"magic_pdf/vscode:/vscode.git/clone" did not exist on "b6710b99886724dada3ce81eb2f48793c01a05a9"
mobilenetv2.py 5.67 KB
Newer Older
1
2
from functools import partial
from typing import Any, Optional, Union
3

4
5
from torch import Tensor
from torch import nn
6
from torch.ao.quantization import QuantStub, DeQuantStub
7
from torchvision.models.mobilenetv2 import InvertedResidual, MobileNetV2, MobileNet_V2_Weights
8

9
from ...ops.misc import Conv2dNormActivation
10
from ...transforms._presets import ImageClassification
11
12
13
from .._api import WeightsEnum, Weights
from .._meta import _IMAGENET_CATEGORIES
from .._utils import handle_legacy_interface, _ovewrite_named_param
14
from .utils import _fuse_modules, _replace_relu, quantize_model
15
16


17
18
19
20
21
__all__ = [
    "QuantizableMobileNetV2",
    "MobileNet_V2_QuantizedWeights",
    "mobilenet_v2",
]
22
23
24


class QuantizableInvertedResidual(InvertedResidual):
25
    def __init__(self, *args: Any, **kwargs: Any) -> None:
26
        super().__init__(*args, **kwargs)
27
28
        self.skip_add = nn.quantized.FloatFunctional()

29
    def forward(self, x: Tensor) -> Tensor:
30
31
32
33
34
        if self.use_res_connect:
            return self.skip_add.add(x, self.conv(x))
        else:
            return self.conv(x)

35
    def fuse_model(self, is_qat: Optional[bool] = None) -> None:
36
        for idx in range(len(self.conv)):
37
            if type(self.conv[idx]) is nn.Conv2d:
38
                _fuse_modules(self.conv, [str(idx), str(idx + 1)], is_qat, inplace=True)
39
40
41


class QuantizableMobileNetV2(MobileNetV2):
42
    def __init__(self, *args: Any, **kwargs: Any) -> None:
43
44
45
46
47
48
        """
        MobileNet V2 main class

        Args:
           Inherits args from floating point MobileNetV2
        """
49
        super().__init__(*args, **kwargs)
50
51
52
        self.quant = QuantStub()
        self.dequant = DeQuantStub()

53
    def forward(self, x: Tensor) -> Tensor:
54
55
56
57
58
        x = self.quant(x)
        x = self._forward_impl(x)
        x = self.dequant(x)
        return x

59
    def fuse_model(self, is_qat: Optional[bool] = None) -> None:
60
        for m in self.modules():
61
            if type(m) is Conv2dNormActivation:
62
                _fuse_modules(m, ["0", "1", "2"], is_qat, inplace=True)
63
            if type(m) is QuantizableInvertedResidual:
64
                m.fuse_model(is_qat)
65
66


67
68
69
70
71
72
73
74
75
76
77
class MobileNet_V2_QuantizedWeights(WeightsEnum):
    IMAGENET1K_QNNPACK_V1 = Weights(
        url="https://download.pytorch.org/models/quantized/mobilenet_v2_qnnpack_37f702c5.pth",
        transforms=partial(ImageClassification, crop_size=224),
        meta={
            "num_params": 3504872,
            "min_size": (1, 1),
            "categories": _IMAGENET_CATEGORIES,
            "backend": "qnnpack",
            "recipe": "https://github.com/pytorch/vision/tree/main/references/classification#qat-mobilenetv2",
            "unquantized": MobileNet_V2_Weights.IMAGENET1K_V1,
78
79
80
81
            "metrics": {
                "acc@1": 71.658,
                "acc@5": 90.150,
            },
82
83
84
85
86
87
88
89
90
91
92
93
94
        },
    )
    DEFAULT = IMAGENET1K_QNNPACK_V1


@handle_legacy_interface(
    weights=(
        "pretrained",
        lambda kwargs: MobileNet_V2_QuantizedWeights.IMAGENET1K_QNNPACK_V1
        if kwargs.get("quantize", False)
        else MobileNet_V2_Weights.IMAGENET1K_V1,
    )
)
95
def mobilenet_v2(
96
97
    *,
    weights: Optional[Union[MobileNet_V2_QuantizedWeights, MobileNet_V2_Weights]] = None,
98
99
100
101
    progress: bool = True,
    quantize: bool = False,
    **kwargs: Any,
) -> QuantizableMobileNetV2:
102
103
    """
    Constructs a MobileNetV2 architecture from
104
    `MobileNetV2: Inverted Residuals and Linear Bottlenecks
105
106
    <https://arxiv.org/abs/1801.04381>`_.

107
108
109
110
    .. note::
        Note that ``quantize = True`` returns a quantized model with 8 bit
        weights. Quantized models only support inference and run on CPUs.
        GPU inference is not yet supported.
111
112

    Args:
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
        weights (:class:`~torchvision.models.quantization.MobileNet_V2_QuantizedWeights` or :class:`~torchvision.models.MobileNet_V2_Weights`, optional): The
            pretrained weights for the model. See
            :class:`~torchvision.models.quantization.MobileNet_V2_QuantizedWeights` below for
            more details, and possible values. By default, no pre-trained
            weights are used.
        progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True.
        quantize (bool, optional): If True, returns a quantized version of the model. Default is False.
        **kwargs: parameters passed to the ``torchvision.models.quantization.QuantizableMobileNetV2``
            base class. Please refer to the `source code
            <https://github.com/pytorch/vision/blob/main/torchvision/models/quantization/mobilenetv2.py>`_
            for more details about this class.
    .. autoclass:: torchvision.models.quantization.MobileNet_V2_QuantizedWeights
        :members:
    .. autoclass:: torchvision.models.MobileNet_V2_Weights
        :members:
        :noindex:
129
    """
130
131
132
133
134
135
136
137
    weights = (MobileNet_V2_QuantizedWeights if quantize else MobileNet_V2_Weights).verify(weights)

    if weights is not None:
        _ovewrite_named_param(kwargs, "num_classes", len(weights.meta["categories"]))
        if "backend" in weights.meta:
            _ovewrite_named_param(kwargs, "backend", weights.meta["backend"])
    backend = kwargs.pop("backend", "qnnpack")

138
139
140
141
142
    model = QuantizableMobileNetV2(block=QuantizableInvertedResidual, **kwargs)
    _replace_relu(model)
    if quantize:
        quantize_model(model, backend)

143
144
    if weights is not None:
        model.load_state_dict(weights.get_state_dict(progress=progress))
145
146

    return model
147
148
149
150
151
152
153
154
155
156
157
158


# The dictionary below is internal implementation detail and will be removed in v0.15
from .._utils import _ModelURLs
from ..mobilenetv2 import model_urls  # noqa: F401


quant_model_urls = _ModelURLs(
    {
        "mobilenet_v2_qnnpack": MobileNet_V2_QuantizedWeights.IMAGENET1K_QNNPACK_V1.url,
    }
)