Re: [Discuss-gnuradio] [SOCIS] Keras, TensorFlow and GNU Radio Blocks

[Christopher Richardson]

Hi Stefan,

I've not tried using Theano as a backend yet, I'll have a look into that,

as at the moment I am using TensorFlow specific features to write

the graph file.

It'd certainly be interesting to compare the performance of both those backends in training.

Cheers

Chris

On Mon, Aug 8, 2016 at 2:44 PM, Stefan Wunsch <address@hidden> wrote:

Hi,

Keras is a very good choice! Does the Theano backend still work? So can
you switch seamlessly between Tensorflow and Theano (altering the
.keras/keras.json)?

I've experienced that getting the GPU support running with Theano was
much easier when using Keras (export THEANO_FLAGS is enough), but
Tensorflow is still always a little bit faster in training. So having
both would be awesome!

Greetings,
Stefan

On 08/08/2016 11:25 AM, Christopher Richardson wrote:

> Hi,
>
> I've now created two GNU Radio blocks, which load a raw CNN model and a
> FAM model, making use of Keras (a TensorFlow API) to create the model
> and TensorFlow serving to save the graph (this requires TensorFlow 0.10,
> to run).
>
> https://signalsintelligence.wordpress.com/2016/08/06/keras-tensorflow-and-gnu-radio-blocks/
>
> Cheers
>
> Chris
>
>

> _______________________________________________
> Discuss-gnuradio mailing list
> address@hidden
> https://lists.gnu.org/mailman/listinfo/discuss-gnuradio
>