tf: Scope model under a model identifier (WIP)

[julien-c]

So what I want to be able to do is:

botFoo = chatbot.Chatbot()
botFoo.main(['--modelTag', "foo", '--test', 'daemon'])

botBar = chatbot.Chatbot()
botBar.main(['--modelTag', "bar", '--test', 'daemon'])

This does not work right now as there are collisions between variable names in the TF symbolic graph. This first commit is – I think – a right step towards fixing this, but is not enough: I think in the model.py file, where we rely on the tf.nn.seq2seq.embedding_rnn_seq2seq implementation, something must not be properly scoped.

What do you guys think?

[eschnou]

Actually, there is a scope argument to the method embedding_rnn_seq2seq, so this might be what you need to use to scope your two models properly.

See here in the TF code:
https://github.com/tensorflow/tensorflow/blob/r0.12/tensorflow/python/ops/seq2seq.py#L292

A specific model scope could thus be added here in DeepQA:
https://github.com/Conchylicultor/DeepQA/blob/master/chatbot/model.py#L169

Haven't tried, but seems like the right place to look into. Also note that the current implementation of loading pre-trained embeddings relies on fixed scope name. This should be solved if you change scoping:
https://github.com/Conchylicultor/DeepQA/blob/master/chatbot/chatbot.py#L390

[julien-c]

Thanks @eschnou ! I will look into it tomorrow and hopefully have a mergeable branch by then.

[Conchylicultor]

As both chatbots would use different sessions, I think they should be on different graphs. If you simply change the scope, as the scope is global, the second session would also allocate memory for the first chatbot. I didn't try but that would be more something like:

g1 = tf.Graph()
with g1.as_default():
    botFoo = chatbot.Chatbot()
    botFoo.main(['--modelTag', "foo", '--test', 'daemon'])

g2 = tf.Graph()
with g2.as_default():
    botBar = chatbot.Chatbot()
    botBar.main(['--modelTag', "bar", '--test', 'daemon'])