Keras: AttributeError: module 'tensorflow' has no attribute 'get_default_session'

[nbro]

System information

• Have I written custom code (as opposed to using example directory): no
• OS Platform and Distribution: 8.04.3 LTS (Bionic Beaver)
• TensorFlow backend (yes / no): yes
• TensorFlow version: 2.0.0 (gpu)
• Keras version: 2.2.5
• Python version: 3.6.8
• CUDA/cuDNN version: 10.1
• GPU model and memory: Tesla K80 (memory limit 11330115994).

Describe the current behavior

I get the following error

AttributeError                            Traceback (most recent call last)
<ipython-input-17-3d00d838479b> in <module>()
      1 from keras import backend as K
----> 2 K.tensorflow_backend._get_available_gpus()

1 frames
/usr/local/lib/python3.6/dist-packages/keras/backend/tensorflow_backend.py in get_session()
    188     global _SESSION
    189 
--> 190     default_session = tf.get_default_session()
    191 
    192     if default_session is not None:

AttributeError: module 'tensorflow' has no attribute 'get_default_session'

when executing the following code

from keras import backend as K
K.tensorflow_backend._get_available_gpus()

or

from tensorflow.keras import backend
backend._get_available_gpus()

AFAIK, sessions should still be supported in TF 2.0. Even though _get_available_gpus() is considered a private method, given the initial _, it does fail with TF 2.0, so I think this is a bug.

Describe the expected behavior

No error and I get the available GPUs. In general, I would like to make sure Keras uses GPU (with TF 2.0). Any reliable method to achieve this?

Code to reproduce the issue

See above.

Other info / logs

See above.

[burakyldrm]

I think the following code will solve the problem
pip install --force-reinstall tensorflow==1.15

[nbro]

@burakyldrm Of course, I don't want to use any other TensorFlow version other than 2.0.0, as I said above.

[Rogan-Clark]

I was having the same problem, but if you modify "tf.get_default_session" to "tf.compat.v1.get_default_session", it should work, at least until the next TensorFlow update.

[nbro]

@Rogan-Clark Have you read my post above? The only lines of code I had were

from keras import backend as K
K.tensorflow_backend._get_available_gpus()

[Rogan-Clark]

You have to go into the /usr/local/lib/python3.6/dist-packages/keras/backend/tensorflow_backend.py file and edit it in that, I think.

[nbro]

@Rogan-Clark Well, I would like to use TensorFlow and Keras out of the box. I don't want to mess with their implementation.

[pargass31]

Hi nbro, did you solve your issue? (I have exactly the same one)
thanks

[zbli]

I also run into the same problem. It would be appreciated if the solving method could be informed.
Thanks

[schustan235]

Same issue here, have also tried to go into the source files with no luck. But I can't find "tf_keras_backend" defined anywhere.

`"""Get a list of available gpu devices (formatted as strings).

# Returns
    A list of available GPU devices.
"""
global _LOCAL_DEVICES
if _LOCAL_DEVICES is None:
    if _is_tf_1():
        **devices = get_session().list_devices()**
        #Source of error^^^
        _LOCAL_DEVICES = [x.name for x in devices]
    else:
        _LOCAL_DEVICES = tf.config.experimental_list_devices()
return [x for x in _LOCAL_DEVICES if 'device:gpu' in x.lower()]

`

You have to go into the /usr/local/lib/python3.6/dist-packages/keras/backend/tensorflow_backend.py file and edit it in that, I think.
^^See above @Rogan-Clark

[advset]

I have the same issue while running

# freeze graph frozen_graph = freeze_session(K.get_session(), output_names=[out.op.name for out in model.outputs])

I imported the correct library I guess and load an h5 model I previously made using Keras and TF2

`from keras import backend as K

from tensorflow.keras.models import load_model

load model

model = load_model('C:/Python36/face-mask-detector/mask_detector.model')`

haha T_T