ssdlite.py

import warnings
from collections import OrderedDict
from functools import partial
from typing import Any, Callable, Dict, List, Optional, Union

import torch
from torch import nn, Tensor

from ...ops.misc import Conv2dNormActivation
from ...transforms._presets import ObjectDetection
from ...utils import _log_api_usage_once
from .. import mobilenet
from .._api import register_model, Weights, WeightsEnum
from .._meta import _COCO_CATEGORIES
from .._utils import _ovewrite_value_param, handle_legacy_interface
from ..mobilenetv3 import mobilenet_v3_large, MobileNet_V3_Large_Weights
from . import _utils as det_utils
from .anchor_utils import DefaultBoxGenerator
from .backbone_utils import _validate_trainable_layers
from .ssd import SSD, SSDScoringHead


__all__ = [
    "SSDLite320_MobileNet_V3_Large_Weights",
    "ssdlite320_mobilenet_v3_large",
]


# Building blocks of SSDlite as described in section 6.2 of MobileNetV2 paper
def _prediction_block(
    in_channels: int, out_channels: int, kernel_size: int, norm_layer: Callable[..., nn.Module]
) -> nn.Sequential:
    return nn.Sequential(
        # 3x3 depthwise with stride 1 and padding 1
        Conv2dNormActivation(
            in_channels,
            in_channels,
            kernel_size=kernel_size,
            groups=in_channels,
            norm_layer=norm_layer,
            activation_layer=nn.ReLU6,
        ),
        # 1x1 projetion to output channels
        nn.Conv2d(in_channels, out_channels, 1),
    )


def _extra_block(in_channels: int, out_channels: int, norm_layer: Callable[..., nn.Module]) -> nn.Sequential:
    activation = nn.ReLU6
    intermediate_channels = out_channels // 2
    return nn.Sequential(
        # 1x1 projection to half output channels
        Conv2dNormActivation(
            in_channels, intermediate_channels, kernel_size=1, norm_layer=norm_layer, activation_layer=activation
        ),
        # 3x3 depthwise with stride 2 and padding 1
        Conv2dNormActivation(
            intermediate_channels,
            intermediate_channels,
            kernel_size=3,
            stride=2,
            groups=intermediate_channels,
            norm_layer=norm_layer,
            activation_layer=activation,
        ),
        # 1x1 projetion to output channels
        Conv2dNormActivation(
            intermediate_channels, out_channels, kernel_size=1, norm_layer=norm_layer, activation_layer=activation
        ),
    )


def _normal_init(conv: nn.Module):
    for layer in conv.modules():
        if isinstance(layer, nn.Conv2d):
            torch.nn.init.normal_(layer.weight, mean=0.0, std=0.03)
            if layer.bias is not None:
                torch.nn.init.constant_(layer.bias, 0.0)


class SSDLiteHead(nn.Module):
    def __init__(
        self, in_channels: List[int], num_anchors: List[int], num_classes: int, norm_layer: Callable[..., nn.Module]
    ):
        super().__init__()
        self.classification_head = SSDLiteClassificationHead(in_channels, num_anchors, num_classes, norm_layer)
        self.regression_head = SSDLiteRegressionHead(in_channels, num_anchors, norm_layer)

    def forward(self, x: List[Tensor]) -> Dict[str, Tensor]:
        return {
            "bbox_regression": self.regression_head(x),
            "cls_logits": self.classification_head(x),
        }


class SSDLiteClassificationHead(SSDScoringHead):
    def __init__(
        self, in_channels: List[int], num_anchors: List[int], num_classes: int, norm_layer: Callable[..., nn.Module]
    ):
        cls_logits = nn.ModuleList()
        for channels, anchors in zip(in_channels, num_anchors):
            cls_logits.append(_prediction_block(channels, num_classes * anchors, 3, norm_layer))
        _normal_init(cls_logits)
        super().__init__(cls_logits, num_classes)


class SSDLiteRegressionHead(SSDScoringHead):
    def __init__(self, in_channels: List[int], num_anchors: List[int], norm_layer: Callable[..., nn.Module]):
        bbox_reg = nn.ModuleList()
        for channels, anchors in zip(in_channels, num_anchors):
            bbox_reg.append(_prediction_block(channels, 4 * anchors, 3, norm_layer))
        _normal_init(bbox_reg)
        super().__init__(bbox_reg, 4)


class SSDLiteFeatureExtractorMobileNet(nn.Module):
    def __init__(
        self,
        backbone: nn.Module,
        c4_pos: int,
        norm_layer: Callable[..., nn.Module],
        width_mult: float = 1.0,
        min_depth: int = 16,
    ):
        super().__init__()
        _log_api_usage_once(self)

        if backbone[c4_pos].use_res_connect:
            raise ValueError("backbone[c4_pos].use_res_connect should be False")

        self.features = nn.Sequential(
            # As described in section 6.3 of MobileNetV3 paper
            nn.Sequential(*backbone[:c4_pos], backbone[c4_pos].block[0]),  # from start until C4 expansion layer
            nn.Sequential(backbone[c4_pos].block[1:], *backbone[c4_pos + 1 :]),  # from C4 depthwise until end
        )

        get_depth = lambda d: max(min_depth, int(d * width_mult))  # noqa: E731
        extra = nn.ModuleList(
            [
                _extra_block(backbone[-1].out_channels, get_depth(512), norm_layer),
                _extra_block(get_depth(512), get_depth(256), norm_layer),
                _extra_block(get_depth(256), get_depth(256), norm_layer),
                _extra_block(get_depth(256), get_depth(128), norm_layer),
            ]
        )
        _normal_init(extra)

        self.extra = extra

    def forward(self, x: Tensor) -> Dict[str, Tensor]:
        # Get feature maps from backbone and extra. Can't be refactored due to JIT limitations.
        output = []
        for block in self.features:
            x = block(x)
            output.append(x)

        for block in self.extra:
            x = block(x)
            output.append(x)

        return OrderedDict([(str(i), v) for i, v in enumerate(output)])


def _mobilenet_extractor(
    backbone: Union[mobilenet.MobileNetV2, mobilenet.MobileNetV3],
    trainable_layers: int,
    norm_layer: Callable[..., nn.Module],
):
    backbone = backbone.features
    # Gather the indices of blocks which are strided. These are the locations of C1, ..., Cn-1 blocks.
    # The first and last blocks are always included because they are the C0 (conv1) and Cn.
    stage_indices = [0] + [i for i, b in enumerate(backbone) if getattr(b, "_is_cn", False)] + [len(backbone) - 1]
    num_stages = len(stage_indices)

    # find the index of the layer from which we won't freeze
    if not 0 <= trainable_layers <= num_stages:
        raise ValueError("trainable_layers should be in the range [0, {num_stages}], instead got {trainable_layers}")
    freeze_before = len(backbone) if trainable_layers == 0 else stage_indices[num_stages - trainable_layers]

    for b in backbone[:freeze_before]:
        for parameter in b.parameters():
            parameter.requires_grad_(False)

    return SSDLiteFeatureExtractorMobileNet(backbone, stage_indices[-2], norm_layer)


class SSDLite320_MobileNet_V3_Large_Weights(WeightsEnum):
    COCO_V1 = Weights(
        url="https://download.pytorch.org/models/ssdlite320_mobilenet_v3_large_coco-a79551df.pth",
        transforms=ObjectDetection,
        meta={
            "num_params": 3440060,
            "categories": _COCO_CATEGORIES,
            "min_size": (1, 1),
            "recipe": "https://github.com/pytorch/vision/tree/main/references/detection#ssdlite320-mobilenetv3-large",
            "_metrics": {
                "COCO-val2017": {
                    "box_map": 21.3,
                }
            },
            "_ops": 0.583,
            "_file_size": 13.418,
            "_docs": """These weights were produced by following a similar training recipe as on the paper.""",
        },
    )
    DEFAULT = COCO_V1


@register_model()
@handle_legacy_interface(
    weights=("pretrained", SSDLite320_MobileNet_V3_Large_Weights.COCO_V1),
    weights_backbone=("pretrained_backbone", MobileNet_V3_Large_Weights.IMAGENET1K_V1),
)
def ssdlite320_mobilenet_v3_large(
    *,
    weights: Optional[SSDLite320_MobileNet_V3_Large_Weights] = None,
    progress: bool = True,
    num_classes: Optional[int] = None,
    weights_backbone: Optional[MobileNet_V3_Large_Weights] = MobileNet_V3_Large_Weights.IMAGENET1K_V1,
    trainable_backbone_layers: Optional[int] = None,
    norm_layer: Optional[Callable[..., nn.Module]] = None,
    **kwargs: Any,
) -> SSD:
    """SSDlite model architecture with input size 320x320 and a MobileNetV3 Large backbone, as
    described at `Searching for MobileNetV3 <https://arxiv.org/abs/1905.02244>`__ and
    `MobileNetV2: Inverted Residuals and Linear Bottlenecks <https://arxiv.org/abs/1801.04381>`__.

    .. betastatus:: detection module

    See :func:`~torchvision.models.detection.ssd300_vgg16` for more details.

    Example:

        >>> model = torchvision.models.detection.ssdlite320_mobilenet_v3_large(weights=SSDLite320_MobileNet_V3_Large_Weights.DEFAULT)
        >>> model.eval()
        >>> x = [torch.rand(3, 320, 320), torch.rand(3, 500, 400)]
        >>> predictions = model(x)

    Args:
        weights (:class:`~torchvision.models.detection.SSDLite320_MobileNet_V3_Large_Weights`, optional): The
            pretrained weights to use. See
            :class:`~torchvision.models.detection.SSDLite320_MobileNet_V3_Large_Weights` 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.
        num_classes (int, optional): number of output classes of the model
            (including the background).
        weights_backbone (:class:`~torchvision.models.MobileNet_V3_Large_Weights`, optional): The pretrained
            weights for the backbone.
        trainable_backbone_layers (int, optional): number of trainable (not frozen) layers
            starting from final block. Valid values are between 0 and 6, with 6 meaning all
            backbone layers are trainable. If ``None`` is passed (the default) this value is
            set to 6.
        norm_layer (callable, optional): Module specifying the normalization layer to use.
        **kwargs: parameters passed to the ``torchvision.models.detection.ssd.SSD``
            base class. Please refer to the `source code
            <https://github.com/pytorch/vision/blob/main/torchvision/models/detection/ssdlite.py>`_
            for more details about this class.

    .. autoclass:: torchvision.models.detection.SSDLite320_MobileNet_V3_Large_Weights
        :members:
    """

    weights = SSDLite320_MobileNet_V3_Large_Weights.verify(weights)
    weights_backbone = MobileNet_V3_Large_Weights.verify(weights_backbone)

    if "size" in kwargs:
        warnings.warn("The size of the model is already fixed; ignoring the parameter.")

    if weights is not None:
        weights_backbone = None
        num_classes = _ovewrite_value_param("num_classes", num_classes, len(weights.meta["categories"]))
    elif num_classes is None:
        num_classes = 91

    trainable_backbone_layers = _validate_trainable_layers(
        weights is not None or weights_backbone is not None, trainable_backbone_layers, 6, 6
    )

    # Enable reduced tail if no pretrained backbone is selected. See Table 6 of MobileNetV3 paper.
    reduce_tail = weights_backbone is None

    if norm_layer is None:
        norm_layer = partial(nn.BatchNorm2d, eps=0.001, momentum=0.03)

    backbone = mobilenet_v3_large(
        weights=weights_backbone, progress=progress, norm_layer=norm_layer, reduced_tail=reduce_tail, **kwargs
    )
    if weights_backbone is None:
        # Change the default initialization scheme if not pretrained
        _normal_init(backbone)
    backbone = _mobilenet_extractor(
        backbone,
        trainable_backbone_layers,
        norm_layer,
    )

    size = (320, 320)
    anchor_generator = DefaultBoxGenerator([[2, 3] for _ in range(6)], min_ratio=0.2, max_ratio=0.95)
    out_channels = det_utils.retrieve_out_channels(backbone, size)
    num_anchors = anchor_generator.num_anchors_per_location()
    if len(out_channels) != len(anchor_generator.aspect_ratios):
        raise ValueError(
            f"The length of the output channels from the backbone {len(out_channels)} do not match the length of the anchor generator aspect ratios {len(anchor_generator.aspect_ratios)}"
        )

    defaults = {
        "score_thresh": 0.001,
        "nms_thresh": 0.55,
        "detections_per_img": 300,
        "topk_candidates": 300,
        # Rescale the input in a way compatible to the backbone:
        # The following mean/std rescale the data from [0, 1] to [-1, 1]
        "image_mean": [0.5, 0.5, 0.5],
        "image_std": [0.5, 0.5, 0.5],
    }
    kwargs: Any = {**defaults, **kwargs}
    model = SSD(
        backbone,
        anchor_generator,
        size,
        num_classes,
        head=SSDLiteHead(out_channels, num_anchors, num_classes, norm_layer),
        **kwargs,
    )

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

    return model