训练计算机视觉模型

Data strategy is product strategy; especially in ML!
数据策略就是产品策略；尤其是在机器学习中！
Getting started with ML is more accessible than you think (thanks in large part to this template).
开始使用 ML 比你想象的要容易得多（在很大程度上要感谢这个模板）。
Define the usecase first. Not starting with a clear end-goal leads to drift.
首先定义用例。不以明确的最终目标开始会导致偏离。
ML is a mature field with a lot of learning and experimentation resources to lean on.
机器学习是一个成熟的领域，有很多学习和实验资源可以依赖。

Last week I came across a post by mfranzon about training a computer vision ML model on a laptop with just a few lines of Python.

上周我看到了一篇由 mfranzon 发布的帖子，内容是关于在笔记本电脑上用几行 Python 训练计算机视觉机器学习模型。

I’ve done a lot of experimenting with foundational models but hadn’t yet experimented with model training much. This seemed like a low-friction opportunity to go ‘up the stack’ and get some hands-on experience to develop better intuition on the topic.

我在基础模型上进行了很多实验，但还没有进行过多的模型训练实验。这似乎是一个低摩擦的机会，可以“向上堆栈”并获得一些实践经验，以便在这个主题上发展更好的直觉。

This post was written in an effort to be more intentional about documenting and sharing my experimentation. I haven’t done one of these in years so I’ll be shaking off some of the rust; bear with me.

这篇文章是为了更有意识地记录和分享我的实验而写的。我已经很多年没有做过这样的事情了，所以我会有些生疏；请多多包涵。

YOLO Training Template

YOLO 训练模板

Mfranzon was generous enough to post his training template here. The readme says it contains

Mfranzon 很慷慨地在这里发布了他的训练模板。自述文件说它包含

a template for training YOLO models on any Kaggle dataset…
一个在任何Kaggle数据集上训练YOLO模型的模板……

I’m still new to this so I started with some definitions:

我还很新，所以我从一些定义开始：

You-Only-Look-Once (YOLO) refers to an object detection algorithm that “looks” at an image only once to find both bounding boxes and probabilities simultaneously (this blog from v7 was helpful).
You-Only-Look-Once (YOLO)指的是一种目标检测算法，它“只看”一次图像，以同时找到边界框和概率（v7的这篇博客很有帮助）。
Kaggle dataset just refers to any one of the many datasets hosted on the Kaggle platform. Kaggle is great resources for learning about ML.
Kaggle 数据集 指的是托管在 Kaggle 平台上的众多数据集中的任何一个。Kaggle 是学习机器学习的绝佳资源。

Ok so now I have a good-enough underst...