Replace stand-alone keras with tensorflow.keras

budach / pysster

pysster: Learning Sequence And Structure Motifs In Biological Sequences Using Convolutional Neural Networks

MIT License

86 stars 23 forks source link

Closed budach closed 6 years ago

budach commented 6 years ago

Simply replacing all occurrences of from keras with from tensorflow.python.keras works like a charm - no crashes - BUT: the loss on my data sets doesn't improve over time anymore. Other users observe similar behavior (https://github.com/tensorflow/tensorflow/issues/15026, https://github.com/tensorflow/tensorflow/issues/15831).

I will try again with a future tensorflow version. Having one less dependency is always nice.

budach commented 6 years ago

Note: Losses still don't develop with Tensorflow 1.6

budach commented 6 years ago

The reason for the differences seems to be that the tensorflow.keras API lags behind stand alone keras API. Not doing it for now.

ghost commented 5 years ago

@budach @fchollet what is your final solution for this important bug.