Aerosol activation NN

[trontrytel]

This PR brings in the first part of the work of @mikhailmints from last year SURF. It:

• Creates an extension for the CM function N_activated_per_mode that will be available if the user has the following packages: DataFrames, Flux, MLJ, MLJFlux, CSV, DataFramesMeta, MLJModels
• Adds a tests that trains a NN model based on PySDM training data and compares the results with the default function.
• It would be nice to have fewer packages here (maybe we could write it without DataFrames? But lets refactor that once we have all models working first.)

If the jls file with the NN model already exists the training is skipped.

[nefrathenrici]

I'm not sure about the MLJ/Flux specifics, but this looks like a good working start.

A stretch goal might be to get rid of DataFrames as a dependency, I am not sure it is necessary

[codecov]

Codecov Report

Attention: Patch coverage is 85.41667% with 7 lines in your changes are missing coverage. Please review.

Project coverage is 98.11%. Comparing base (13b011d) to head (d38910d).
Report is 1 commits behind head on main.

Files	Patch %	Lines
ext/Common.jl	82.05%	7 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #318      +/-   ##
==========================================
- Coverage   98.65%   98.11%   -0.55%     
==========================================
  Files          33       35       +2     
  Lines        1117     1165      +48     
==========================================
+ Hits         1102     1143      +41     
- Misses         15       22       +7     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[trontrytel]

I have applied all the comments. Thank you @nefrathenrici for all the help with the extensions.

Every now and then I get this warning when training the NN. Do you see where the Float64 is coming in?

[ Info: Training a new ML model. This may take a minute or two.
┌ Warning: Layer with Float32 parameters got Float64 input.
│   The input will be converted, but any earlier layers may be very slow.
│   layer = Dense(11 => 250, sigmoid_fast)  # 3_000 parameters
│   summary(x) = "11×1000 Matrix{Float64}"

[trontrytel]

Seems like codecov gets confused about the code inside the ext folder. I guess there is no easy way to fix it?

[nefrathenrici]

This looks good! I think DataFrames really shouldn't be needed here, I can try and remove it when I find the time next week.

I think the Codecov just ignores extensions, I have not been able to find any way around it.

[trontrytel]

This looks good! I think DataFrames really shouldn't be needed here, I can try and remove it when I find the time next week.

Thanks! That would be great, if you have the time.

I think the Codecov just ignores extensions, I have not been able to find any way around it.

So sad. I'm worried my codecov stats will suffer and I will not be able to brag about test coverage anymore. Could we open an issue somewhere on the intercept of codecov and julia extensions communities (assuming it's not empty)...