def propagate_broadcast(op, const_value_by_tensor, dtype=''):
# type: (TFOperation, _ConstValueByTensorT, str)->typing.Tuple[typing.List[typing.List[int]], typing.List[str]]
return ([
infer.elementwise(inputs=[input.shape for input in op.inputs],
broadcast=infer.Broadcast.FROM_RIGHT)
], [dtype if dtype else get_op_t(op)])
def propagate_broadcast(op, dtype_from_index=0):
# type: (ONNXOperation, int)->typing.Tuple[typing.List[typing.List[int]], typing.List[str]]
return ([
infer.elementwise(inputs=[input.shape for input in op.inputs],
broadcast=infer.Broadcast.FROM_RIGHT)
], [op.inputs[dtype_from_index].dtype])
def propagate_broadcast_with_axis(op, dtype=None):
# type: (ONNXOperation, typing.Optional[str])->typing.Tuple[typing.List[typing.List[int]], typing.List[str]]
x, y = op.inputs
if 'axis' in op.attribs:
axis = op.attribs['axis']
if axis < 0:
axis += x.rank
return ([
infer.elementwise(inputs=[x.shape, axis * [1] + y.shape],
broadcast=infer.Broadcast.FROM_LEFT)
], [x.dtype])
else:
return ([
infer.elementwise(inputs=[x.shape, y.shape],
broadcast=infer.Broadcast.FROM_RIGHT)
], [x.dtype if dtype is None else dtype])
def propagate_expand(op):
# type: (ONNXOperation)->typing.Tuple[typing.List[typing.List[int]], typing.List[str]]
input, shape = op.inputs
return [
infer.elementwise(
inputs=[input.shape,
evaluate_shape_tensor_simple(shape)],
broadcast=infer.Broadcast.FROM_RIGHT)
], [input.dtype]
def propagate_expand(op):
# type: (ONNXOperation)->typing.Tuple[typing.List[typing.List[int]], typing.List[str]]
input, shape = op.inputs
input_shape = list(input.shape)
other_shape = evaluate_shape_tensor_simple(shape)
if len(other_shape) > len(input_shape):
input_shape = [1] * (len(other_shape) - len(input_shape)) + input_shape
if len(input_shape) > len(other_shape):
other_shape = [1] * (len(input_shape) - len(other_shape)) + other_shape
# undocumented feature in ONNX, probably comes from pytorch
other_shape = [
i if o == -1 else o for i, o in zip(input_shape, other_shape)
]
op.attribs['shape'] = other_shape
op.inputs = (input, )
return [
infer.elementwise(inputs=[input_shape, other_shape],
broadcast=infer.Broadcast.SAME_RANK)
], [input.dtype]
def test_elementwise(self):
a = [1, 2, 3]
b = [1, 2, 3]
c = infer.elementwise([a, b], infer.Broadcast.NONE)
self.assertEqual(c, a)
self.assertIsNot(c, a)
self.assertIsNot(c, b)
a = [1, 2, 3]
b = [3]
ab = [a, b]
c = infer.elementwise([a, b], infer.Broadcast.FROM_RIGHT)
self.assertEqual([1, 2, 3], c)
self.assertEqual([1, 2, 3], a)
self.assertEqual([3], b)
self.assertEqual([a, b], ab)
with self.assertRaises(AssertionError):
infer.elementwise([[1, 2], [1]], infer.Broadcast.NONE)
with self.assertRaises(AssertionError):
infer.elementwise([[1, 2], [1, 3]], infer.Broadcast.NONE)
self.assertEqual([2, 4, 3], infer.elementwise([[1, 4, 1], [2, 1, 3]], infer.Broadcast.SAME_RANK))
with self.assertRaises(AssertionError):
infer.elementwise([[1, 2], [1]], infer.Broadcast.SAME_RANK)
self.assertEqual([4, 2, 3], infer.elementwise([[1, 2, 3], [4, 2]], infer.Broadcast.FROM_LEFT))
with self.assertRaises(AssertionError):
infer.elementwise([[2, 3], [3]], infer.Broadcast.FROM_LEFT)
self.assertEqual([1, 2, 3], infer.elementwise([[1, 1, 3], [2, 3]], infer.Broadcast.FROM_RIGHT))
with self.assertRaises(AssertionError):
infer.elementwise([[2, 3], [2]], infer.Broadcast.FROM_RIGHT)
