[minor] Speed up dataloader

[leoniewgnr]

🔬 Background

We all know the fit function is super slow. This is due to some bottlenecks that are responsible for most of the slow code.

🔮 Key changes

Speed up the following bottlenecks

Training Dataloader

• 'init_after_tabularized' in time_dataset.py
• first for loop 'drop_nan_after_init' in time_dataset.py --> from 2 s to 0.00x s per call

Tune hyperparams and train

• bottlenecks to be found --> not bottlenecks found as completely in pytorch lightning, use accelerator option to speed up

IDs	Dataloader before	Dataloader after
8	33	12
16	69	19
32	112	37
64	200	84

[github-actions]

Model Benchmark

Benchmark	Metric	main	current	diff
YosemiteTemps	MAE_val	1.34899	1.3442	-0.36%	✅
YosemiteTemps	RMSE_val	2.00817	2.00245	-0.28%	✅
YosemiteTemps	Loss_val	0.00078	0.00077	-0.58%	✅
YosemiteTemps	MAE	1.32133	1.3192	-0.16%	✅
YosemiteTemps	RMSE	2.13713	2.13518	-0.09%	✅
YosemiteTemps	Loss	0.00064	0.00064	-0.17%	✅
YosemiteTemps	time	74.7397	112.07	49.95%	❌
PeytonManning	MAE_val	0.58162	0.58159	-0.0%	✅
PeytonManning	RMSE_val	0.72218	0.72216	-0.0%	✅
PeytonManning	Loss_val	0.01239	0.01239	-0.01%	✅
PeytonManning	MAE	0.41671	0.41671	0.0%	✅
PeytonManning	RMSE	0.55961	0.55961	0.0%	✅
PeytonManning	Loss	0.00612	0.00612	0.0%	✅
PeytonManning	time	15.8071	23.14	46.39%	❌
AirPassengers	MAE_val	13.0627	13.0626	-0.0%	✅
AirPassengers	RMSE_val	15.9453	15.9452	-0.0%	✅
AirPassengers	Loss_val	0.00131	0.00131	-0.0%	✅
AirPassengers	MAE	9.88153	9.88156	0.0%	✅
AirPassengers	RMSE	11.7354	11.7354	0.0%	✅
AirPassengers	Loss	0.00052	0.00052	0.0%	✅
AirPassengers	time	6.65026	9.74	46.46%	❌

Model training plots

Model Training

PeytonManning

YosemiteTemps

AirPassengers

[codecov]

Codecov Report

Merging #1360 (e15bb53) into main (d236471) will decrease coverage by 0.01%.
The diff coverage is 95.83%.

@@            Coverage Diff             @@
##             main    #1360      +/-   ##
==========================================
- Coverage   89.85%   89.85%   -0.01%     
==========================================
  Files          38       38              
  Lines        5068     5066       -2     
==========================================
- Hits         4554     4552       -2     
  Misses        514      514              

Files Changed	Coverage Δ
neuralprophet/time_dataset.py	95.51% <95.83%> (-0.03%)	⬇️

[LeonieFreisinger]

LGTM