From 0baf4e1b249ff03333da152022db18cf2ca05452 Mon Sep 17 00:00:00 2001 From: qiaolongfei Date: Wed, 14 Mar 2018 19:48:05 +0800 Subject: [PATCH] update format --- doc/design/distributed_lookup_table_design.md | 13 ++++++------- 1 file changed, 6 insertions(+), 7 deletions(-) diff --git a/doc/design/distributed_lookup_table_design.md b/doc/design/distributed_lookup_table_design.md index bf959bd097..695a9831e2 100644 --- a/doc/design/distributed_lookup_table_design.md +++ b/doc/design/distributed_lookup_table_design.md @@ -41,17 +41,16 @@ All the operators above access the `lookup_table` directly in memory. If we chan ## TODO 1. Implement `distributed lookup table`, with service part and client part. The client should provide four interfaces: - - `Pull(ids) -> embedding_values` pull embedding_values according to ids. - - `Push(grad, update_method)` push `grad` to the distributed lookup table and update it to the parameter using `update_method `, this interface use is asynchronous. - - `Save()` save the model to a persistent file system, such as HDFS. - - `Load()` load the model from a persistent file system, such as HDFS. - - The details will be proposed in another PR. + - `Pull(ids) -> embedding_values` pull embedding_values according to ids. + - `Push(grad, update_method)` push `grad` to the distributed lookup table and update it to the parameter using `update_method `, this interface use is asynchronous. + - `Save()` save the model to a persistent file system, such as HDFS. + - `Load()` load the model from a persistent file system, such as HDFS. + The details will be proposed in another PR. 1. Design and implement `lookup_table_op` and `lookup_table_grad_op ` with distributed lookup table client. 1. Implement the Python wrapper for above ops, users can choose and config to use these ops. 1. The distributed Fluid should support this `distributed lookup table service` on kubernetes. 1. Implement a `distributed transpiler` that can change the program into a distributed one which will use the `distributed lookup table service`. ## Things need to be discussed -In the above design, the parameter update is done within `distributed lookup table service`, the interface is `Push(grad, update_method)`, this is different than the current design of PaddlePaddle Fluid, Currently, parameter update is done by Operators. So how should we impelement these update_method? +In the above design, the parameter update is done within `distributed lookup table service`, the interface is `Push(grad, update_method)`, this is different than the current design of PaddlePaddle Fluid. Currently, parameter update is done by Operators. How should we impelement these update_method? -- GitLab