Coverage for tvo/utils/param_init.py: 0%

1# -*- coding: utf-8 -*- 

2# Copyright (C) 2019 Machine Learning Group of the University of Oldenburg. 

3# Licensed under the Academic Free License version 3.0 

4 

5import tvo 

6import numpy as np 

7import torch as to 

8from torch import Tensor 

9 

10 

11def init_W_data_mean( 

12 data: Tensor, 

13 H: int, 

14 std_factor: float = 0.25, 

15 dtype: to.dtype = to.float64, 

16 device: to.device = None, 

17) -> Tensor: 

18 """Initialize weights W based on noisy mean of the data points. 

19 

20 :param data: Data set, is (N, D). 

21 :param H: Number of basis functions to be generated. 

22 :param std_factor: Scalar to control amount of standard deviation of additive noise 

23 :param dtype: dtype of output Tensor. Defaults to torch.float64. 

24 :param device: torch.device of output Tensor. Defaults to tvo.get_device(). 

25 :returns: Weight matrix W with shape (D,H). 

26 """ 

27 device_ = tvo.get_device() if device is None else device 

28 data_nanmean = to.from_numpy(np.nanmean(data.detach().cpu().numpy(), axis=0)).to( 

29 dtype=dtype, device=device_ 

30 ) 

31 var = init_sigma2_default(data, dtype, device_) 

32 return data_nanmean.repeat((H, 1)).t() + std_factor * to.sqrt(var) * to.randn( 

33 (len(data_nanmean), H), dtype=dtype, device=device_ 

34 ) 

35 

36 

37def init_sigma2_default( 

38 data: Tensor, dtype: to.dtype = to.float64, device: to.device = None 

39) -> Tensor: 

40 """Initialize scalar sigma parameter based on variance of the data points. 

41 

42 :param data: Data set, is (N, D). 

43 :param dtype: dtype of output Tensor. Defaults to torch.float64. 

44 :param device: torch.device of output Tensor. Defaults to tvo.get_device(). 

45 :returns: Scalar sigma parameter. 

46 

47 Returns the mean of the variance in each dimension d=1,...,D. 

48 """ 

49 _device = tvo.get_device() if device is None else device 

50 var = to.from_numpy(np.nanvar(data.detach().cpu().numpy(), axis=0)).to( 

51 device=_device, dtype=dtype 

52 ) 

53 return to.mean(var) + to.tensor([0.001], device=_device, dtype=dtype) 

54 

55 

56def init_pies_default( 

57 H: int, crowdedness: float = 2.0, dtype: to.dtype = to.float64, device: to.device = None 

58): 

59 """Initialize pi parameter based on given crowdedness. 

60 

61 :param H: Length of pi vector. 

62 :param crowdedness: Average crowdedness corresponding to sum of elements in vector pi. 

63 :param dtype: dtype of output Tensor. Defaults to torch.float64. 

64 :param device: torch.device of output Tensor. Defaults to tvo.get_device(). 

65 :returns: Vector pi. 

66 """ 

67 

68 if device is None: 

69 device = tvo.get_device() 

70 return to.full((H,), fill_value=crowdedness / H, dtype=dtype, device=device)