Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/ruhyadi/yolov5-registry
https://github.com/ruhyadi/yolov5-registry
Last synced: about 10 hours ago
JSON representation
- Host: GitHub
- URL: https://github.com/ruhyadi/yolov5-registry
- Owner: ruhyadi
- License: apache-2.0
- Created: 2022-06-27T04:09:29.000Z (over 2 years ago)
- Default Branch: main
- Last Pushed: 2022-06-27T04:09:30.000Z (over 2 years ago)
- Last Synced: 2024-11-08T20:51:04.635Z (about 2 months ago)
- Language: Python
- Size: 151 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# CVAT Model Registry Template
A template to kickstart cvat model deployment
Click on [Use this template](https://github.com/ashleve/lightning-hydra-template/generate) to initialize new repository.
## đź“ŚÂ Â Introduction
The template can be used as a model registry for the model to be deployed to CVAT. This template contains several scripts (`main.py`, `model_handler.py`, `ci.yaml`, etc) you must customize according to the model used. This template also uses CI mechanisms to: build docker images; debug models in Nuclio; and delete GitHub releases when CI failure.
An example of using this template to develop a registry model can be seen on the branch:
- YOLOv5
- YOLOR
- DETR
## 🚀  Quickstart
You can clone the repository and put the script dependency model in it.
```bash
git clone https://github.com/ruhyadi/model-registry-cvat-template ./your_model
cd ./your_model
git clone https://github.com/username/your_model ./
```
You must configure three important scripts to deploy the model to cvat: `main.py`, `model_handler.py`, and `function.yaml`.
### Main Script (`main.py`)
`main.py` is a script used to initiate and inference models. Inside `main.py` there are two functions: `init_context` and `handler`.The `init_context` function initiates the model, while the `handler` function is used for the inference model. The goal of `main.py` is to generate a response as `json`.
The author has provided a `#TODO` that you can follow along:
1. In the `init_context` function, load `model` according to the load model function of `model_handler`.
2. In the `handler` function, change the message in `context.logger.info()`.
3. Set the desired threshold value (default: 0.5).
4. Change the inference model mechanism; the final result will be `results` which contains a dictionary (NOTE: for object detection task):
```yaml
results = {
"confidence": 0.8,
"label": "car",
"xmin": 0.1,
"ymin": 0.2,
"xmax": 0.3,
"ymax": 0.4,
}
```
### Model Handler (`model_handler.py`)
The purpose of `model_handler.py` is to simplify the process of loading and inference models. This process will usually require script dependencies according to each model. There are no special write conditions in `model_handler.py`. However, the handler model must be able to be used for this simplification process.
### Function (`function.yaml`)
`function.yaml` contains the configuration that nuclio will use to deploy the model. The `function.yaml` file already contains information about how to fill the file. The parameter `build.baseImage` should be noted here, which contains the docker image. This parameter will be automatically generated by **system model deployment**, so users do not need to change the model version every time there is an update. However, filling in the initial version of the model is still recommended.
```yaml
build:
image: cvat/model-name
baseImage: username/model-name:version # automatic generated
```
### CI/CD (`workflows/ci.yaml`)
The CI/CD function in this template is helps automate:
- build a docker image model (including model weights);
- debug model with nuclio;
- delete release if any of the functions fail.
**GitHub Release** triggers the CI/CD function by including model weights on Assets. GitHub Release can be done quickly using the **Model Registry** function.
## ℹ️  Questions?
Don't hesitate to make an issue.