xref: /aosp_15_r20/external/pytorch/torch/onnx/symbolic_opset7.py (revision da0073e96a02ea20f0ac840b70461e3646d07c45) 
1# mypy: allow-untyped-defs
2"""
3Note [ONNX operators that are added/updated from opset 7 to opset 8]
4~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
5New operators:
6  Expand
7
8Updated operators:
9  Min, Max, Sum, Mean: supports multidirectional broadcasting.
10  MaxPool: added optional indices output.
11  Scan
12"""
13
14import functools
15import warnings
16
17from torch.onnx import symbolic_helper, symbolic_opset9 as opset9
18from torch.onnx._internal import jit_utils, registration
19
20
21_onnx_symbolic = functools.partial(registration.onnx_symbolic, opset=7)
22
23block_listed_operators = (
24    "scan",
25    "expand",
26    "expand_as",
27    "meshgrid",
28    "adaptive_max_pool1d",
29    "adaptive_max_pool2d",
30    "adaptive_max_pool3d",
31    "max_pool1d_with_indices",
32    "max_pool2d_with_indices",
33    "max_pool3d_with_indices",
34)
35
36
37# NOTE: max, min, sum, mean: broadcasting is not supported in opset 7.
38# torch.max (same for torch.min) actually has two interfaces smashed together:
39# torch.max(x, dim, keepdim) and torch.max(x, y)
40@_onnx_symbolic("aten::max")
41def max(g: jit_utils.GraphContext, self, dim_or_y=None, keepdim=None):
42    # torch.max(input, other)
43    if keepdim is None and dim_or_y is not None:
44        warnings.warn(
45            "Multidirectional broadcasting is not supported in opset 7. "
46            "This might cause the onnx model to be incorrect, if inputs to max operators "
47            "have different shapes"
48        )
49    return opset9.max(g, self, dim_or_y, keepdim)
50
51
52@_onnx_symbolic("aten::min")
53def min(g: jit_utils.GraphContext, self, dim_or_y=None, keepdim=None):
54    # torch.min(input, other)
55    if keepdim is None and dim_or_y is not None:
56        warnings.warn(
57            "Multidirectional broadcasting is not supported in opset 7. "
58            "This might cause the onnx model to be incorrect, if inputs to min operators "
59            "have different shapes"
60        )
61    return opset9.min(g, self, dim_or_y, keepdim)
62
63
64for block_listed_op in block_listed_operators:
65    _onnx_symbolic(f"aten::{block_listed_op}")(
66        symbolic_helper._block_list_in_opset(block_listed_op)
67    )
68