Hacktoberfest 2026：維護者為十月標記出來的 issue，仍然開放、適合新手。 瀏覽 Hacktoberfest issue

Need a way to pass trainable Variables to Optimizer

未關閉
  • 哪裡不對？
#307 12 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

還沒有人認領這個 Issue。

評估

難度
5/5
預估耗時
一週以上
新手友好度
30/100
Issue 類型
功能
描述清晰度
需要釐清
活躍度
停滯
技術堆疊
java, tensorflow
領域
backend-api-design, machine-learning

研究方向

Start with Optimizer.minimize(loss), its addGradients path, and the Graph/GraphOperation graph-walking work in #232. Compare the two proposed approaches and inspect how Model and Metric variables are represented. Done means a decided API or graph traversal that excludes non-trainable variables and avoids the reported unreachable-node error.

由索引模型根據 Issue 內容生成。

描述

In woking with Model training, an issue on Optimizer has shown its head.

Currently, when calling minimize(loss) on the Optimizer instance, the Optimizer code walks the entire Graph and pulls out all the defined Variables in the graph. The idea is when you call minimize(loss), the Optimizer builds gradients based on all the variables. However, when working with Model, this "all variables approach" breaks down, because some variables are not referenced in the loss operand execution path. This produces the following error:

org.tensorflow.exceptions.TFInvalidArgumentException: Cannot compute the partial derivative for node 'model/mse_total' as it's unreachable from the output node(s).

This specific error is because the MSE metric's internal variables are not within the loss execution path. This pattern of "non-trainable variables (weights)" is in most Metric classes, and in the Model itself, so it is wide spread. What we need is a way to distinguish between trainable and non-trainable variables. Trainable variables would then be used to calculate the gradient values in the Optimizer.

In Python tensorflow, the Keras Layers track the trainable variables as an attribute list, the Model then passes the collected lists to the Optimizer's minimize method.

There are a couple of options here:

  1. Mimic TF Keras, and have each Layer identify its trainable variables, Then, pass the trainable variables as a List<Variable<?> list using a call like, Optimizer.minimize(loss, trainableVariables), then have the Optimizer minimize routine call addGradients with this variable list, rather than walk the whole Graph, to compute the gradients.
  2. Within Optimzier.minimize(loss), walk the loss operand execution path to locate any variables contributing to the loss calculation, then pass these to addGradients. A solution based on this option may be facilitated using #232, "Add graph walking functions to Graph and GraphOperation".
主要語言
Java
星號
927
分支
227
30 天內沒有已合併 PR

環境準備

  • 沒有 Dockerfile 或 Docker Compose 檔案
  • 沒有 Pull Request 範本
  • 閱讀貢獻指南

從這裡開始

  1. 先讀完整個 Issue，再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫，在一個分支上完成修改。
  4. 送出 Pull Request，並在描述裡引用這個 Issue 編號。

tensorflow/java 的其他 Issue

查看 tensorflow/java 的全部 Issue

相似的 Issue

更多 Java Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。