此页面由 AI 自动翻译。查看英文原版
最后更新

本页目录

  • 模型
  • Models from training
  • Model File(s)
  • Model Card
最后更新
本节介绍将帮助您有效使用模型注册表平台的关键概念、思想和术语。

模型

模型是 Hub AI 和模型领域中的基本元素,由以下部分组成:
  • 模型文件 - 模型文件或 NN 归档文件(如果模型文件与 JSON 元数据一起打包),以及
  • 模型卡片 - 模型的全面描述。
模型可以是公共的或私有的(团队拥有),并且可以自由地上传到集合或从集合中下载。 有关详细信息,请参阅 上传/下载 指南。每个模型文件构成一个单独的变体,其特征在于其训练配置和使用上下文。为了支持高效的模型开发和在部署应用程序中精确引用,我们将模型文件分为三个嵌套级别。这种结构有助于您:
  • 跟踪模型的演变,
  • 区分语义上不同的模型,
  • 并通过适当的标识符引用正确的工件。
每个嵌套级别——变体、版本和实例——都有其自己的唯一 ID,可用于您的 DepthAI 管道中,以从 Hub 获取相应的模型。下图说明了层次结构:
Markdown
1├── Variant 1
2|   ├── Version 1
3|   |   ├── Instance 1
4|   |   └── Instance 2
5|   └── Version 2
6|       ├── Instance 1
7|       └── Instance 2
8├── Variant 2
9...

Variant

A Variant is a specific version of a model, typically distinguished by input resolution (e.g., 224x224), dataset used for training (e.g., COCO), or some smaller variation in model architecture. Each variant is associated with a Variant ID.

Version

Each model Variant can have multiple Versions. Semantically, this can be used to track different weights of your model that you are developing. You can have an initial version, then a version finetuned, and a version that is, for example, used in production. Each Version has its Version ID.

Instance

An Instance is a compiled, platform-targeted version of a Variant — the actual artifact deployed to hardware. For more information about the conversion to a platform-specific instance, see the Detailed Conversion guideEach instance is identified by a unique Instance ID that points to a specific platform export rather than a general model variant/version. Therefore, when integrating an Instance into your pipeline, ensure the Instance ID corresponds to the correct platform for your deployment environment.

Models from training

A successful Hub AI training run automatically creates a new version under the model linked to its project. The version name is based on the source dataset and model input dimensions, for example Warehouse Pallets 512x384, while Hub advances the model's semantic version automatically.The Training Run column links the model version back to the experiment that produced it. Each successful run registers two instances:
  • an ONNX NN Archive containing the trained inference model and its metadata;
  • a PyTorch checkpoint used to resume training, evaluate the model, or pre-annotate a compatible dataset.
Version tags such as candidate or production can identify the currently promoted result without removing the immutable version and run history. Reusing a tag within one model moves that tag from its previous version.See Training & Evaluation Runs for how these artifacts are created and The Hub AI workflow for how they return to pre-annotation and deployment.

Model File(s)

Model file(s) (also referred to as model executable(s)) come standalone or packed into a NN Archive. The model file(s) must be either in ONNX (.onnx), OpenVINO IR (.xml and .bin), or TensorFlow Lite (.tflite) or one of the RVC compiled formats.

Model Card

A Model Card is a collection of information about the model. You can choose your own structure for private models. For public models, we suggest the following structure (check out the Model Card Template ):
Markdown
1# Model Details
2
3## Model Description
4... Description of the model functionality.
5
6- Developed By - Name of the model developer(s).
7- Shared By - Source of the model file/weights.
8- Model Type - General model type (e.g. computer vision).
9- License - Link to the license that governs the use/distribution of the model.
10- Resources - Link(s) to the model resources (e.g. paper, sourecode, etc.)
11
12# Training Details
13
14## Training Data
15... Describe and link to the data used to train the model.
16
17# Testing Details
18
19## Metrics
20... Describe and report the metrics used to validate the model.
21
22# Technical Specifications
23
24## Input/Output Details
25... Name and a short description for each input/output tensor.
26
27## Model Architecture
28... Describe the model architecture (e.g. backbone, head, etc.).
29
30## Throughput
31... Report model throughput on RVC platform(s) for which the model is converted.
32
33## Quantization
34... Describe quantization data if quantization is used during conversion.
35
36# Utilization
37... Explain how to utilize the model in a DepthAI pipeline. Ideally provide or link to an example.