YOLOX源码阅读收货

---

程序

1、 优化器分别对模型不同的结构进行优化

            pg0, pg1, pg2 = [], [], []  # optimizer parameter groups

            for k, v in self.model.named_modules():
                if hasattr(v, "bias") and isinstance(v.bias, nn.Parameter):
                    pg2.append(v.bias)  # biases
                if isinstance(v, nn.BatchNorm2d) or "bn" in k:
                    pg0.append(v.weight)  # no decay
                elif hasattr(v, "weight") and isinstance(v.weight, nn.Parameter):
                    pg1.append(v.weight)  # apply decay

            optimizer = torch.optim.SGD(
                pg0, lr=lr, momentum=self.momentum, nesterov=True
            )
            optimizer.add_param_group(
                {"params": pg1, "weight_decay": self.weight_decay}
            )  # add pg1 with weight_decay
            optimizer.add_param_group({"params": pg2})
            self.optimizer = optimizer

2、detect时对output进行NMS可以使用torchvision.ops.nms 或者 torchvision.ops.batched_nms

#detections shape: B,8400,7 其中7=（xmin,ymin,xmax,ymax,conf,class_score,class)
 nms_out_index = torchvision.ops.nms(
                detections[:, :4],
                detections[:, 4] * detections[:, 5],
                nms_thre,
            )

 nms_out_index = torchvision.ops.batched_nms(
                detections[:, :4],
                detections[:, 4] * detections[:, 5],
                detections[:, 6],
                nms_thre,
            )

3、得到异步的正则化结构及其参数

def get_async_norm_states(module):
    async_norm_states = OrderedDict()
    for name, child in module.named_modules():
        if isinstance(child, ASYNC_NORM):
            for k, v in child.state_dict().items():              #named_modules()里面还可以使用state_dict()
                async_norm_states[".".join([name, k])] = v
    return async_norm_states

4、 批量IOU的计算方式

def bboxes_iou(bboxes_a, bboxes_b, xyxy=True):
    if bboxes_a.shape[1] != 4 or bboxes_b.shape[1] != 4:
        raise IndexError

    if xyxy:
        tl = torch.max(bboxes_a[:, None, :2], bboxes_b[:, :2])
        br = torch.min(bboxes_a[:, None, 2:], bboxes_b[:, 2:])
        area_a = torch.prod(bboxes_a[:, 2:] - bboxes_a[:, :2], 1)
        area_b = torch.prod(bboxes_b[:, 2:] - bboxes_b[:, :2], 1)
    else:
        tl = torch.max(
            (bboxes_a[:, None, :2] - bboxes_a[:, None, 2:] / 2),
            (bboxes_b[:, :2] - bboxes_b[:, 2:] / 2),
        )
        br = torch.min(
            (bboxes_a[:, None, :2] + bboxes_a[:, None, 2:] / 2),
            (bboxes_b[:, :2] + bboxes_b[:, 2:] / 2),
        )

        area_a = torch.prod(bboxes_a[:, 2:], 1)
        area_b = torch.prod(bboxes_b[:, 2:], 1)
    en = (tl < br).type(tl.type()).prod(dim=2)
    area_i = torch.prod(br - tl, 2) * en  # * ((tl < br).all())
    return area_i / (area_a[:, None] + area_b - area_i) 

5、init_normalization 简单使用

        def init_yolo(M):
            for m in M.modules():
                if isinstance(m, nn.BatchNorm2d):
                    m.eps = 1e-3
                    m.momentum = 0.03

        if getattr(self, "model", None) is None:
            in_channels = [256, 512, 1024]
            backbone = YOLOPAFPN(self.depth, self.width, in_channels=in_channels, act=self.act)
            head = YOLOXHead(self.num_classes, self.width, in_channels=in_channels, act=self.act)
            self.model = YOLOX(backbone, head)

        self.model.apply(init_yolo)
        self.model.head.initialize_biases(1e-2)