graph4nlp
graph4nlp copied to clipboard
graph4ai
[Roadmap] Graph4NLP v0.6 plan
Task List
- [New Feature] A brand new configuration system based on OmegaConf.
- [Feature] Upgrade
stanford CoreNLPto Stanza. - [Feature] Improve the dataset custermize experiments.
- [Feature] Add relational GNN support.
@AlanSwift looking forward to the relational GNN support.
By the way, without the support to run in a multiple-GPU environment, not sure how this library could be useful for large datasets. Do you plan to add support for multiple GPUs?
@AlanSwift @hugochan @teddylfwu @SaizhuoWang this is a fantastic library. And to the best of my knowledge, this is the only GNN library that is NLP focused.
Would like to give some feedback about the library:
-
I think the effort for the library is misdirected. For example, the library has severe bug like Beam search is broken. I think any
Majorbug should be prioritized as that completely limits the use of the library. -
Even if I look into the current Roadmap, I see effort is seriously misguided. The only major feature I see is adding
relational GNN supportwhich is great. But what's the point of so many bookkeeping tasks like add abrand new configuration system based on OmegaConf. These are just busy work that does not do much for an average user of the library. I am pretty sure a lot of effort would then be directed just to modify/update existing examples and to use this new configuration system.
Unfortunately that does not do much for a typical user. No one really cares that you are using OmegaConf configuration library. What matters is adding actual useful feature into the library. And it is extremely important to ensure the features that are already into the library actually works and there is no bug. Lets say you added R-GCN and it is buggy. Then the priority should be to fix that bug instead of updating a configuration system.
If there is a major bug, there is no point of adding more and more features into the library. To summarize, for any software engineering tasks, there needs to be a balance between what is "actual feature" vs "doing chores" (like updating configuration system) which does not add much to the library.
This is just my 2 cents. And again - thanks for this fantastic library.
@AlanSwift @hugochan @teddylfwu @SaizhuoWang Furthermore, I see in the description work on configuration system is stated as a New Feature:
[New Feature] A brand new configuration system based on OmegaConf
This is not a feature. Feature is something that adds value to a project. So the understanding of what is called as a "feature" is wrong. It should be stated "Code Debt" or "Maintenance" work.
@AlanSwift @hugochan @teddylfwu @SaizhuoWang @xiao03 wondering when this library version would be released 🙏 🙏 🙏 ?
Reply to all: We will have a meeting this weekend. I think the new release will come in two weeks. Really sorry for the delay.
@AlanSwift this is a great library for GNN. Looking forward to the release specially the implementation of RGCN.
Also if you could support transformer in the future, would a great feature for this useful library.
Yes. The full support of the transformer is on schedule. The transformer encoder will release in the next coming version. And the transformer decoder will come soon (on test). @code-rex1
@AlanSwift looking forward to use the transformer encoder 🥇. When would it be release?
We are still discussing it #578. We still need 2-3 weeks to make it ready since it covers many features. @smith-co
@AlanSwift wondering when the RGCN implementation would be available?
We have almost finished the implementation and currently are testing on down tasks. We are sorry we delayed too much time since it is a big feature migration from the homogeneous graph to the heterogeneous graph.
