AttributeError with limit_gpu_memory

[flevet]

Hi,

I just installed CSBDeep and I would be interested in using the limit_gpu_memory. When uncommented in the jupyter notebook, I have this error :

AttributeError                            Traceback (most recent call last)
<ipython-input-2-4d7e1208218d> in <module>
----> 1 limit_gpu_memory(fraction=0.5)

~\Anaconda3\envs\tf\lib\site-packages\csbdeep\utils\tf.py in limit_gpu_memory(fraction, allow_growth)
     43     fraction is None or (np.isscalar(fraction) and 0<=fraction<=1) or _raise(ValueError('fraction must be between 0 and 1.'))
     44 
---> 45     if K.tensorflow_backend._SESSION is None:
     46         config = tf.ConfigProto()
     47         if fraction is not None:

AttributeError: module 'keras.backend.tensorflow_backend' has no attribute '_SESSION'

Is it a problem related to versions of tensorflow-gpu, cuda and cuDNN ?

For information I'm using Anaconda with Python 3.7 and I created a specific environement:

conda create -n tf
conda activate tf
conda install tensorflow=1.14.0
conda install tensorflow-gpu=1.14.0
pip install csbdeep

Best,
Florian

[maweigert]

Hi Florian,

This looks a keras vs tensorflow version mismatch. Can you try to reinstall keras ?

pip uninstall keras
pip install -U keras

[flevet]

Hi Martin,

No change unfortunately. For information, it is keras-2.3.1 that is being installed.
With:
tensorflow 1.14.0
tensorflow-gpu 1.14.0
cudatoolkit 10.0.130
cudnn 7.6.4

[maweigert]

Hi Florian,

You are right, it does fail for keras > 2.2.5 (the _SESSION attribute was removed at some point).
We will look into that, thanks for pointing this out!

In the meantime, after downgrading keras it should work again

pip uninstall keras csbdeep
pip install keras==2.2.5 csbdeep

[flevet]

Hi Martin,

Thanks, it resolved the problem !