[fix] Reduce saved model size

[leoniewgnr]

🔬 Background

See #1415

🔮 Key changes

removed attribute _trainer

[leoniewgnr]

I removed some attributes, but really unsure where the data is saved. In model.logger.history is saved some big tensor, but the model size is still not the same:

2023-09-19 13:04:06 [INFO]: size of saved fitted model: 288791 (test_utils.py:41)
2023-09-19 13:04:06 [INFO]: testing: save forecast model (test_utils.py:45)
2023-09-19 13:04:06 [INFO]: size of saved forecast model: 43850 (test_utils.py:47)

[codecov]

Codecov Report

Merging #1425 (dcb55d7) into main (b04e141) will decrease coverage by 0.04%.
The diff coverage is 85.00%.

@@            Coverage Diff             @@
##             main    #1425      +/-   ##
==========================================
- Coverage   89.88%   89.85%   -0.04%     
==========================================
  Files          38       38              
  Lines        5071     5084      +13     
==========================================
+ Hits         4558     4568      +10     
- Misses        513      516       +3     

Files Changed	Coverage Δ
neuralprophet/utils.py	79.55% <85.00%> (-0.10%)	⬇️

[github-actions]

Model Benchmark

Benchmark	Metric	main	current	diff
PeytonManning	MAE_val	0.58162	0.58162	0.0%	✅
PeytonManning	RMSE_val	0.72218	0.72218	0.0%	✅
PeytonManning	Loss_val	0.01239	0.01239	0.0%	✅
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	12.8696	13.27	3.11%	⚠️
AirPassengers	MAE_val	13.0627	13.0627	0.0%	✅
AirPassengers	RMSE_val	15.9453	15.9453	0.0%	✅
AirPassengers	Loss_val	0.00131	0.00131	0.0%	✅
AirPassengers	MAE	9.88153	9.88153	0.0%	✅
AirPassengers	RMSE	11.7354	11.7354	0.0%	✅
AirPassengers	Loss	0.00052	0.00052	0.0%	✅
AirPassengers	time	5.46486	5.62	2.84%	✅
YosemiteTemps	MAE_val	1.34899	1.34899	0.0%	✅
YosemiteTemps	RMSE_val	2.00817	2.00817	0.0%	✅
YosemiteTemps	Loss_val	0.00078	0.00078	0.0%	✅
YosemiteTemps	MAE	1.32133	1.32133	0.0%	✅
YosemiteTemps	RMSE	2.13713	2.13713	0.0%	✅
YosemiteTemps	Loss	0.00064	0.00064	0.0%	✅
YosemiteTemps	time	61.0225	62.58	2.55%	✅

Model training plots

Model Training

PeytonManning

YosemiteTemps

AirPassengers

[leoniewgnr]

I think I found it! In _trainer there was saved some unecessary information!
2023-09-19 13:51:06 [INFO]: size of saved fitted model: 18095 (test_utils.py:43)
2023-09-19 13:51:06 [INFO]: testing: save forecast model (test_utils.py:47)
2023-09-19 13:51:06 [INFO]: size of saved forecast model: 18180 (test_utils.py:50)

[ourownstory]

Good fix!

Do you think we should just include "_trainer" in attrs_to_remove = ["trainer"]?
Or is there a relevant scenario where "_trainer" needs to be saved, but "trainer" not?
If not, we can skip the 'minimal' option and keep the code more lean.

[leoniewgnr]

@ourownstory no probably not, I think so too, we should drop the minimal version and keep the code lean, will change that soon