Deployment concerns

[bob-mcrae]

We've used this excellent library for training using the InceptionTimePlus model and we're now looking to deploy saved models for inference in a production environment. For this inference, all we need is to load the model & make predictions; basically two library commands. The trouble I am currently facing is that it's taking a long time to load the library in what I thought was a minimal form. For example, the following command is taking a full 45-seconds on my local PC and ~35 seconds in Google Colab.
from tsai.all import load_learner

I realize that this import will only need to occur once per service startup, but I was hoping to make the service as light-weight and fast as possible. Are there any tips on improving the speed of what we're trying to do? Thanks in advance.

[oguiza]

Hi @bob-mcrae,
Thanks for creating this discussion.
I've created an issue based on it (#273)

[bob-mcrae]

...and, inference is working as expected.

VERY, VERY nice work Ignacio! Thank you!

[bob-mcrae]

Hi Ignacio,

I am working on my docker image now. Any idea on the timeline for the pip release that addressed this issue (273)? Thanks in advance.

[oguiza]

Hi Bob,
Either later today if I can find the time or tomorrow.

[oguiza]

Hi @bob-mcrae,
FYI, I've just created the new pip release (0.2.23).

[bob-mcrae]

Thanks so much Ignacio. Great work!

[oguiza]

Great! I'm glad it works as expected! 😀

I'll close this issue now. Please, reopen it if necessary.

And thanks a lot for bringing this up to my attention.