【Hacker News搬运】苹果发布用于训练深度神经网络的库CoreNet

hackernews

Title: Apple releases CoreNet, a library for training deep neural networks

苹果发布用于训练深度神经网络的库CoreNet

Text:

Url: https://github.com/apple/corenet

CoreNet是一个用于训练深度神经网络的库，它允许研究人员和工程师训练标准和小规模以及大规模的新颖模型，用于各种任务，包括基础模型（如CLIP和LLM）、物体分类、物体检测和语义分割。

最新的更新是2024年4月，CoreNet库的0.1.0版本包括了OpenELM、CatLIP和MLX示例。苹果公司使用CoreNet进行的研究工作包括OpenELM、CatLIP、Reinforce Data, Multiply Impact、CLIP meets Model Zoo Experts、FastVit、Bytes Are All You Need、MobileOne、RangeAugment、Separable Self-attention for Mobile Vision Transformers (MobileViTv2)和CVNets等。

CoreNet的安装需要Git LFS，并且建议在Linux上使用Python 3.10+和PyTorch (版本 >= v2.1.0)，在macOS上使用Python 3.9+。CoreNet的目录结构包括入门指南、训练配方、MLX示例、模型实现、数据集和其他关键目录。

维护者包括Sachin、Maxwell Horton、Mohammad Sekhavat和Yanzi Jin。此外，CoreNet从CVNets发展而来，扩展了其应用范围，包括基础模型的训练，如LLM。如果觉得这个项目有用，可以引用以下论文：Mehta, Sachin and Abdolhosseini, Farzad and Rastegari, Mohammad, CVNets: High Performance Library for Computer Vision, Proceedings of the 30th ACM International Conference on Multimedia, MM '22。

Post by: rocauc

Comments:

gbickford: > Relationship with CVNets> CoreNet evolved from CVNets, to encompass a broader range of applications beyond computer vision. Its expansion facilitated the training of foundational models, including LLMs.We can expect it to have grown from here: <a href="https://apple.github.io/ml-cvnets/index.html" rel="nofollow">https://apple.github.io/ml-cvnets/index.html</a>It looks like a mid-level implementations of training and inference. You can see in their "default_trainer.py"[1] that the engine uses Tensors from torch but implements its own training method. They implement their own LR scheduler and optimizer; the caller can optionally use Adam from torch.It's an interesting (maybe very Apple) choice to build from the ground up instead of partnering with existing frameworks to provide first class support in them.The MLX examples seem to be inference only at this point. It does look like this might be a landing ground for more MLX specific implementations: e.g. <a href="https://github.com/apple/corenet/blob/5b50eca42bc97f6146b812a3e3469959da5da0ec/mlx_examples/clip/model.py#L402">https://github.com/apple/corenet/blob/5b50eca42bc97f6146b812...</a>It will be interesting to see how it tracks over the next year; especially with their recent acquisitions:Datakalab <a href="https://news.ycombinator.com/item?id=40114350">https://news.ycombinator.com/item?id=40114350</a>DarwinAI <a href="https://news.ycombinator.com/item?id=39709835">https://news.ycombinator.com/item?id=39709835</a>1: <a href="https://github.com/apple/corenet/blob/main/corenet/engine/default_trainer.py">https://github.com/apple/corenet/blob/main/corenet/engine/de...</a>

gbickford: &gt；与CVNets的关系&gt；CoreNet从CVNets演变而来，涵盖了计算机视觉之外更广泛的应用。它的扩展促进了基础模型的培训，包括LLM 我们可以期待它从这里开始生长：<a href=“https://；&#x2F；apple.github.io&#x2F：ml cvnets&#x2F！index.html”rel=“nofollow”>https://&#x2F；apple.github.io&#x2F；ml cvnets；index.html</a>它看起来像是训练和推理的中级实现。你可以在他们的“；default_trainer.py”；[1] 引擎使用torch中的张量，但实现自己的训练方法。它们实现了自己的LR调度器和优化器；呼叫者可以选择性地使用来自torch的Adam 它；这是一个有趣的（可能非常像苹果）选择，从头开始构建，而不是与现有框架合作，在其中提供一流的支持 MLX的例子似乎只是在这一点上的推论。看起来这可能是更多MLX特定实现的着陆点：例如，<a href=“https://；&#x2F；github.com&#x2F：apple&#x2F，corenet&#x2F。blob&#x2F！5b50eca42bc97f6146b812a3e3469959da5da0ec&#x2F…MLX_examples&#x2F？clip&#x2Fmodel.py#L402”>https://&#x2F；github.com&#x2F；apple&#x2F；corenet；blob；5b50eca42bc97f6146b812…</a>看看它在明年的跟踪情况会很有趣；尤其是最近的收购：Datakalab<a href=“https://；&#x2F；news.ycombinator.com&#x2F？id=40114350”>https://&#x2F；news.ycombinator.com&#x2F；项目id=40114350</a>DarwinAI<a href=“https:&#x2F；新闻.ycombinator.com&#x2F？项目？id=39709835”>https:&#x20F&#x2F；news.ycombinator.com&#x2F；项目id=39709835</a>1:<a href=“https://；&#x2F；github.com&#x2F：apple&#x2F，corenet&#x2F。blob&#x2F！main&#x2F&#x2F；github.com&#x2F；apple&#x2F；corenet；blob；main；corenet；发动机；de</a>

leodriesch: How does this compare to MLX? As far as I understand MLX is equivalent to PyTorch but optimized for Apple Silicon.Is this meant for training MLX models in a distributed manner? Or what is its purpose?

leodriesch: 这与MLX相比如何？据我所知，MLX相当于PyTorch，但针对Apple Silicon进行了优化 这是为了以分布式方式训练MLX模型吗？或者它的目的是什么？

ipsum2: It's interesting that Apple also actively develops <a href="https://github.com/apple/axlearn">https://github.com/apple/axlearn</a>, which is a library on top of Jax. Seems like half the ML teams at Apple use PyTorch, and the other half uses Jax. Maybe its split between Google Cloud and AWS?

ipsum2: 它；有趣的是，苹果还积极开发<a href=“https://；&#x2F；github.com&#x2F，Apple&#x2F！axlearn”>https://&#x2F；github.com&#x2F；apple&#x2F；axlearn</a>，它是Jax之上的一个库。苹果的ML团队似乎有一半使用PyTorch，另一半使用Jax。也许是谷歌云和AWS之间的分歧？

coder543: They also mention in the README:> CatLIP: CLIP-level Visual Recognition Accuracy with 2.7x Faster Pre-training on Web-scale Image-Text DataThis is the first I’m hearing of that, and the link seems broken.

coder543: 他们还在自述中提到：&gt；CatLIP:CLIP级别的视觉识别精度，在网络规模的图像文本数据上的预训练速度快2.7倍这是我第一次听说，链接似乎断了。

mxwsn: Built on top of pytorch.

mxwsn: 建在pytorch的顶部。