├── README.md
└── GAF.py


/README.md:
--------------------------------------------------------------------------------
1 | # Gramian-Angular-Field-in-Pytorch
2 | A implementation of Gramian Angular Field (GAF) in Pytorch.
3 | 


--------------------------------------------------------------------------------
/GAF.py:
--------------------------------------------------------------------------------
 1 | import torch
 2 | class GramianAngularFieldPytorch(object):
 3 |     """Gramian Angular Field.
 4 | 
 5 |     Parameters
 6 |     ----------
 7 |     method : 'summation' or 'difference' (default = 'summation')
 8 |         Type of Gramian Angular Field. 's' can be used for 'summation'
 9 |         and 'd' for 'difference'.
10 | 
11 |     References
12 |     ----------
13 |     .. [1] Z. Wang and T. Oates, "Encoding Time Series as Images for Visual
14 |            Inspection and Classification Using Tiled Convolutional Neural
15 |            Networks." AAAI Workshop (2015).
16 | 
17 |     """
18 | 
19 |     def __init__(self, method='summation'):
20 |         self.method = method
21 | 
22 |     def min_max_norm(self, X):
23 |         min_val = torch.min(X, dim=-1, keepdim=True)[0]
24 |         max_val = torch.max(X, dim=-1, keepdim=True)[0]
25 |         res = (X - min_val) / (max_val - min_val)
26 |         return res * 2 - 1
27 | 
28 |     @staticmethod
29 |     def _gasf(X_cos, X_sin):
30 | 
31 |         X_cos_L = X_cos.unsqueeze(-1)
32 |         X_cos_R = X_cos.unsqueeze(-2)
33 | 
34 |         X_sin_L = X_sin.unsqueeze(-1)
35 |         X_sin_R = X_sin.unsqueeze(-2)
36 | 
37 |         X_gasf = torch.bmm(X_cos_L, X_cos_R) - torch.bmm(X_sin_L, X_sin_R)
38 | 
39 |         return X_gasf
40 | 
41 |     @staticmethod
42 |     def _gadf(X_cos, X_sin):
43 | 
44 |         X_sin_L = X_sin.unsqueeze(-1)
45 |         X_cos_R = X_cos.unsqueeze(-2)
46 | 
47 |         X_cos_L = X_cos.unsqueeze(-1)
48 |         X_sin_R = X_sin.unsqueeze(-2)
49 | 
50 |         X_gadf = torch.bmm(X_sin_L, X_cos_R) - torch.bmm(X_cos_L, X_sin_R)
51 | 
52 |         return X_gadf
53 | 
54 |     def transform(self, X):
55 |         """Transform each time series into a GAF image.
56 | 
57 |         Parameters
58 |         ----------
59 |         X : array-like, shape = (n_samples, n_timestamps)
60 | 
61 |         Returns
62 |         -------
63 |         X_new : array-like, shape = (n_samples, image_size, image_size)
64 |             Transformed data. If ``flatten=True``, the shape is
65 |             `(n_samples, image_size * image_size)`.
66 | 
67 |         """
68 |         X_cos = self.min_max_norm(X)
69 |         X_sin = torch.sqrt(torch.clip(1 - X_cos ** 2, 0, 1))
70 |         if self.method in ['s', 'summation']:
71 |             X_new = self._gasf(X_cos, X_sin)
72 |         else:
73 |             X_new = self._gadf(X_cos, X_sin)
74 |         return X_new


--------------------------------------------------------------------------------