pytorch

1
"""This module contains utility method for mobile model optimization and lint."""
2

3
import torch
4
from enum import Enum
5
from torch._C import _MobileOptimizerType as MobileOptimizerType
6
from typing import Optional, Set, List, AnyStr
7

8
class LintCode(Enum):
9
    BUNDLED_INPUT = 1
10
    REQUIRES_GRAD = 2
11
    DROPOUT = 3
12
    BATCHNORM = 4
13

14
def optimize_for_mobile(
15
        script_module: torch.jit.ScriptModule,
16
        optimization_blocklist: Optional[Set[MobileOptimizerType]] = None,
17
        preserved_methods: Optional[List[AnyStr]] = None,
18
        backend: str = 'CPU') -> torch.jit.RecursiveScriptModule:
19
    """
20
    Optimize a torch script module for mobile deployment.
21

22
    Args:
23
        script_module: An instance of torch script module with type of ScriptModule.
24
        optimization_blocklist: A set with type of MobileOptimizerType. When set is not passed,
25
            optimization method will run all the optimizer pass; otherwise, optimizer
26
            method will run the optimization pass that is not included inside optimization_blocklist.
27
        preserved_methods: A list of methods that needed to be preserved when freeze_module pass is invoked
28
        backend: Device type to use for running the result model ('CPU'(default), 'Vulkan' or 'Metal').
29
    Returns:
30
        A new optimized torch script module
31
    """
32
    if not isinstance(script_module, torch.jit.ScriptModule):
33
        raise TypeError(
34
            f'Got {type(script_module)}, but ScriptModule is expected.')
35

36
    if optimization_blocklist is None:
37
        optimization_blocklist = set()
38

39
    if preserved_methods is None:
40
        preserved_methods = []
41

42
    # Convert potential byte arrays into strings (if there is any) to pass type checking
43
    # Here we use a new name as assigning it back to preserved_methods will invoke
44
    # mypy errors (i.e. List[AnyStr] = List[str])
45
    preserved_methods_str: List[str] = [str(method) for method in preserved_methods]
46

47
    bundled_inputs_attributes = _get_bundled_inputs_preserved_attributes(script_module, preserved_methods_str)
48
    if all(hasattr(script_module, method) for method in bundled_inputs_attributes):
49
        preserved_methods_str = list(set(preserved_methods_str + bundled_inputs_attributes))
50

51
    non_exist_methods = []
52
    for method in preserved_methods_str:
53
        if not hasattr(script_module, method):
54
            non_exist_methods.append(method)
55
    if non_exist_methods:
56
        raise AttributeError(
57
            f"The following methods to preserve do not exist in script_module: {', '.join(non_exist_methods)}")
58

59
    backend = backend.lower()
60
    if backend == 'cpu':
61
        optimized_cpp_module = torch._C._jit_pass_optimize_for_mobile(
62
            script_module._c,
63
            optimization_blocklist,
64
            preserved_methods_str)
65
    elif backend == 'vulkan':
66
        optimized_cpp_module = torch._C._jit_pass_vulkan_optimize_for_mobile(
67
            script_module._c,
68
            optimization_blocklist,
69
            preserved_methods_str)
70
    elif backend == 'metal':
71
        optimized_cpp_module = torch._C._jit_pass_metal_optimize_for_mobile(script_module._c, preserved_methods_str)
72
    else:
73
        raise TypeError("Unknown backend, must be one of 'CPU', 'Vulkan' or 'Metal'")
74

75
    return torch.jit._recursive.wrap_cpp_module(optimized_cpp_module)
76

77

78
def generate_mobile_module_lints(script_module: torch.jit.ScriptModule):
79
    """
80
    Generate a list of lints for a given torch script module.
81

82
    Args:
83
        script_module: An instance of torch script module with type of ScriptModule.
84

85
    Returns:
86
        lint_map: A list of dictionary that contains modules lints
87
    """
88
    if not isinstance(script_module, torch.jit.ScriptModule):
89
        raise TypeError(
90
            f'Got {type(script_module)}, but ScriptModule is expected.')
91

92
    lint_list = []
93

94
    if not hasattr(script_module, "_generate_bundled_inputs_for_forward"):
95
        lint_list.append({"name": LintCode.BUNDLED_INPUT.name, "message": "No bundled input for forward, please add bundled inputs "
96
                          "before saving the module using torch.utils.bundled_inputs.augment_model_with_bundled_inputs."})
97

98
    for name, param in script_module.named_parameters():
99
        if param.requires_grad:
100
            lint_list.append({"name": LintCode.REQUIRES_GRAD.name, "message": f"Param {name} requires grad, "
101
                             "please set torch.no_grad() to reduce memory usage and improve computation speed during "
102
                              "inference phase."})
103

104
    op_names = torch.jit.export_opnames(script_module)
105
    for op_name in op_names:
106
        if "dropout" in op_name:
107
            lint_list.append({"name": LintCode.DROPOUT.name, "message": "Operator {} exists, remember to call eval() before "
108
                              "saving the module.and call torch.utils.mobile_optimizer.optimize_for_mobile to drop dropout "
109
                              "operator.".format(op_name)})
110
        if "batch_norm" in op_name:
111
            lint_list.append({"name": LintCode.BATCHNORM.name, "message": "Operator {} exists, remember to call eval() before "
112
                              "saving the module and call torch.utils.mobile_optimizer.optimize_for_mobile to drop batch_norm "
113
                              "operator.".format(op_name)})
114

115
    return lint_list
116

117
def _get_bundled_inputs_preserved_attributes(script_module: torch.jit.ScriptModule, preserved_methods: List[str]) -> List[str]:
118

119
    bundled_inputs_attributes = []
120
    # Has bundled inputs for forward
121
    if hasattr(script_module, 'get_all_bundled_inputs'):
122
        bundled_inputs_attributes.append('get_all_bundled_inputs')
123
        bundled_inputs_attributes.append('get_num_bundled_inputs')
124

125
    # Bundled inputs in module after the change that introduced bundled inputs for multiple functions
126
    if hasattr(script_module, 'get_bundled_inputs_functions_and_info'):
127
        bundled_inputs_attributes.append('get_bundled_inputs_functions_and_info')
128
        all_info = script_module.get_bundled_inputs_functions_and_info()
129
        for function_name in all_info:
130
            if function_name not in preserved_methods:
131
                bundled_inputs_attributes.append(function_name)
132
            bundled_inputs_attributes.append("get_all_bundled_inputs_for_" + function_name)
133
            bundled_inputs_attributes.append("_bundled_inputs_deflated_" + function_name)
134

135
    return bundled_inputs_attributes
136