Coverage for mlprodict/onnxrt/ops_cpu/op_conv.py: 85%

1# -*- encoding: utf-8 -*- 

2# pylint: disable=E0203,E1101,C0111 

3""" 

4@file 

5@brief Runtime operator. 

6""" 

7import numpy 

8from ._op import OpRun 

9from ..shape_object import ShapeObjectFct 

10from .op_conv_ import ConvFloat, ConvDouble # pylint: disable=E0611,E0401 

11 

12 

13class Conv(OpRun): 

14 

15 atts = {'auto_pad': 'NOTSET', 'group': 1, 

16 'dilations': [1, 1], 

17 'kernel_shape': [], 

18 'pads': [], 

19 'strides': [1, 1]} 

20 

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

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

23 expected_attributes=Conv.atts, 

24 **options) 

25 self._init() 

26 

27 def _init(self): 

28 self.rt32_ = ConvFloat() 

29 self.rt64_ = ConvDouble() 

30 for rt in [self.rt32_, self.rt64_]: 

31 rt.init(self.auto_pad, 

32 numpy.array(self.dilations, dtype=numpy.int64), 

33 self.group, 

34 numpy.array(self.kernel_shape, dtype=numpy.int64), 

35 numpy.array(self.pads, dtype=numpy.int64), 

36 numpy.array(self.strides, dtype=numpy.int64)) 

37 

38 def _run(self, X, W, B=None): # pylint: disable=W0221 

39 if X is None: 

40 raise ValueError( # pragma: no cover 

41 "X cannot be None for operator %r, ONNX=%r" % ( 

42 type(self), self.onnx_node)) 

43 if min(X.shape) == 0: 

44 raise RuntimeError( # pragma: no cover 

45 "Unable to run operator Conv on an empty matrix. " 

46 "X.shape=%r." % (X.shape, )) 

47 if min(W.shape) == 0: 

48 raise RuntimeError( # pragma: no cover 

49 "Unable to run operator Conv on an empty matrix. " 

50 "W.shape=%r." % (W.shape, )) 

51 if B is not None and min(B.shape) == 0: 

52 raise RuntimeError( # pragma: no cover 

53 "Unable to run operator Conv on an empty matrix. " 

54 "B.shape=%r." % (B.shape, )) 

55 if X.dtype == numpy.float32: 

56 return (self.rt32_.compute(X, W, B), ) 

57 return (self.rt64_.compute(X, W, B), ) 

58 

59 def _infer_shapes(self, X, W, B=None): # pylint: disable=W0221 

60 

61 def compute_shape(xshape, wshape, bshape): 

62 xs = numpy.ones(xshape, dtype=numpy.float32) 

63 ws = numpy.ones(wshape, dtype=numpy.float32) 

64 bs = (numpy.ones(bshape, dtype=numpy.float32) 

65 if bshape is not None else None) 

66 res = self.rt32_.compute(xs, ws, bs) 

67 return res.shape 

68 

69 return (ShapeObjectFct( 

70 compute_shape, X, W, B, name="Conv", dtype=X.dtype), ) 

71 

72 def _infer_types(self, X, W, B=None): # pylint: disable=W0221 

73 return (X, ) 

74 

75 def _infer_sizes(self, X, W, B=None): # pylint: disable=W0221 

76 res = self.run(X, W, B=None) 

77 C = X.shape[1] 

78 kernel_size = numpy.prod(self.kernel_shape) 

79 kernel_dim = C / self.group * kernel_size 

80 temp = kernel_dim * res[0].size 

81 return (dict(temp=temp * X.dtype.itemsize), ) + res