-
-
Notifications
You must be signed in to change notification settings - Fork 4
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Converting ONNX models to Flux Chains #21
Comments
As you note, this approach was already discussed in #10. Is there anything to add here that you think wouldn't be relevant to that thread? |
The discussion there is in more general related to ONNX to Flux for all models. Here, I plan to consider only models that can be easily represented by Flux chains. This issue can be considered as a sub-topic to that discussion. (Although, in the future, this can be further expanded to include all relevant points/features mentioned in that issue.) But for now, my main focus here is on providing a tool for quick prototyping using already available ONNX models. |
I believe the approach we're trying now is to directly track work via PRs on Flux (and related repos) and to move non-tracking discussion (e.g. around design) to discussions. Discussions in particular are nice because you can have multiple sub-topics (e.g. ONNX -> ONNX on Flux Chains). If you already have an idea of what you're planning to write about/track on this issue thread, I can help you to migrate it to one or both of the places above. Then this issue can be closed so that we don't have two duplicate ones about ONNX support. |
Let's move to discussions then ;) |
Flux chains are easy to work with and most of the models used as starting points for ML tasks can be represented as Flux chains. So, a feature that would take the corresponding ONNX model and convert it directly to Flux chains would be useful.
This would especially be helpful for vision tasks for doing a quick test run of an architecture. Moreover, providing such a feature as part of Metalhead.jl can also be considered.
Such a feature (ONNX2Flux) would be limited in what models it can support (due to Flux chains) but can provide a good starting/testing point for pre-existing ONNX models with ease of use.
If interest exists in creating such a feature (ONNX2Flux.jl), I would be happy to work on it.
Since ONNXmutable.jl already does a good job at converting ONNX to CompGraph, adding support for conversion directly to Flux chains would use that.
Suggestions are welcome.
References:
This discussion thread #10
https://github.com/DrChainsaw/ONNXmutable.jl
The text was updated successfully, but these errors were encountered: