vision_transformer
typeshed
vision_transformer | typeshed | |
---|---|---|
7 | 24 | |
9,287 | 4,076 | |
2.2% | 1.2% | |
5.5 | 9.9 | |
about 2 months ago | about 5 hours ago | |
Jupyter Notebook | Python | |
Apache License 2.0 | 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.
vision_transformer
-
Can I use CLIP to tag my picture collection?
And one last thing, should I even be thinking of using CLIP for these tasks when Google has released a better model here: https://github.com/google-research/vision_transformer/blob/main/model_cards/lit.md
-
When the client's management is happy but their dev team is a pain
Google's vision transformers are type hinted.
-
Improving Search Quality for Non-English Queries with Fine-tuned Multilingual CLIP Models
We’re going to look at a model that Open AI has trained with a broad multilingual dataset: The xlm-roberta-base-ViT-B-32 CLIP model, which uses the ViT-B/32image encoder, and the XLM-RoBERTa multilingual language model. Both of these are pre-trained:
-
[R] How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers
JAX Code: https://github.com/google-research/vision_transformer
- [D] (Paper Overview) MLP-Mixer: An all-MLP Architecture for Vision
-
[P] Animesion: a framework, for anime (and related) character recognition. It uses Vision Transformers trained on a subset of Danbooru2018, that we rebranded as DAF:re, and can classify a given image into one of more than 3000 characters! Source code and checkpoints included.
For this project I used the pretrained models released by Google in Jax, using this particular PyTorch custom implementation. Those were pretrained on ImageNet21k with 14 M images among 21 K classes. Then yes I finetune on two datasets: one with 15 K images and 170 characters, and one with 3 K characters and almost 500 K images.
- Short term memory solutions for video tasks?
typeshed
-
What's the point of using `Any` in Union, such as `str | Any`
"csv.pyi is from VS Code Pylance extension" is misleading. Yes, it's included in the code base of the extension, but it's likely originally from python/typeshed. I diffed csv.pyi in the extension and the repository, and they're exactly the same.
-
Importing python libraries "Cannot find implementation or library stub for module named ..."
You can check the typeshed library that offers stubs for many packages.
-
Ask HN: Will we see a TypeScript for Python?
https://github.com/python/typeshed is Python's equivalent of DefinitelyTyped. I'm not 100% sure why it's not more of a popular thing the way DefinitelyTyped is; I think there might, to some extent, be different attitudes around the appropriateness of having third-party typings for packages, when the actual maintainer of the package isn't interested in providing first-party ones.
-
Why Type Hinting Sucks!
https://github.com/python/mypy same with typeshed https://github.com/python/typeshed
-
When the client's management is happy but their dev team is a pain
Here's the tensorflow type stubs on typeshed. https://github.com/python/typeshed/tree/main/stubs/tensorflow
-
Offer to Type Hint API's, or Start a Statically Typed Python?
Also, be aware that there is already a central place for stubs files. If you are going to take the time to write one, contributing it there will help everyone if the package owners aren't already including some type hints.
-
Ruby 3.2’s YJIT is Production-Ready
Python's type hints are definitely an improvement and they're getting better all the time, but they're still frustrating to use at anything approaching the edge. I long for something as elegant and functional as TypeScript.
One hurdle I've stumbled over recently is the question "what is a type?", the answer can be surprising. Unions, for example, are types but not `Type`s. A function that takes an argument of type `Type` will not accept a Union. So if you want to write a function that effectively "casts" a parameter to a specified type, you can't. The best you can do is have an overload that accepts `Type` and does an actual cast, and then another that just turns it into `Any`. This is, in fact, how the standard library types its `cast` function [1]. The argument I've seen for the current behavior is that `Type` describes anything that can be passed to isinstance, but that's not a satisfying answer. Even then, `Union` can be passed to isinstance and still does not work with `Type`. Talk currently is to introduce a new kind of type called `TypeForm` or something to address this, which is certainly an improvement over nothing, but still feels like technical debt.
[1]: https://github.com/python/typeshed/blob/main/stdlib/typing.p...
-
GitHub stars won't pay your rent
>Ultimately if you care enough about Fody to spend over a hundred dollars worth of your time contributing to it, you probably care enough about Fody to drop them three dollars.
No, I really don't.
https://github.com/keepassxreboot/keepassxc/pull/8500 - I was randomly reading keepassxc's manpage and spotted a curious option, spent some time spelunking through the code and history to discover that it was an outdated option, sent a PR.
https://github.com/python/typeshed/pull/8617 - I converted one of the scripts I use in my DE from shell to Python, saw that VSCode has this new fancy typing support for Python, quickly found a basic bug in the type definitions for the os module, tested a fix locally, sent a PR.
https://gitlab.gnome.org/GNOME/gtk/-/issues/5250 - I found an issue with copy-paste on my phone, investigated it all the way through to the GTK stack, found the commits that introduced the issue, created a distro patch for it while discussing it with GTK upstream.
https://gitlab.alpinelinux.org/alpine/aports/-/merge_request... - I noticed that gnome-passwordsafe crashes some times, debugged it to discover that it was missing a dependency, sent a PR to the distro package to update the dependencies.
etc etc. I've made lots of fixes like these. I have no interest in paying for each and every one of them. The projects are all better off for fixes like mine and gatekeeping them on payment would've been nothing but their loss.
-
Wrapping my head around type hinting
The csv module is one of those standard library modules that doesn't provide its own type hints, but instead gets them through the external typeshed project, and (for compatibility/implementation reasons, I surmise) the name of these types sometimes don't quite align with the objects they correspond to. So, for all intents and purposes, _csv._reader is the correct name of the type that csv.reader() returns, as ugly as it is.
-
Using Mypy in Production
You have to do handling like that in other languages like TypeScript anyway.
Painpoint with type annotations:
- not being able to reuse "shapes" of data: TypedDict, NamedTuple, dataclasses.dataclass, and soon kwargs (PEP 692 [1]) all have named, typed fields now. You have to
- Since there's no generic "shape" structure that works across data types, there isn't a way to load up a JSON / YAML / TOML into a dictionary, upcast it via a `TypedGuard`, and pass it into a TypedDict / NamedTuple / Dataclass. dataclasses.asdict() or dataclasses.astuple() return naive / untyped tuples and dicts. Also the factory functions will not work with TypedDict or NamedTuple, respectively, even if you duplicate the fields by hand. See my post here: https://github.com/python/typeshed/issues/8580
- Standard library doesn't have runtime validation (e.g. pydantic / https://github.com/pydantic/pydantic).
- pytest fixtures are hard.
- Django is hard. PEP 681 may not be a saving grace either. [3]
[1] https://peps.python.org/pep-0692/
What are some alternatives?
pytorch-image-models - PyTorch image models, scripts, pretrained weights -- ResNet, ResNeXT, EfficientNet, NFNet, Vision Transformer (ViT), MobileNet-V3/V2, RegNet, DPN, CSPNet, Swin Transformer, MaxViT, CoAtNet, ConvNeXt, and more
pyre-check - Performant type-checking for python.
nerfstudio - A collaboration friendly studio for NeRFs
mypy - Optional static typing for Python
ImageNet21K - Official Pytorch Implementation of: "ImageNet-21K Pretraining for the Masses"(NeurIPS, 2021) paper
NumPy - The fundamental package for scientific computing with Python.
Fashion12K_german_queries
flask-parameter-validation - Get and validate all Flask input parameters with ease.
TorchSharp - A .NET library that provides access to the library that powers PyTorch.
dactyl-keyboard - Web generator for dactyl keyboards.
fashion-200k - Fashion 200K dataset used in paper "Automatic Spatially-aware Fashion Concept Discovery."
Nuitka - Nuitka is a Python compiler written in Python. It's fully compatible with Python 2.6, 2.7, 3.4, 3.5, 3.6, 3.7, 3.8, 3.9, 3.10, and 3.11. You feed it your Python app, it does a lot of clever things, and spits out an executable or extension module.