tensorflow - LSTM implementation with peephole -
i have been reading papers lstm , checking implementations. there 1 point not clear me.
in of papers mentioned weight matrices cell gate vectors should diagonal(ex: alex page 5, 2013), haven't seen in implementation.
example :
1 2 example mila lab. 3
are these people implementing wrongly or missing something?
the tensorflow implementation use diagonal matrix, see here. note means in practice peepholes go cell itself, , you're doing elementwise vector multiplies.
Comments
Post a Comment