def propagate_tile(op, const_value_by_tensor):
# type: (TFOperation, _ConstValueByTensorT)->typing.Tuple[typing.List[typing.List[int]], typing.List[str]]
input, multiples = op.inputs
multiples = const_value_by_tensor[multiples].tolist() # type: typing.List[int]
return [infer.tile(input=input.shape,
repeat=multiples)], [op.attribs['T']]
def tile_shape(op):
# type: (Caffe2Operation)->ShapeResult
assert len(op.inputs) in [1, 2, 3]
tiles = op.attribs.get('tiles', 1)
axis = op.attribs.get('axis', 0)
if len(op.inputs) >= 2:
if op.inputs[1].data is None:
raise utils.NNEFToolsException(
'Tile is not supported with calculated sizes.')
tiles = op.inputs[1].data.item()
if len(op.inputs) >= 3:
if op.inputs[2].data is None:
raise utils.NNEFToolsException(
'Tile is not supported with calculated sizes.')
axis = op.inputs[2].data.item()
repeats = [1] * op.inputs[0].rank
repeats[axis] = tiles
op.attribs['tiles'] = tiles
op.attribs['axis'] = axis
op.inputs = (op.inputs[0], )
return infer.tile(op.inputs[0].shape, repeats), op.inputs[0].dtype
def propagate_tile(op):
# type: (ONNXOperation)->typing.Tuple[typing.List[typing.List[int]], typing.List[str]]
if len(op.inputs) == 3:
input, tiles, axis = op.inputs
tiles = evaluate_scalar_int_tensor_simple(tiles)
axis = evaluate_scalar_int_tensor_simple(axis)
output_shape = list(input.shape)
output_shape[axis] *= tiles
return [output_shape], [op.input.dtype]
elif len(op.inputs) == 2:
input, repeats = op.inputs
return [
infer.tile(input=input.shape,
repeat=evaluate_shape_tensor_simple(repeats))
], [input.dtype]
else:
assert False, 'Tile must have 2 or 3 inputs'
def propagate_tile(op):
# type: (ONNXOperation)->typing.Tuple[typing.List[typing.List[int]], typing.List[str]]
if len(op.inputs) == 3:
input, tiles, axis = op.inputs
tiles = evaluate_scalar_int_tensor_simple(tiles)
axis = evaluate_scalar_int_tensor_simple(axis)
repeats = [1] * input.rank
repeats[axis] = tiles
op.inputs = (input, )
op.attribs['repeats'] = repeats
elif len(op.inputs) == 2:
input, repeats = op.inputs
op.inputs = (input, )
op.attribs['repeats'] = evaluate_shape_tensor_simple(repeats)
else:
assert False, 'Tile must have 2 or 3 inputs'
return [infer.tile(input=input.shape,
repeat=op.attribs['repeats'])], [input.dtype]
def test_tile(self):
self.assertEqual([4, 6, 6, 4], infer.tile(input=[1, 2, 3, 4], repeat=[4, 3, 2, 1]))
def tile_shape(op):
rank = op.inputs[0].shape
axis = op.attrib['axis']
return shapes.tile(
op.inputs[0].shape,
repeat=[op.attrib['tiles'] if i == axis else 1 for i in range(rank)])
