[GSK-1617] ModelCache does not load cached results correctly on Windows

[mattbit]

This is caused by bad formatting of newlines in csv read/write, see here.

Fixing also a series of other issues:

• Fix cache loading when cache is disabled
• Fix broken cache for text generation models (and added a test)
• Fix broken cache path on Windows
• Fix broken CSV writing on Windows
• Ensure regression models output is casted to float

[linear]

GSK-1617 Bug with ModelCache on Windows

Model cache fails on Windows with regression models, breaking all models.

[andreybavt]

Why not warm up at creation time instead of read_from_cache ?

[mattbit]

Because I don’t want to warm up if the cache is disabled. But it’s still better to create the ModelCache object in the model, so that it can be used if the cache is enabled at a later time.

[andreybavt]

I think in case cache is disabled the cleanest will be not to initialize _cache property at all and then warm it up in the ModelCache constructor depending on a cache type (in memory only vs FS backed), WDYT?

[mattbit]

But cache can be disabled temporarily. So if we don’t initialize the _cache property at all, we would need to check if it exists at every prediction, initializing it there and possibly warm it up at prediction time, in addition to checking that cache is enabled. That’s because cache could have been disabled when model was initialized, but enabled later.

In any case I would avoid doing expensive and persistent operations upon construction of the instance.

[andreybavt]

ok let's keep it this way

[sonarqubecloud]

Kudos, SonarCloud Quality Gate passed!   

0 Bugs
0 Vulnerabilities
0 Security Hotspots
0 Code Smells

94.7% Coverage
0.0% Duplication