of existence of constant. - using
nn.paralleltable, expose constant model consumers. - using
nn.linearno bias , overwriting weights. i'm not sure how prevent optimizer performing update.
you can create nn.linear , override :accgradparameters no-op function
m = nn.linear(100,200) -- copy weights / bias m.weight / m.bias m.accgradparameters = function() end -- m constant multiplier thing 