jerinphilip / lemonade Goto Github PK

#11 disables tcmalloc which is not strictly preferable. Should revisit at some point.

May be of help:

• https://stackoverflow.com/questions/47599162/pybind11-how-to-package-c-and-python-code-into-a-single-package

Pybindings are expected to merge with bergamot-translator, which generates documentation with doxygen, exhale and sphinx. Python generally works well with sphinx. If we could add the basic wiring here, moving bindings to bergamot-translator can have one more additional feature.

We may potentially use translateLocally if there's a daemonization prospect, but going to go with the python package and the C + Python connect commonly used to implement iBus engines.

Kind-of like eat your own dog-food, so those things live.

https://github.com/google/pytype appears to be good at static analyzing and catching a few things. It gets ugly at failed import from library on applying to raw source.

Need to look into docs. One workaround is to fetch the wheels which contain the so and use those to run pytype.

It's possible to eventually absorb tagtransfer in here. However, requirements should be simple and optional so as to not make a bloated package.

https://github.com/jerinphilip/tagtransfer

The following kind-of tools are common-place in editing non Latin scripts through a US Keyboard layout, including my own mother tongue.

The existing iBus implementation for a demonstration of outbound translation uses an open source provision for the same. It is possible to attach something like iBus (more generally an input method) into a text area. https://swanalekha.smc.org.in/browser/ is such an in-browser implementation I am already aware of and I see source available in https://github.com/smc/swanalekha.

Since https://mozilla.github.io/translate/ is a static page consuming WebAssembly, it is possible to completely deliver the models from a separate URL and the JS through the existing web systems.

I believe JavaScript could potentially provide a richer environment with an ability to store the original source text etc, operating on larger content editable text areas. A verification UI can pop up on hover of the elements presented above, which is what's currently deemed missing in lemonade native.

We need better JavaScript primitives, which are reusable.

The keyboard event and the state machine is at best hacky. It also has some issues with enter keys, with only the first enter key being registered.

The wheels built on ubuntu straighforward work with the respective distributions. However python community seems to have imposed standards on what wheels can be published, and prefers a "manylinux" wheel.

https://github.com/pypa/manylinux

Good thing to have is a GitHub actions environment setting up manylinux, and creating the wheel for the same. We'll have to see if there are any issues of speed or symbol missing or such.

There is a way to connect to pandoc from Python: https://pypi.org/project/pandoc/

Given we already do HTML conversions, it might be possible for us to quickly throw around odt/docx/whatever translations using HTML and pandoc for the rest.

Each returned response, we have information on how translation splits sentences. We can use this to provide seamless typing experience. After maybe some window hyperparameter of consistent start of a new sentence, we may commit text.

The capability already exists in some capacity. It will break backwards compatibility, but issok.
Currently we use only translateLocally.

pytest?

Maybe attempt cross-platform. We have to compile bergamot-translator, so ccache will come in handy. WebAssembly is allowed upstream and won't be duplicated here.

There's something amiss in implementing https://pybind11.readthedocs.io/en/stable/advanced/cast/stl.html, having to do VectorString or VectorResponse explicitly in scripts. This is not a blocker for now, but is bad user experience which we may be able to fix.