Coverage for /builds/ase/ase/ase/utils/abc.py: 100.00%

1from abc import ABC, abstractmethod 

2 

3import numpy as np 

4 

5# Due to the high prevalence of cyclic imports surrounding ase.optimize, 

6# we define the Optimizable ABC here in utils. 

7# Can we find a better way? 

8 

9 

10class Optimizable(ABC): 

11 @abstractmethod 

12 def ndofs(self) -> int: 

13 """Return number of degrees of freedom.""" 

14 

15 @abstractmethod 

16 def get_x(self) -> np.ndarray: 

17 """Return current coordinates as a flat ndarray.""" 

18 

19 @abstractmethod 

20 def set_x(self, x: np.ndarray) -> None: 

21 """Set flat ndarray as current coordinates.""" 

22 

23 @abstractmethod 

24 def get_gradient(self) -> np.ndarray: 

25 """Return gradient at current coordinates as flat ndarray. 

26 

27 NOTE: Currently this is still the (flat) "forces" i.e. 

28 the negative gradient. This must be fixed before the optimizable 

29 API is done.""" 

30 # Callers who want Nx3 will do ".get_gradient().reshape(-1, 3)". 

31 # We can probably weed out most such reshapings. 

32 # Grep for the above expression in order to find places that should 

33 # be updated. 

34 

35 @abstractmethod 

36 def get_value(self) -> float: 

37 """Return function value at current coordinates.""" 

38 

39 @abstractmethod 

40 def iterimages(self): 

41 """Yield domain objects that can be saved as trajectory. 

42 

43 For example this can yield Atoms objects if the optimizer 

44 has a trajectory that can write Atoms objects.""" 

45 

46 def converged(self, gradient: np.ndarray, fmax: float) -> bool: 

47 """Standard implementation of convergence criterion. 

48 

49 This assumes that forces are the actual (Nx3) forces. 

50 We can hopefully change this.""" 

51 assert gradient.ndim == 1 

52 return self.gradient_norm(gradient) < fmax 

53 

54 def gradient_norm(self, gradient): 

55 forces = gradient.reshape(-1, 3) # XXX Cartesian 

56 return np.linalg.norm(forces, axis=1).max() 

57 

58 def __ase_optimizable__(self) -> 'Optimizable': 

59 """Return self, being already an Optimizable.""" 

60 return self