Core concepts（核心概念）¶
This page introduces the core concepts behind Model Studio.
Models¶
Models provide a common interface for integrating models in Foundry. Model Studio training runs produce a model that can be used in downstream applications like Python transforms, Pipeline Builder, functions, and more.
Experiments¶
Experiments are artifacts that represent a collection of metrics produced during a model training job. Each model version produced by Model Studio will have an associated experiment that displays representative parameters and metrics about the training job. For example, some trainers may render a graph of a model ensemble that was produced during training under the Plots tab.
Trainers¶
Model studio trainers are the actual model training implementation that is used to train a model. Each trainer is targeted at a specific task.
- Time series forecasting: Predicts future values by analyzing patterns in training data.
- Regression: Predicts continuous numeric values by learning relationships between input features and target variables in the training data.
- Classification: Assigns input data to predefined categories or classes by identifying patterns and distinctions in training data.
Training jobs¶
Training jobs launched from Model Studio will always run against the latest training configuration. Running multiple jobs in a row will continue to use the same configuration, although outputs may change due to changing input data. Each run is tracked and can be viewed in the Model Studio home page.
Training jobs run as standard transforms in Foundry, meaning that data lineage is respected and any markings applied to input datasets will be applied to the output model.
Compute usage is measured and reported as Foundry compute-seconds. Review our usage types documentation for more details.
Parameters¶
Each trainer defines a set of parameters that can be used to control the training job. Trainers also offer distinct parameters that are specific to that trainer.
When configuring a model studio, in-platform documentation will provide details about given parameters. You can view a trainer's documentation page to learn more about that trainer's parameters.
中文翻译¶
核心概念¶
本页介绍 Model Studio 背后的核心概念。
模型¶
模型 为在 Foundry 中集成模型提供了通用接口。Model Studio 训练运行生成的模型可用于下游应用，例如 Python transforms、Pipeline Builder、函数 等。
实验¶
实验 是代表模型训练作业期间生成的指标集合的产物。Model Studio 生成的每个模型版本都会关联一个实验，显示训练作业的代表性参数和指标。例如，某些训练器可能会在 Plots 选项卡下呈现训练期间生成的模型集成图。
训练器¶
Model Studio 训练器是用于训练模型的实际模型训练实现。每个训练器都针对特定任务。
- 时间序列预测： 通过分析训练数据中的模式来预测未来值。
- 回归： 通过学习输入特征与训练数据中目标变量之间的关系来预测连续数值。
- 分类： 通过识别训练数据中的模式和差异，将输入数据分配到预定义的类别或分类中。
训练作业¶
从 Model Studio 启动的训练作业始终基于最新的训练配置运行。连续运行多个作业将继续使用相同的配置，但由于输入数据的变化，输出可能会有所不同。每次运行都会被跟踪，并可在 Model Studio 主页 中查看。
训练作业在 Foundry 中作为标准 transforms 运行，这意味着数据血缘关系得到保留，应用于输入数据集的任何标记都将应用于输出模型。
计算使用量以 Foundry 计算秒数（compute-seconds）进行衡量和报告。有关更多详细信息，请参阅我们的使用类型文档。
参数¶
每个训练器都定义了一组可用于控制训练作业的参数。训练器还提供特定于该训练器的独特参数。
配置 Model Studio 时，平台内的文档将提供有关给定参数的详细信息。您可以查看训练器的文档页面，以了解有关该训练器参数的更多信息。