Does not work with Tensorflow

[Sohojoe]

I get the following error when running with TensorFlow (with GPU)

current context was not created by the StreamExecutor cuda_driver API: 0x2ccf300; a CUDA runtime call was likely performed without using a StreamExecutor context

Issue is that the code is referencing Theano directly:- from theano.gradient import disconnected_grad

reference:- https://github.com/tensorflow/tensorflow/issues/916

[sherjilozair]

it should be possible to use tf.stop_gradient to do this. However, this change should ideally belong to Keras backend, so that the DQN code becomes backend-agnostic.

[calclavia]

Use K.stop_gradient?