AttributeError: 'module' object has no attribute 'clear_session' about dist-keras

[mashaye]

Thanks, it works fine now.

[JoeriHermans]

Are you using Keras 2? Or did you run a trainer multiple times, e.g., after training?

[mashaye]

Yes, I am using keras 2.
I tried SingleTrainer, AEASGD, and EAMSGD, same problem for all.

[JoeriHermans]

I'll fix this today :) I'll let you know when it is done.

[JoeriHermans]

Should be fixed now, let me know. However, I couldn't test it on Keras 2 atm since my TensorFlow is quite outdated due to an old libc.

[mashaye]

Thanks for your quick responses.
It is working for the SingleTrainer, but for AEASGD I got the following error:

trained_model = trainer.train(training_set)
File "build/bdist.linux-x86_64/egg/distkeras/trainers.py", line 575, in train
File "build/bdist.linux-x86_64/egg/distkeras/trainers.py", line 425, in allocate_parameter_server
File "build/bdist.linux-x86_64/egg/distkeras/parameter_servers.py", line 230, in init
AttributeError: 'Sequential' object has no attribute 'get_weight'

[JoeriHermans]

Again, typo on my part. This is fixed now in 465c43c I pushed it to the master branch.

I recommend running the ADAG optimizer (see MNIST example). This optimizer usually is more stable and converges faster than the others.

Joeri