Coverage for anfis_toolbox / optim / __init__.py: 100%

1"""Optimization algorithms for ANFIS. 

2 

3This module contains pluggable training algorithms (optimizers/trainers) 

4that can be used with the ANFIS model. 

5 

6Design goals: 

7- Decouple training algorithms from the model class 

8- Keep a simple API similar to scikit-learn fit(X, y) 

9- Allow power users to instantiate and pass custom trainers 

10 

11Example: 

12 from anfis_toolbox.optim import SGDTrainer, RMSPropTrainer, HybridTrainer 

13 trainer = SGDTrainer(learning_rate=0.01, epochs=200) 

14 losses = trainer.fit(model, X, y) 

15 

16Task compatibility and guidance: 

17-------------------------------- 

18- HybridTrainer implements the original Jang (1993) hybrid learning and is intended 

19 for regression with the regression ANFIS (single-output). It is not compatible 

20 with the classification head. 

21 

22- SGDTrainer, RMSPropTrainer and AdamTrainer perform generic backprop updates and now 

23 accept pluggable loss functions (see ``anfis_toolbox.losses``). They default to mean 

24 squared error for regression, but can minimize other differentiable objectives such as 

25 categorical cross-entropy when used with ``ANFISClassifier``. Targets are adapted via the 

26 selected loss' ``prepare_targets`` helper, so integer labels or one-hot matrices are both 

27 supported seamlessly. 

28""" 

29 

30from .adam import AdamTrainer 

31from .base import BaseTrainer 

32from .hybrid import HybridTrainer 

33from .hybrid_adam import HybridAdamTrainer 

34from .pso import PSOTrainer 

35from .rmsprop import RMSPropTrainer 

36from .sgd import SGDTrainer 

37 

38__all__ = [ 

39 "BaseTrainer", 

40 "SGDTrainer", 

41 "HybridTrainer", 

42 "AdamTrainer", 

43 "HybridAdamTrainer", 

44 "RMSPropTrainer", 

45 "PSOTrainer", 

46]