gradCheck test about torch-autograd

[lukealonso]

Hi @ronanmoynihan ,

The following is not safe:

if type(inputs.T)~='table' then

Autograd wraps all your tensor input parameters, to allow them to be tracked throughout the function. So while you can apply the usual torch functions / operators to them, you shouldn't check the types.

It seems like you were trying to work around the fact that assignment in the form of:

delta[torch.gt(delta,0)] =

Is not currently a supported operation. Let me talk to the team and see what can be done.

Luke

[lukealonso]

Also, using the : operator is not valid. You'll have to rewrite:

inputs.x:size(1)

to

torch.size(inputs.x, 1)

etc

[ronanmoynihan]

Hi @luketwitter,
OK, yeah I was thinking those commands (.gt & .lt) were not supported when I first started using it but managed to get a model training somewhat as I expected which gave similar predictions to a model that wasn't using Autograd.

This is a simple test on the Boston housing data and similar to the gradCheck linked above using a normal Torch NN with an Autograd criterion.

https://github.com/ronanmoynihan/autograd-criterion/blob/master/main.lua

th main.lua -under

[alexbw]

@ronanmoynihan Are you still having issues or can I close?

[ronanmoynihan]

Hi @alexbw you can close it.
Thanks.