codekingpro/portable-devtools

1# Copyright (c) Microsoft Corporation. All rights reserved.
2# Licensed under the MIT License.
3
4from __future__ import annotations
5
6import inspect
7from collections import abc
8
9import torch
10
11
12def _parse_inputs_for_onnx_export(all_input_parameters, inputs, kwargs):
13    # extracted from https://github.com/microsoft/onnxruntime/blob/239c6ad3f021ff7cc2e6247eb074bd4208dc11e2/orttraining/orttraining/python/training/ortmodule/_io.py#L433
14
15    def _add_input(name, input):
16        """Returns number of expanded inputs that _add_input processed"""
17
18        if input is None:
19            # Drop all None inputs and return 0.
20            return 0
21
22        num_expanded_non_none_inputs = 0
23        if isinstance(input, abc.Sequence):
24            # If the input is a sequence (like a list), expand the list so that
25            # each element of the list is an input by itself.
26            for i, val in enumerate(input):
27                # Name each input with the index appended to the original name of the
28                # argument.
29                num_expanded_non_none_inputs += _add_input(f"{name}_{i}", val)
30
31            # Return here since the list by itself is not a valid input.
32            # All the elements of the list have already been added as inputs individually.
33            return num_expanded_non_none_inputs
34        elif isinstance(input, abc.Mapping):
35            # If the input is a mapping (like a dict), expand the dict so that
36            # each element of the dict is an input by itself.
37            for key, val in input.items():
38                num_expanded_non_none_inputs += _add_input(f"{name}_{key}", val)
39
40            # Return here since the dict by itself is not a valid input.
41            # All the elements of the dict have already been added as inputs individually.
42            return num_expanded_non_none_inputs
43
44        # InputInfo should contain all the names irrespective of whether they are
45        # a part of the onnx graph or not.
46        input_names.append(name)
47
48        # A single input non none input was processed, return 1
49        return 1
50
51    input_names = []
52    var_positional_idx = 0
53    num_expanded_non_none_positional_inputs = 0
54
55    for input_idx, input_parameter in enumerate(all_input_parameters):
56        if input_parameter.kind == inspect.Parameter.VAR_POSITIONAL:
57            # VAR_POSITIONAL parameter carries all *args parameters from original forward method
58            for args_i in range(input_idx, len(inputs)):
59                name = f"{input_parameter.name}_{var_positional_idx}"
60                var_positional_idx += 1
61                inp = inputs[args_i]
62                num_expanded_non_none_positional_inputs += _add_input(name, inp)
63        elif (
64            input_parameter.kind == inspect.Parameter.POSITIONAL_ONLY
65            or input_parameter.kind == inspect.Parameter.POSITIONAL_OR_KEYWORD
66            or input_parameter.kind == inspect.Parameter.KEYWORD_ONLY
67        ):
68            # All positional non-*args and non-**kwargs are processed here
69            name = input_parameter.name
70            inp = None
71            input_idx += var_positional_idx  # noqa: PLW2901
72            is_positional = True
73            if input_idx < len(inputs) and inputs[input_idx] is not None:
74                inp = inputs[input_idx]
75            elif name in kwargs and kwargs[name] is not None:
76                inp = kwargs[name]
77                is_positional = False
78            num_expanded_non_none_inputs_local = _add_input(name, inp)
79            if is_positional:
80                num_expanded_non_none_positional_inputs += num_expanded_non_none_inputs_local
81        elif input_parameter.kind == inspect.Parameter.VAR_KEYWORD:
82            # **kwargs is always the last argument of forward()
83            for name, inp in kwargs.items():
84                if name not in input_names:
85                    _add_input(name, inp)
86
87    return input_names
88
89
90def _flatten_module_input(names, args, kwargs):
91    """Flatten args and kwargs in a single tuple of tensors."""
92    # extracted from https://github.com/microsoft/onnxruntime/blob/239c6ad3f021ff7cc2e6247eb074bd4208dc11e2/orttraining/orttraining/python/training/ortmodule/_io.py#L110
93
94    def is_primitive_type(value):
95        return type(value) in {int, bool, float}
96
97    def to_tensor(value):
98        return torch.tensor(value)
99
100    ret = [to_tensor(arg) if is_primitive_type(arg) else arg for arg in args]
101    ret += [
102        to_tensor(kwargs[name]) if is_primitive_type(kwargs[name]) else kwargs[name] for name in names if name in kwargs
103    ]
104
105    # if kwargs is empty, append an empty dictionary at the end of the sample inputs to make exporter
106    # happy. This is because the exporter is confused with kwargs and dictionary inputs otherwise.
107    if not kwargs:
108        ret.append({})
109
110    return tuple(ret)
111
112
113def infer_input_info(module: torch.nn.Module, *inputs, **kwargs):
114    """
115    Infer the input names and order from the arguments used to execute a PyTorch module for usage exporting
116    the model via torch.onnx.export.
117    Assumes model is on CPU. Use `module.to(torch.device('cpu'))` if it isn't.
118
119    Example usage:
120    input_names, inputs_as_tuple = infer_input_info(module, ...)
121    torch.onnx.export(module, inputs_as_type, 'model.onnx', input_names=input_names, output_names=[...], ...)
122
123    :param module: Module
124    :param inputs: Positional inputs
125    :param kwargs: Keyword argument inputs
126    :return: Tuple of ordered input names and input values. These can be used directly with torch.onnx.export as the
127            `input_names` and `inputs` arguments.
128    """
129    module_parameters = inspect.signature(module.forward).parameters.values()
130    input_names = _parse_inputs_for_onnx_export(module_parameters, inputs, kwargs)
131    inputs_as_tuple = _flatten_module_input(input_names, inputs, kwargs)
132
133    return input_names, inputs_as_tuple
134