Show HN: Autocomplete Python with Deep Learning

[vpj]

https://github.com/vpj/python_autocomplete

[bobajeff]

I've been thinking about using something like this for Swype-like keyboard for coding on phones.

[sho → bobajeff]

I've been waiting for Apple to at least apply some ML in text prediction (and moreso, spellchecking) in its onscreen keyboards. What they have now is worse than useless and I would have thought it would be low-hanging fruit?

[etaioinshrdlu]

Another fun idea is to make a "code denoiser". Removing noise (bugs) from your code.

It's quite simple to synthetically inject lots of noise into code, simply use the wrong variables, operators, numbers, structures, reverse statement ordering, transforming code at the AST level.

Then you can learn mappings of bad code => good code!

The same idea has worked for natural languages as well, correcting typos.

[make3]

I wished they had an online demo, or jupyter notebook with an easy to download pretrained model so I can test it in colab

[nikeee]

Is this similar to Microsoft's IntelliCode? How does it differ?

[veselin → nikeee]

This one is a character level language model. The Microsoft one is per token.

This one should be the closest to the Microsoft model: https://miltos.allamanis.com/publicationfiles/allamanis2018l...

Btw, simple neural networks are far from doing the best for code. The kind of accuracy LSTM gives is already achieved with good n-gram models. E.g. see this one on a Linux code prediction: https://arxiv.org/pdf/1506.02078.pdf Also, there are other techniques that improve on top of the n-gram models: https://files.sri.inf.ethz.ch/website/papers/charmodel-iclr2...

[ma2rten → veselin]

I only skimmed the paper you linked, but I don't think it says what you say it says.

"We found that scaling up the model almost entirely eliminates errors in the n-gram category."

[p1esk → ma2rten]

See table 1 in his last link.

[p1esk → veselin]

What about gpt-2?

[logicprog]

This is pretty cool, but I'm not sure how much better this approach would be compared to language server based completion (something like RLS).

[Matumio → Der_Einzige]

This looks more like a research project. In contrast, tabNine is a closed-source commercial package. No need to compete with that.