search

microsoft / ELL

Embedded Learning Library
https://microsoft.github.io/ELL
Other
2.29k stars 295 forks source link

CNTK import - activation function on fully connected layers #173

Closed sdudeck closed 6 years ago

sdudeck commented 6 years ago

Hello, I am trying to import a simple cntk-model consisting of one fully connected layer only (18 inputs, 2 outputs, relu-activation). It seems as if the ELL-import skips the activation function at all. Independend of the activation function I define for the layer (cntk.tanh, cntk.sigmoid, cntk.relu, None) the resulting ELL-json-file looks exactly the same (beside the weights). Is this a known issue? I think my problems with importing a darknet model a couple of month ago was quite similar, that the activation layer was not imported properly.

Thank you very much, Sven

Model_ell.zip Model_cntk.zip

lovettchris commented 6 years ago

Yes indeed, there is a bug in CntkDenseConverter, in cntk_converters.py. This line:

if self.activation:

should be:

if self.activation is not None:

I'll push a fix.

sdudeck commented 6 years ago

thanks a lot - changing this line did the trick :-)