Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 83.62%. Comparing base (9f00320) to head (af75564).
Report is 1 commits behind head on main.

@@           Coverage Diff           @@
##             main    #2752   +/-   ##
=======================================
  Coverage   83.62%   83.62%           
=======================================
  Files         825      825           
  Lines      108686   108686           
=======================================
  Hits        90893    90893           
  Misses      17793    17793           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

Here is a test result, the results maybe a little different each time because the datasets were randomly generated.

This implementation is complementary to Burn's official LSTM, users can choose either one depends on the project's specific needs.

Good! I am not so familiar with Rust and Burn as you. I am glad that you can revise the code. I'll learn from you on the design logic and formatting.

Good! I am not so familiar with Rust and Burn as you. I am glad that you can revise the code. I'll learn from you on the design logic and formatting.
I am not clear about the working mechanisms of github. I find that you have modified the code in my pull request. Do I need modify the source code and pull a request again, or you will make all the changes you mentioned instead of me ? If I need to make the changes you suggest, I will modify and test it in my fork and pull a request again.

I am not clear about the working mechanisms of github. I find that you have modified the code in my pull request.

The specific changes for the Cargo.toml were pretty short so I was able to suggest the code changes from the github PR review, and as you can see it automatically committed them to your fork when you applied them.

Do I need modify the source code and pull a request again, or you will make all the changes you mentioned instead of me ? If I need to make the changes you suggest, I will modify and test it in my fork and pull a request again.

The contribution process usually happens in a sequential manner. A user (you) submits code changes, a contributor (me) reviews the code and requests modifications where needed, and the user can provide his thoughts on the requests and make the necessary changes before asking for another round of review. And the process is typically repeated. But everything remains as part of the same pull request 🙂

In your last PR, the only changes required were structural (change how the example needed to be exposed in the workspace) so I made the changes myself since I had time.

I'll let you make the changes, but if you need help let me know!

It's weired that the model performance decreased a little after these changes. Is the latest burn still unstable or the crates in the Cargo.toml will affect the performance ? I didn't change any parameters and the model. The only things changed are Cargo.toml and the shell commands. I tested the orginal one again, it indeed worked better than this new one.
This is the new one:

I think I find the reason. I copied the code from the wgan-generate.rs (https://github.com/tracel-ai/burn/blob/main/examples/wgan/examples/wgan-generate.rs), where the autodiffbackend was used. But in modern-lstm inference, I think the backend without autodiff must be used, otherwise the predictions will be erroneous because of the layernorm and dropout. Now it works as expected when using backend without autodiff for inference.

I think I find the reason. I copied the code from the wgan-generate.rs (https://github.com/tracel-ai/burn/blob/main/examples/wgan/examples/wgan-generate.rs), where the autodiffbackend was used.

Oh good catch! The structure likely was carried over from another example, but I just fixed it in #2736.

I'll fix the wgan example.