pytorch/ignite

Event handler to "unlock" GCPs Vertex.AI hyperparameter tuning service

Open

#3,023 创建于 2023年8月14日

在 GitHub 查看
 (2 评论) (0 反应) (1 负责人)Python (4,313 star) (602 fork)batch import
help wanted

描述

Dear PyTorch team,

I have developed a custom event handler that would make ignite code (more specifically: MONAI code) accessible to hyperparamer tuning jobs in GCP Vertex AI. It is an Ignite-tified version of this Link. As you can see from the code in the link, the metrics are simply saved at a certain time, in a certain place with a certain syntax. Once this is possible the Vertex HPO orchestration kicks in. Input arguments are controlled via Vertex.AI custom training jobs and output model performance can be extracted from that output file written at the end of the training.

Why is this useful? With this handler Ignite code can be subject to "outsourced" hyperparameter screening with just adding this handler and a few lines of Vertex config files. I found the outsourcing of HPO to the cloud platform way easier than coding it myself.

If you want I can contribute my solution to the codebase via a PR. Just let me know. image003_obf image002_obf image001

贡献者指南

技术栈
python
领域
machine learningcloud
议题类型
feature
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
3
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
1-3 hours
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
needs maintainer response
清晰度议题是否清楚说明期望改动、验收标准和下一步。
clear
前置要求
PythonPyTorch IgniteGCP Vertex AI
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
50
研究方向
Review the existing event handlers in the pytorch/ignite codebase (e.g., in ignite/handlers/) to understand the interface. Then, adapt the code from the provided GCP hypertune link (https://github.com/GoogleCloudPlatform/cloudml hypertune/blob/master/hypertune/hypertune.py) to create a new handler. The handler should write metric values to a temporary file at the end of each epoch, following the Vertex AI HPO expected format. The main contribution is to package this as a PR integrating the handler into ignite's handler module.