FluxML/Flux.jl

'cell_clip' and 'proj_clip' for Flux: LSTM

Open

#653 建立於 2019年3月1日

在 GitHub 查看
 (1 留言) (0 反應) (0 負責人)Julia (4,725 star) (619 fork)batch import
enhancementhelp wanted

描述

Looking through the docs. , I found that this feature is missing for LSTM layers. It comes as an optional parameter in tensorflow LSTMCell This will be useful for porting stuff to flux while achieving the same results as from tensorflow implementations for some popular algorithms (like ELMo). Looking in the code base of tensorflow if says that it clip the value of cell states between [-cell_clip, cell_clip] in it's clip_by_value method.

貢獻者指南

技術棧
無
領域
machine learning
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。
3
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。
1-3 hours
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。
mostly clear
前置要求
Familiarity with LSTM internalsJulia basicsFlux.jl library structure
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。
30
研究方向
The issue requests adding cell clip and proj clip parameters to LSTM layers in Flux.jl. To implement, first locate the existing LSTM implementation in the Flux source, likely in `src/layers/stateless.jl` or similar. Study how TensorFlow's LSTMCell applies clipping using `clip by value`. Then modify the Flux LSTM forward pass to optionally clip cell and hidden states. Consider adding a new keyword argument to the LSTM constructor and integrating it into the recurrent cell computation. Finally, write tests to verify the clipping behavior matches TensorFlow's. Review any linked PRs or comments for additional context.