Open bzz opened 4 years ago
Check out this pull request on
You'll be able to see Jupyter notebook diff and discuss changes. Powered by ReviewNB.
I figured that reviewing ToC in a notebook diff is not easy, so here it is, the collapsable cell structure with titles:
And here it is customized for a new workshop, and what we discussed already:
Data: exploration
Data: problem definition for CodeSearchNet dataset
Data: generate tensor representation
Data: inspect
Plot subtoken sequence lengths
Train the model
Visualize the training
Inference using trained model
Visualize the attention
Attention Utils
Display Attention
Serving: export the model
Serving: serve predictions over HTTP
Interactive predictions (local webapp)
Compare the results with the literature
Change the model to GGNN (advanced)
Shall we merge this?
Sorry @bzz for the lack of review, I had in mind that it was WIP for some reason even though you asked for review super long time ago, my bad. I think we can merge as is: it seems to need a rebase against the new docker image but so do the other notebooks so we can take care of that in a future PR. As for the experimental plan, we can also discuss it together with the PyTorch version and come to a parallel plan that will be great for both, using this one as the starting point (it seems very good to me).
Previous name suggestion notebook using OpenNMT-tf + youtokenme for BPE was overexposing the accidental complexity of "tenzorisation" of the source code.
This uses https://github.com/tensorflow/tensor2tensor as a library to archive the same, and thus also works on TPU with Colab.
This version is not the final, workshop-ready one but rather an intermediate one that was used on Colab.
We agreed that the scarce time of workshop prep would better be spent not improving this one, but rather including a pyTourch version instead, where it should be easy to incorporate custom models e.g based on GGNN, and contrast the results to seq2seq ones.
@m09 Please, review the structure of the notebook though - I'm planing to reuse it for a new version, so any methodological feedback on it would be very appreciated.