search

ablaom / MLJTutorial.jl

Notebooks for introducing the machine learning toolbox MLJ (Machine Learning in Julia)
MIT License
25 stars 3 forks source link

Pluto notebook for part 02 adapted - is now functional #6

Closed roland-KA closed 2 years ago

roland-KA commented 2 years ago

Hi @ablaom, now the notebook for part 02 is ready.

There are a few points to look out for in this part:

roland-KA commented 2 years ago

... and perhaps you leave this pull request open, so I can deliver the remaining parts on this basis?

roland-KA commented 2 years ago

Part 03 ist also ready:

roland-KA commented 2 years ago

In part 02 I've changed the code in "Warm restart", so that the output of the fit! operations shows in the notebook (and not in the terminal, where Pluto has been started).

ablaom commented 2 years ago

Hi @ablaom, now the notebook for part 02 is ready.

Awesome, many thanks. They look very nice.

There are a few points to look out for in this part:

  • In 'Step 1', after loading the Iris dataset, there is no need to coerce the data. It seems that CategoricalArrays became smarter. When I did the conversion of the original tutorial about a week ago, there was still the need to do it. I left a comment in the notebook about this change (in case people are watching the YouTube video from 2020).

Thanks for flagging. I've fixed this in the other notebooks as well.

  • In 'Warm start' I'm using the machine mach for further processing (not mach3). I think it actually doesn't matter which of both machines to use, but as Pluto has some problems with retrieving a machine from disk (see the note in 'Save and retrieve a machine'), it's safer to use mach here. And in case you wonder: If have inserted a few sub-titles in this section, to make the notebook more accessible from the TOC.

👍🏾

  • Please have a look on the calculation of misclassification_rate in this notebook when you test it. On one run it calculated the value 0; on further runs the values seemed to be reasonable. I don't know why it returned 0 on the first run.

The model is not reproducible and so this is not totally unexpected. I don't think it's a bug. Unfortunately we can't make this reproducible without using a different Flux chain builder, because you cannot currently pass an RNG to Flux.dropout. But I've flagged this here, thanks!

Unfortunately, the new pipeline type does not support type of transformations of the target, so I deleted that part of the tutorial, and added an explanation. Feel free to tweak this if you like. The plan is to have a separate model wrapper to handle target transformations. I guess I need to make this a higher priority 😄 .

I got some merge conflict headaches, and so am closing this in favour of #10, which I am going to merge. Probably best to make new PR's for each section anyway. I won't publicise the pluto tutorials too much until we get the pipeline stuff finished.

roland-KA commented 2 years ago

More people seem to run in that macro-related problem with Pluto when using @pipeline (see: Pluto.jl #1658). I've added a comment within that issue, because I think it's also a general Pluto problem, that has to be resolved (since the code runs outside of Pluto without errors).

So it would be nice, if we could see progress on both fields 😊.