[Fix] loading model on machine with different accelerator

[McOffsky]

🔬 Background

During my experiments with NP I've found out, that utils.load() is using trainer config is retrived from model, including accelerator. This caused an error on cpu machine while loading model fitted on GPU.

🔮 Key changes

• NeuralProphet.restore_trainer() now takes optional argument, accelerator: str = None. If not provided, value stored in model will be used.

• utils.load() now creates torch.device(map_location) into separate variable, and that variable is used in torch.load(). Orginal value is additionaly passed to NeuralProphet.restore_trainer() as accelerator

📋 Review Checklist

• I have performed a self-review of my own code.
• I have commented my code, added docstrings and data types to function definitions.
• [n/c] I have added pytests to check whether my feature / fix works.

Please make sure to follow our best practices in the Contributing guidelines.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (9493458) 88.33% compared to head (65f3289) 88.41%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1509      +/-   ##
==========================================
+ Coverage   88.33%   88.41%   +0.08%     
==========================================
  Files          38       38              
  Lines        5099     5103       +4     
==========================================
+ Hits         4504     4512       +8     
+ Misses        595      591       -4     

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

[ourownstory]

Great addition, thank you!

[ourownstory]

Can be merged after passing of tests.

[McOffsky]

Happy to help (a bit) with such a great library!
I'll fix the issues tomorrow, it'a a bit late in my timezone ;)

[ourownstory]

Hi @McOffsky Thank you!
The failed tests are not related to your contribution, we can safely ignore them.

The pyright type checker did want a change of line 2737, I fixed it quickly - save your efforts for a later contribution.
(For your reference you can read about the type logic here: https://stackoverflow.com/questions/51710037/how-should-i-use-the-optional-type-hint)

Best,
Oskar

[ourownstory]

Update: waiting for rerun of tests to finish before merge.

[github-actions]

Model Benchmark

Benchmark	Metric	main	current	diff
AirPassengers	MAE_val	31.6338	31.6337	-0.0%	✅
AirPassengers	RMSE_val	32.6669	32.6669	-0.0%	✅
AirPassengers	Loss_val	0.01372	0.01372	-0.0%	✅
AirPassengers	MAE	6.24137	6.24136	-0.0%	✅
AirPassengers	RMSE	7.82929	7.82928	-0.0%	✅
AirPassengers	Loss	0.00063	0.00063	-0.0%	✅
AirPassengers	time	10.6782	10.72	0.39%	✅
YosemiteTemps	MAE_val	0.64178	0.64178	0.0%	✅
YosemiteTemps	RMSE_val	0.92787	0.92787	0.0%	✅
YosemiteTemps	Loss_val	0.0005	0.0005	0.0%	✅
YosemiteTemps	MAE	1.092	1.092	0.0%	✅
YosemiteTemps	RMSE	1.89124	1.89124	0.0%	✅
YosemiteTemps	Loss	0.00155	0.00155	0.0%	✅
YosemiteTemps	time	59.4511	62.74	5.53%	⚠️
EnergyPriceDaily	MAE_val	5.55889	5.55889	-0.0%	✅
EnergyPriceDaily	RMSE_val	7.12233	7.12233	-0.0%	✅
EnergyPriceDaily	Loss_val	0.02829	0.02829	-0.0%	✅
EnergyPriceDaily	MAE	6.37172	6.37172	-0.0%	✅
EnergyPriceDaily	RMSE	8.53016	8.53016	-0.0%	✅
EnergyPriceDaily	Loss	0.02907	0.02907	0.0%	✅
EnergyPriceDaily	time	15.2592	15.59	2.17%	✅
PeytonManning	MAE_val	0.3466	0.3466	0.0%	✅
PeytonManning	RMSE_val	0.49808	0.49808	0.0%	✅
PeytonManning	Loss_val	0.01752	0.01752	0.0%	✅
PeytonManning	MAE	0.34735	0.34735	0.0%	✅
PeytonManning	RMSE	0.48788	0.48788	0.0%	✅
PeytonManning	Loss	0.01457	0.01457	-0.0%	✅
PeytonManning	time	20.8968	22.12	5.85%	⚠️

Model training plots

Model Training

PeytonManning

YosemiteTemps

AirPassengers

EnergyPriceDaily