To see memory slot probabilities

[joecheriross]

I am trying to see the memory slot probabilities(probabilities associated with different sentences) for a particular query. Is there a way to visualize them ? Please help.

Thanks,
Joe

[domluna]

I think it would be cool, thought about it in the past but I can't guarantee I'll get around to it. If you come up with something feel free to send a pull request!

[joecheriross]

Thanks Dominique.

A few questions to make sure my understanding is correct.
During testing, providing stories and questions.

• In memn2n.py
  probs = tf.nn.softmax(dotted)
  The probs here get the probability values corresponding to each memory slot(sentence in the story) for the given query in each hop right ?
  If there are 3 sentences in the story, then probs will have 3 probability values
• Finding a way to print probs and stories should help to get an idea on how the sentences are picked for answering a question ?

[domluna]

That's correct. Maybe
https://www.tensorflow.org/versions/r0.10/api_docs/python/control_flow_ops.html#Print
could help in this regard.

On Tue, Sep 27, 2016 at 3:01 AM, Joe Cheri Ross [email protected]
wrote:

A few questions to make sure my understanding is correct.
During testing, providing stories and questions.

In memn2n.py
probs = tf.nn.softmax(dotted)
The probs here get the probability values corresponding to each memory
slot(sentence in the story) for the given query in each hop right ?
If there are 3 sentences in the story, then probs will have 3
probability values

Finding a way to print probs and stories should help to get an idea on
how the sentences are picked for answering a question ?

—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
#16 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/AB0aF2BC37b0zK7DbVGLpmNBqAFXovR7ks5quL84gaJpZM4KGoKO
.

[joecheriross]

I tried out some modification (in memn2n.py) to get the memory probabilities (attentions). The last probs obtained, is assigned to a variable self.mem_probs after the hops loop in _inference(). This is returned by the predict_proba() function.

But each time when I run, I get different values for the memory probabilities for the same test instances. I was trying to get accuracy based on attentions. But this gives different accuracy each time. Can you please see if I am doing something wrong ? Please let me know if something is not clear.

modification in memn2n.py

            probs= None #added by joe
            for _ in range(self._hops):
                m_emb = tf.nn.embedding_lookup(self.A, stories)
                m = tf.reduce_sum(m_emb * self._encoding, 2) + self.TA
                # hack to get around no reduce_dot
                u_temp = tf.transpose(tf.expand_dims(u[-1], -1), [0, 2, 1])
                dotted = tf.reduce_sum(m * u_temp, 2)

                # Calculate probabilities
                probs = tf.nn.softmax(dotted)
                #pr= tf.Print(probs, [probs], message="probs: ", summarize=1000000000)
                #self.image_summary_t= tf.image_summary("probs", probs)

                probs_temp = tf.transpose(tf.expand_dims(probs, -1), [0, 2, 1])
                c_temp = tf.transpose(m, [0, 2, 1])
                o_k = tf.reduce_sum(c_temp * probs_temp, 2)


                u_k = tf.matmul(u[-1], self.H) + o_k
                # nonlinearity
                if self._nonlin:
                    u_k = nonlin(u_k)

                u.append(u_k)

            self.mem_probs= probs #added by joe

modification in predict_proba

return self._sess.run([self.predict_proba_op, self.mem_probs], feed_dict=feed_dict)

[domluna]

Does #pr= tf.Print(probs, [probs], message="probs: ", summarize=1000000000) give the same probs as self.mem_probs?

[joecheriross]

Yes. I checked that as well, to make sure that I am not doing something wrong.
It also gives the same value.