vespa-engine/vespa

Support a properties-file (available during runtime) for ml-models (Stateless evaluation in container)

Open

#9,552 建立於 2019年5月24日

在 GitHub 查看
 (0 留言) (0 反應) (0 負責人)Java (4,948 star) (561 fork)batch import
enhancementgood first issue

描述

Hello Vespa Team, Can you please consider support a properties-file which is available during run-time along with the model ? So some meta-data e.g. threshold/label, etc can be associated with the model. This is for the stateless evaluation of the models like (XGBoost, TensorFlow, Onnx, etc) which is supported in vespa.

Thank you, Pinank

貢獻者指南

技術棧
javatensorflow
領域
machine learningbackend
議題類型
feature
難度面向新貢獻者的預計實作難度，1 表示很小改動，5 表示專家級工作。
4
預計時間有經驗貢獻者完成調查、實作、測試並準備 pull request 的粗略時間範圍。
over 1 week
活動狀態議題目前的可參與程度：新鮮、活躍、陳舊、阻塞或等待維護者輸入。
stale
清晰度議題是否清楚說明預期改動、驗收標準和下一步。
unclear
前置要求
Understanding of Vespa architectureFamiliarity with ML model evaluation in Vespa
新手友善度1-100 的估計分數，表示該議題對首次貢獻者的友善程度。
10
研究方向
The issue requests adding a properties file for ML models (XGBoost, TensorFlow, ONNX) during stateless evaluation. Since the issue is from 2019 and has no comments, the requester's intent is to associate metadata like thresholds/labels with models. To implement, one should investigate Vespa's model serving infrastructure, specifically how models are packaged and evaluated in containers (e.g., using the model evaluation API). The key challenge is to define a format for the properties file and integrate it into the model loading process. The solution would likely involve modifications to the model configuration and evaluation pipeline.