returnn
cinder
returnn | cinder | |
---|---|---|
4 | 43 | |
349 | 3,379 | |
0.6% | 0.6% | |
9.8 | 9.4 | |
10 days ago | 5 days ago | |
Python | Python | |
GNU General Public License v3.0 or later | GNU General Public License v3.0 or later |
Stars - the number of stars that a project has on GitHub. Growth - month over month growth in stars.
Activity is a relative number indicating how actively a project is being developed. Recent commits have higher weight than older ones.
For example, an activity of 9.0 indicates that a project is amongst the top 10% of the most actively developed projects that we are tracking.
returnn
-
Keras Core: Keras for TensorFlow, Jax, and PyTorch
That looks very interesting.
I actually have developed (and am developing) sth very similar, what we call the RETURNN frontend, a new frontend + new backends for our RETURNN framework. The new frontend is supporting very similar Python code to define models as you see in PyTorch or Keras, i.e. a core Tensor class, a base Module class you can derive, a Parameter class, and then a core functional API to perform all the computations. That supports multiple backends, currently mostly TensorFlow (graph-based) and PyTorch, but JAX was something I also planned. Some details here: https://github.com/rwth-i6/returnn/issues/1120
(Note that we went a bit further ahead and made named dimensions a core principle of the framework.)
(Example beam search implementation: https://github.com/rwth-i6/i6_experiments/blob/14b66c4dc74c0...)
One difficulty I found was how design the API in a way that works well both for eager-mode frameworks (PyTorch, TF eager-mode) and graph-based frameworks (TF graph-mode, JAX). That mostly involves everything where there is some state, or sth code which should not just execute in the inner training loop but e.g. for initialization only, or after each epoch, or whatever. So for example:
- Parameter initialization.
- Anything involving buffers, e.g. batch normalization.
- Other custom training loops? Or e.g. an outer loop and an inner loop (e.g. like GAN training)?
- How to implement sth like weight normalization? In PyTorch, the module.param is renamed, and then there is a pre-forward hook, which on-the-fly calculates module.param for each call for forward. So, just following the same logic for both eager-mode and graph-mode?
- How to deal with control flow context, accessing values outside the loop which came from inside, etc. Those things are naturally possible eager-mode, where you would get the most recent value, and where there is no real control flow context.
- Device logic: Have device defined explicitly for each tensor (like PyTorch), or automatically eagerly move tensors to the GPU (like TensorFlow)? Moving from one device to another (or CPU) is automatic or must be explicit?
I see that you have keras_core.callbacks.LambdaCallback which is maybe similar, but can you effectively update the logic of the module in there?
-
Python’s “Type Hints” are a bit of a disappointment to me
> warnings of IDEs are simple to ignore
This is unusual. In my experience, of codebases I have worked with or have seen, when there are type hints, there are almost all perfectly correct.
Also, you can setup the CI to check also for IDE warnings. For example, we use this script for PyCharm: https://github.com/rwth-i6/returnn/blob/master/tests/pycharm...
The test for PyCharm inspections only passes when there are no warnings.
Although, I have to admit, we explicitly exclude type warnings because here we have a couple of false positives. So in this respect, it actually agrees with the article.
But then we also do code review and there we are strict about having it all correct.
Yes, I see the argument of the article that the typing in Python is not perfect and you can easily fool it if you want, so you cannot 100% trust the types. But given good standard practice, it will only rarely happen that the type is not as expected and typing helps a lot. And IDE type warnings, or mypy checks still are useful tools and catch bugs for you, just not maybe 100% of all typing bugs but still maybe 80% of them or so.
> Isn’t it better to detect at least some errors than to detect none at all?
-
How to cleanup a branch (PR) with huge number of commits
I was trying to implement some new feature in some larger somewhat messy project (RETURNN but not so relevant).
So I created a new branch, also made a GitHub draft PR (here), and started working on it.
cinder
-
Meta Used Monolithic Architecture to Ship Threads in Only Five Months
Meta is actually contributing directly to upstream cpython. If you really wanted to, the internal fork is also open source: https://github.com/facebookincubator/cinder
- Meta pledges Three-Year sponsorship for Python if GIL removal is accepted
- Back end of Meta Threads is built with Python 3.10 with some interesting tweaks
-
Lessons from Mojo for PHP 10+ ?
Just one example: last year Meta open-sourced Cinder, which powers Instagram and provides sizeable speedups compared to CPython.
- Python true static typing
-
Best book on writing an optimizing compiler (inlining, types, abstract interpretation)?
I used to work on the Cinder JIT and can help document any passes you find interesting or confusing.
-
Python-based compiler achieves orders-of-magnitude speedups
You might enjoy Cinder then. It's based on CPython so it is nearly 100% compatible.
https://github.com/facebookincubator/cinder/
Disclaimer: I used to work on it.
-
beartype: It has documentation now. It only took two years, my last hair follicle, precious sanity points (SPs), and working with Sphinx. Don't be like @leycec. Go hard on documentation early.
I think Cinder's Static Python, which also performs runtime type checking, is more ambitious. Though it's not production ready yet.
-
If there’s gonna be a Python 4.0 one day, what’s a breaking change you’d like to see? Let’s explore the ideas you have that can make Python even better!
Here's a fork that implements that https://github.com/facebookincubator/cinder - it might be nice to one day get that up streamed but obviously it'll be controversial and it certainly needs more time to bake. Hopefully at some point we can make it a pip installable extension though.
-
Is it time for Python to have a statically-typed, compiled, fast superset?
The other thing that was interesting to me, was the potential of type annotations to help make for a faster, safer experience on the compiler end of things. One example is seen in Meta’s Cinder project, on the docs it explains how typing can be used to reduce the number of steps for the compiler ([cinder/static_python.rst at cinder/3.8 · facebookincubator/cinder · GitHub](https://github.com/facebookincubator/cinder/blob/cinder/3.8/CinderDoc/static_python.rst)), making it more effective.
What are some alternatives?
punctuator2 - A bidirectional recurrent neural network model with attention mechanism for restoring missing punctuation in unsegmented text
faster-cpython - How to make CPython faster.
enforce - Python 3.5+ runtime type checking for integration testing and data validation
Pyjion - Pyjion - A JIT for Python based upon CoreCLR
keras-nlp - Modular Natural Language Processing workflows with Keras
Pyjion
recurrent-fwp - Official repository for the paper "Going Beyond Linear Transformers with Recurrent Fast Weight Programmers" (NeurIPS 2021)
graalpython - A Python 3 implementation built on GraalVM
keras-core - A multi-backend implementation of the Keras API, with support for TensorFlow, JAX, and PyTorch.
MonkeyType - A Python library that generates static type annotations by collecting runtime types
i6_experiments
hpy - HPy: a better API for Python