openwhisk-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From James Thomas <>
Subject Re: AI Actions as a first-class citizen in OpenWhisk
Date Tue, 14 Aug 2018 14:39:12 GMT

Great wiki page! Since I've been playing with TensorFlow.js on OpenWhisk, I
definitely think there's a sweet spot for running certain ML tasks using
serverless platforms. The suggestions in the wiki page all make sense to me.

Access to the GPU is one of the biggest barriers to using more complex
models and operations (and even training longer-term). Nvida does have a
Docker version that allows you to pass through GPUs (

Another issue I found was the performance difference between warm and cold
activations when using customised run images. This could have been resolved
it either there was a pre-warmed Node.js image with TF-JS libraries or we
have a mechanism to create action packages larger than 48MB. This might
include creating actions from packages at an external HTTP address or
object storage URI.

On 13 August 2018 at 22:17, Dragos Dascalita Haut <
> wrote:

> Once you've experienced FaaS, you don't wanna go back. This has been my
> experience with AI and FaaS.
> In particular, running AI inferences in FaaS proved to be a great match:
> - Each function processes one request at a time. A model usually takes 1
> data input and produces 1 data output.
> - Enough code to fit into a function. An AI action loads a model, runs the
> inference, and returns the result.
> - In addition, FaaS provides a model to scale to 0 and scale to millions
> with the traffic.
> With OpenWhisk I think we're very close to make AI Actions a first-class
> citizen for developers, and I've created a wiki to explore what it would
> take to get there [1].  Coincidently James Thomas also published today his
> experience with Tensorflow and OpenWhisk [2]
> I'm interested in your thoughts, and see if there's enough interest in our
> community to make this a reality.
> Feel free to contribute to the wiki with edits, comments, anything you'd
> wanna add.
> [1] -
> [2] -
> with-tensorflow-js-4aa24494a9b4
> Thanks,
> dragos

James Thomas

  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message