mobilenetv2.py

from functools import partial
from typing import Any, Optional, Union

from torch import nn, Tensor
from torch.ao.quantization import DeQuantStub, QuantStub
from torchvision.models.mobilenetv2 import InvertedResidual, MobileNet_V2_Weights, MobileNetV2

from ...ops.misc import Conv2dNormActivation
from ...transforms._presets import ImageClassification
from .._api import register_model, Weights, WeightsEnum
from .._meta import _IMAGENET_CATEGORIES
from .._utils import _ovewrite_named_param, handle_legacy_interface
from .utils import _fuse_modules, _replace_relu, quantize_model


__all__ = [
    "QuantizableMobileNetV2",
    "MobileNet_V2_QuantizedWeights",
    "mobilenet_v2",
]


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

    def forward(self, x: Tensor) -> Tensor:
        if self.use_res_connect:
            return self.skip_add.add(x, self.conv(x))
        else:
            return self.conv(x)

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


class QuantizableMobileNetV2(MobileNetV2):
    def __init__(self, *args: Any, **kwargs: Any) -> None:
        """
        MobileNet V2 main class

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

    def forward(self, x: Tensor) -> Tensor:
        x = self.quant(x)
        x = self._forward_impl(x)
        x = self.dequant(x)
        return x

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


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,
            "_metrics": {
                "ImageNet-1K": {
                    "acc@1": 71.658,
                    "acc@5": 90.150,
                }
            },
            "_ops": 0.301,
            "_file_size": 3.423,
            "_docs": """
                These weights were produced by doing Quantization Aware Training (eager mode) on top of the unquantized
                weights listed below.
            """,
        },
    )
    DEFAULT = IMAGENET1K_QNNPACK_V1


@register_model(name="quantized_mobilenet_v2")
@handle_legacy_interface(
    weights=(
        "pretrained",
        lambda kwargs: MobileNet_V2_QuantizedWeights.IMAGENET1K_QNNPACK_V1
        if kwargs.get("quantize", False)
        else MobileNet_V2_Weights.IMAGENET1K_V1,
    )
)
def mobilenet_v2(
    *,
    weights: Optional[Union[MobileNet_V2_QuantizedWeights, MobileNet_V2_Weights]] = None,
    progress: bool = True,
    quantize: bool = False,
    **kwargs: Any,
) -> QuantizableMobileNetV2:
    """
    Constructs a MobileNetV2 architecture from
    `MobileNetV2: Inverted Residuals and Linear Bottlenecks
    <https://arxiv.org/abs/1801.04381>`_.

    .. 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.

    Args:
        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:
    """
    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")

    model = QuantizableMobileNetV2(block=QuantizableInvertedResidual, **kwargs)
    _replace_relu(model)
    if quantize:
        quantize_model(model, backend)

    if weights is not None:
        model.load_state_dict(weights.get_state_dict(progress=progress, check_hash=True))

    return model