search

Enny1991 / PLSTM

137 stars 32 forks source link

Manually set and constant tau, r_on and s #7

Closed jamessergeant closed 7 years ago

jamessergeant commented 7 years ago

As per the lip reading experiment in the original paper, have added an input argument to allow for manually setting constant (non-trainable) values of tau, r_on and s. Additionally, a hard gating is applied during r_on. #6

Enny1991 commented 7 years ago

Hey @jamessergeant, I think is a nice add-on to put non trainable parameters! But I'm not sure why can't we just set the trainable=False? Why would you set the initialization to constant? In this way all the units will have the same parameters. I would leave it initialize as it was before and just set the trainable to false. What do you think?

+Enea

jamessergeant commented 7 years ago

@Enny1991, in the paper they manually set the period (and I presumed r_on and s, will need to double check the paper) to constants to match the input frequencies of the audio and video streams. I could separate this functionality using another input arg.

Enny1991 commented 7 years ago

Yeah I would rather do that. Maybe have the choice between not train the parameters and set them to a precise value. Someone might want to not train but still have some variability in the different units.

jamessergeant commented 7 years ago

Sounds good. I'll modify tomorrow and update my fork.

jamessergeant commented 7 years ago

@Enny1991 Have updated with the discussed change.

jamessergeant commented 7 years ago

Hi @Enny1991 just though I'd follow up on this.

Enny1991 commented 7 years ago

Yup, sorry for the delay!

jamessergeant commented 7 years ago

Thanks @Enny1991 :) Cheers for the implementation btw. Is coming in very handy.