"mnist.createdatasets" not working properly about rbm_toolbox_lua HOT 6 CLOSED

I think I specified the wrong path in the examples..

try chainging mnist_folder to:

mnist_folder = '../mnist-th7'

from rbm_toolbox_lua.

Thanks

Apparently that was the thing I missed.

Also, I wanted to suggest you add "accuracy" for both train and test in the
display progress.

Regards,
Siavash

On Wed, Sep 3, 2014 at 3:01 PM, Søren Kaae Sønderby <
[email protected]> wrote:

I think I specified the wrong path in the examples..

try chainging mnist_folder to:

mnist_folder = '../mnist-th7'

Reply to this email directly or view it on GitHub
#1 (comment).

from rbm_toolbox_lua.

It should print both

from rbm_toolbox_lua.

Its TR and VAL

from rbm_toolbox_lua.

Based on the code, yes.

But it is giving me a constant zero when I run with my own data. When I get
the final network after the epochs finish and I try the accuracy function,
it gives me a number.

Also, I had to change something in he code to get it to work for my own
data. Decause my data is doubleTensor, I had to remove all parts of the
code which have floatTensor in them. Can it be that this change is causing
problems?

On Wed, Sep 3, 2014 at 7:24 PM, Søren Kaae Sønderby <
[email protected]> wrote:

Its TR and VAL

Reply to this email directly or view it on GitHub
#1 (comment).

from rbm_toolbox_lua.

Its true that there is a problem with double/float. Anyway if you plan to run on a GPU you could
just stick with floats.

I dont know whats the problem is without seeing and example.
The training error is always evaluated. If you specify val data set then validation should be evaluated as well.

If the validation performance is not evaluated then I think the code prints -1

from rbm_toolbox_lua.