Entailment pr

[vyskoto4]

The entailment model implementation is not 100% final, I had to replace some hardcoded code by the SequenceSplit layer and the performance is not as good as before.

[pasky]

softmax on 1 dim?

[vyskoto4]

Ok, it was a typo, replaced it with linear

[pasky]

Thanks a lot for the update! I'm a lot happier I think, a bit wary about the ugly looking code in rnn_input but I understand the neccessity and we'll be able to throw away that complexity again with the Keras1.0 port. So it's okay. :)

But I'm still a bit leery about the ptscorer business. I think the basic sanity check is:

• What happens when we use a different model with the rte task? (E.g. re the sigmoid -> linear activation change. But I don't understand that anyway?)
• What happens when we use the model with a different task?
  Are these cases both okay?

The other problem I see is that we use completely different ptscoring strategies when the to_n is involved - I don't understand that. I think we discussed just making the output dimension a parameter of mlp_ptscorer? Would that be inferior to the current solution?

[vyskoto4]

• The sigmoid-linear change in RTE task has been done only because of the new model, as the reached accuracy was the best with these settings. I also tried RNN, AVG (and maybe CNN) and their performance seemed to be the same as before.
• I did not test it with a different task. That is on my todo list.
• I'l test the modified mlp_scorer as soon as I get some GPU time.

EDIT:

• I deleted the custom scorer and the mlp_scorer is being used. I did not notice any significant performance changes after this.
• Now testing the new (ok, its actually old today) on anssel/wang. I did not run full eval but the best val MRR during traning was 0.845.