Variable.ref() no longer exists as of TensorFlow v0.12

[shoyer]

It's now a private method Variable._ref())

Edward uses this in edward/inferences/klpq.py and edward/inferences/klqp.py for computing gradients, in lines that look like:

    grads = tf.gradients(
        -tf.reduce_mean(q_log_prob * tf.stop_gradient(w_norm)),
        [v.ref() for v in var_list])

I believe it should work if [v.ref() for v in var_list] is simply replaced by var_list.

[dustinvtran]

the proposed solution makes sense. we should check this; i forget why but i recall having to add ref() for a reason. maybe it was because tf.gradients() didn't work explicitly on tensorflow variables and do such a conversion internally? (vague recollection)

[gokceneraslan]

I addressed this and a few other TF 0.12 incompatibilities in #374.

[dustinvtran]

looking back, i used [v.ref() for v in var_list] instead of var_list because the compute_gradients() method in TensorFlow optimizers does it: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/training/optimizer.py#L347

this looks like overkill for us so indeed we can drop it.