Machinelearning: Printing training statistics by default discussion

Created on 24 Oct 2018 · 4Comments · Source: dotnet/machinelearning

Before the ML.Net Environment changes to Local environment, PR #923 during training of a pipeline, we would see output in the console.

I am trying a long-ish to train model today; takes about 4 minutes, and I find the absence of any progress reporting unsettling. I am not sure whether things are working as expected, or working at all.
I think this will be off-putting for the users.

API usability

Source

sfilipi

Most helpful comment

I very much like having the intermediate progress output.

Specifically, this gives me:

Percent done (or a way to guess)
Indication of _any_ progress
Knowledge of which component is taking time
Warnings/errors insitu

I would like us to go a bit further and have all trainable components output a more useful friendly (ComponentX ##.#% complete). This is only doable for components with known end points and not components which run until convergence. This also gets a bit nasty with threading.

justinormont on 25 Oct 2018

👍3

All 4 comments

I very much like having the intermediate progress output.

Specifically, this gives me:

Percent done (or a way to guess)
Indication of _any_ progress
Knowledge of which component is taking time
Warnings/errors insitu

justinormont on 25 Oct 2018

👍3

@Zruty0 since you authored the #923 PR, do you know if the change was intentional?

artidoro on 27 Oct 2018

It was an intentional change to NOT write to console willy-nilly.
Right now, ML.NET's MLContext doesn't write to console.

If you do mlContext.Log += Console.WriteLine, we will have the output back in the console, but in a non-colorized format that is more applicable for parsing than for human readability.

I am not sure whether the colorized pretty console printing is something we want to have in the library at all.