Coverage for mlprodict/onnxrt/ops_cpu/op_gru.py: 97%

1# -*- encoding: utf-8 -*- 

2# pylint: disable=E0203,E1101,C0111 

3""" 

4@file 

5@brief Runtime operator. 

6""" 

7import numpy 

8from ._op import OpRun 

9 

10 

11class CommonGRU(OpRun): 

12 

13 def __init__(self, onnx_node, expected_attributes=None, desc=None, 

14 **options): 

15 OpRun.__init__(self, onnx_node, desc=desc, 

16 expected_attributes=expected_attributes, 

17 **options) 

18 self.nb_outputs = len(onnx_node.output) 

19 self.number_of_gates = 3 

20 

21 def f(self, x): 

22 return 1 / (1 + numpy.exp(-x)) 

23 

24 def g(self, x): 

25 return numpy.tanh(x) 

26 

27 def _step(self, X, R, B, W, H_0): 

28 seq_length = X.shape[0] 

29 hidden_size = H_0.shape[-1] 

30 batch_size = X.shape[1] 

31 

32 Y = numpy.empty( 

33 [seq_length, self.num_directions, batch_size, hidden_size]) 

34 h_list = [] 

35 

36 [w_z, w_r, w_h] = numpy.split(W, 3) # pylint: disable=W0632 

37 [r_z, r_r, r_h] = numpy.split(R, 3) # pylint: disable=W0632 

38 [w_bz, w_br, w_bh, r_bz, r_br, r_bh] = numpy.split( # pylint: disable=W0632 

39 B, 6) # pylint: disable=W0632 

40 gates_w = numpy.transpose(numpy.concatenate((w_z, w_r))) 

41 gates_r = numpy.transpose(numpy.concatenate((r_z, r_r))) 

42 gates_b = numpy.add(numpy.concatenate((w_bz, w_br)), 

43 numpy.concatenate((r_bz, r_br))) 

44 

45 H_t = H_0 

46 for x in numpy.split(X, X.shape[0], axis=0): 

47 gates = numpy.dot(x, gates_w) + numpy.dot(H_t, gates_r) + gates_b 

48 z, r = numpy.split(gates, 2, -1) # pylint: disable=W0632 

49 z = self.f(z) 

50 r = self.f(r) 

51 h_default = self.g(numpy.dot(x, numpy.transpose( 

52 w_h)) + numpy.dot(r * H_t, numpy.transpose(r_h)) + w_bh + r_bh) 

53 h_linear = self.g(numpy.dot(x, numpy.transpose( 

54 w_h)) + r * (numpy.dot(H_t, numpy.transpose(r_h)) + r_bh) + w_bh) 

55 h = h_linear if self.linear_before_reset else h_default 

56 H = (1 - z) * h + z * H_t 

57 h_list.append(H) 

58 H_t = H 

59 

60 concatenated = numpy.concatenate(h_list) 

61 if self.num_directions == 1: 

62 Y[:, 0, :, :] = concatenated 

63 

64 if self.layout == 0: 

65 Y_h = Y[-1] 

66 else: 

67 Y = numpy.transpose(Y, [2, 0, 1, 3]) 

68 Y_h = Y[:, :, -1, :] 

69 

70 return Y, Y_h 

71 

72 def _run(self, X, W, R, B=None, attributes=None, sequence_lens=None, # pylint: disable=W0221 

73 initial_h=None, verbose=0, fLOG=None): 

74 self.num_directions = W.shape[0] 

75 

76 if self.num_directions == 1: 

77 R = numpy.squeeze(R, axis=0) 

78 W = numpy.squeeze(W, axis=0) 

79 if B is not None: 

80 B = numpy.squeeze(B, axis=0) 

81 if sequence_lens is not None: 

82 sequence_lens = numpy.squeeze(sequence_lens, axis=0) 

83 if initial_h is not None: 

84 initial_h = numpy.squeeze(initial_h, axis=0) 

85 

86 hidden_size = R.shape[-1] 

87 batch_size = X.shape[1] 

88 

89 b = (B if B is not None else 

90 numpy.zeros(2 * self.number_of_gates * hidden_size, dtype=X.dtype)) 

91 h_0 = (initial_h if initial_h is not None else 

92 numpy.zeros((batch_size, hidden_size), dtype=X.dtype)) 

93 

94 B = b 

95 H_0 = h_0 

96 else: 

97 raise NotImplementedError( # pragma: no cover 

98 "Unsupported value %r for num_directions and operator %r." % ( 

99 self.num_directions, self.__class__.__name__)) 

100 

101 Y, Y_h = self._step(X, R, B, W, H_0) 

102 

103 return (Y, ) if self.nb_outputs == 1 else (Y, Y_h) 

104 

105 

106class GRU(CommonGRU): 

107 

108 atts = { 

109 'activation_alpha': [0.], 

110 'activation_beta': [0.], 

111 'activations': [b'Tanh', b'Tanh'], 

112 'clip': [], 

113 'direction': b'forward', 

114 'hidden_size': None, 

115 'layout': 0, 

116 'linear_before_reset': 0, 

117 } 

118 

119 def __init__(self, onnx_node, desc=None, **options): 

120 CommonGRU.__init__(self, onnx_node, desc=desc, 

121 expected_attributes=GRU.atts, 

122 **options)