def test_split(self):
self.assertEqual([[1, 2, 2], [1, 2, 2], [1, 2, 2]], infer.split([1, 2, 6], -1, num=3))
self.assertEqual([[1, 2, 2], [1, 2, 4]], infer.split([1, 2, 6], -1, ratios=[1, 2]))
self.assertEqual([[1, 2, 2], [1, 2, 4]], infer.split([1, 2, 6], -1, sizes=[2, 4]))
self.assertEqual([[1, 2, 2], [1, 2, 4]], infer.split([1, 2, 6], -1, split_points=[2]))
self.assertEqual([[1, 2, 2], [1, 2, 1], [1, 2, 2], [1, 2, 1]],
infer.split([1, 2, 6], -1, split_points=[2, 3, 5]))
def propagate_split(op):
# type: (ONNXOperation)->typing.Tuple[typing.List[typing.List[int]], typing.List[str]]
if 'split' in op.attribs:
return (infer.split(input=op.input.shape,
axis=op.attribs.get('axis', 0),
sizes=op.attribs['split']),
[op.input.dtype] * len(op.outputs))
else:
return (infer.split(input=op.input.shape,
axis=op.attribs.get('axis', 0),
num=len(op.outputs)),
[op.input.dtype] * len(op.outputs))
def propagate_splitv(op, const_value_by_tensor):
# type: (TFOperation, _ConstValueByTensorT)->typing.Tuple[typing.List[typing.List[int]], typing.List[str]]
input, size_splits, axis = op.inputs
axis = const_value_by_tensor[axis].item()
return (infer.split(input=input.shape,
axis=axis,
num=op.attribs['num_split']),
[op.attribs['T']] * op.attribs['num_split'])
def split_shape(op):
# type: (Caffe2Operation)->ShapeResult
if len(op.inputs) == 1:
sizes = op.attribs['split']
elif len(op.inputs) == 2:
if op.inputs[1].data is None:
raise utils.NNEFToolsException(
'Split is not supported with calculated sizes.')
sizes = op.inputs[1].data.tolist()
op.attribs['split'] = sizes
else:
assert False
op.inputs = (op.inputs[0], )
output_shapes = tuple(
infer.split(input=op.inputs[0].shape,
axis=op.attribs['axis'],
sizes=sizes))
return output_shapes, (op.inputs[0].dtype, ) * len(output_shapes)
def slice_shape(op):
return shapes.split(op.inputs[0].shape,
axis=op.attribs['axis'],
split_points=op.attribs['slice_point'])
