singa-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Cliff Berg (JIRA)" <>
Subject [jira] [Commented] (SINGA-333) Add support for Open Neural Network Exchange (ONNX) format
Date Mon, 23 Oct 2017 12:04:00 GMT


Cliff Berg commented on SINGA-333:

Thanks for this info Wangwei. I will use the Python API. I accept your assessment that using
the Python APIs is easier for a complex network, but it surprises me. I would have expected
that a declarative specification would be easier to design, examine, and adjust, and it seemed
to me that the gRPC message types were a declarative specification.

I am from an earlier generation of programmers - I started during the 1980s - and back then
there was a strong cultural preference for defining languages. (I myself was on the team that
designed the VHDL language.) The approach would have been to define a neural network specification
language. Indeed, Googling that, I see that there are some (e.g., EpsiloNN). I have not used
any of those, so I can't comment on them. I suppose that is the intention of ONNX. The advantages
of that approach are that (1) you can analyze designs, using static analysis tools, (2) you
can build a grammar-defined (i.e., standard object model) AST and optimize a design using
an optimization algorithm, (3) you can explain a design, by explaining its elements in the
language syntax or the standard object model, and (4) the design is statically check-able
- you don't have to run it in order to discover that you are missing an attribute.

I therefore hope that ONNX or something statically analyzable will be considered.

Very best, - Cliff

> Add support for Open Neural Network Exchange (ONNX) format
> ----------------------------------------------------------
>                 Key: SINGA-333
>                 URL:
>             Project: Singa
>          Issue Type: New Feature
>            Reporter: Moaz Reyad
> The [Open Neural Network Exchange (ONNX)|] is a format for interchanging
neural network models between AI systems. 
> The ONNX code is available here:
> It can be a useful feature for SINGA to support this format. However, ONNX is graph based
and some extra work will be required to convert SINGA models to this format. May be some core
parts of SINGA must to be upgraded first.
> This issue is open for studying and discussing the proposed feature of supporting ONNX.

This message was sent by Atlassian JIRA

View raw message