Exporting portable models

[bvanessen]

Lets add support for exporting a trained model into a portable network format. This should enable us to pre-train a network, export it, and allow someone to fine tune it in a different toolkit.

[adammoody]

Haven't had a chance to look at it deeply, but ONNX seems to have good support so far:
https://research.fb.com/onnx-v1-released/

[davidHysom]

This seems interesting ... but it seems so recent (early release: Sept of this year) that I would hesitate to commit much energy. If it's a going thing six months from now, maybe. I think most or all of the interoperability API efforts are or will be python based. Original post, Brian: "Lets add support for exporting a trained model into a portable network format" Am thinking we shouldn't roll-our-own at this point in time, or jump on any bandwagon. Unless there's a motivating (political) need.

…

________________________________ From: Adam Moody <[email protected]> Sent: Thursday, December 21, 2017 10:59:02 AM To: LLNL/lbann Cc: Subscribed Subject: Re: [LLNL/lbann] Exporting portable models (#151) Haven't had a chance to look at it deeply, but ONNX seems to have good support so far: https://research.fb.com/onnx-v1-released/ — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub<#151 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AI8DH2aXlZuhiV_l7Q5BesKA3RaAGtaOks5tCqp1gaJpZM4RJf_6>.

[bvanessen]

At the moment all I want is a numpy array per weight matrix that can be loaded into TensorFlow or something like that. Brian C. Van Essen [email protected]<mailto:[email protected]> (w) 925-422-9300 (c) 925-290-5470

…

---

Sent from my iPhone On Dec 21, 2017, at 5:12 PM, davidHysom <[email protected]<mailto:[email protected]>> wrote: This seems interesting ... but it seems so recent (early release: Sept of this year) that I would hesitate to commit much energy. If it's a going thing six months from now, maybe. I think most or all of the interoperability API efforts are or will be python based. Original post, Brian: "Lets add support for exporting a trained model into a portable network format" Am thinking we shouldn't roll-our-own at this point in time, or jump on any bandwagon. Unless there's a motivating (political) need.

________________________________ From: Adam Moody <[email protected]<mailto:[email protected]>> Sent: Thursday, December 21, 2017 10:59:02 AM To: LLNL/lbann Cc: Subscribed Subject: Re: [LLNL/lbann] Exporting portable models (#151) Haven't had a chance to look at it deeply, but ONNX seems to have good support so far: https://research.fb.com/onnx-v1-released/ — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub<#151 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AI8DH2aXlZuhiV_l7Q5BesKA3RaAGtaOks5tCqp1gaJpZM4RJf_6>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#151 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AF7Ce6EIhuV-Wk-e4Z2iRYfUTkygFydJks5tCwIJgaJpZM4RJf_6>.