How does RTMDet ensure that the outputs of reg branch are non-negative?

[tkhe]

RTMDet use ltrb as the output, which should all be non-negative. otherwise when transform ltrb to x1 y1 x2 y2, there could be a situation where x2 < x1.

mmdet use exp to ensure this, see in https://github.com/open-mmlab/mmdetection/blob/cfd5d3a985b0249de009b67d04f37263e11cdf3d/mmdet/models/dense_heads/rtmdet_head.py#L143

reg_dist = scale(self.rtm_reg(reg_feat).exp()).float() * stride[0]

but it seems to be missing in mmyolo. rtm_reg is initialized with normal distribution, and there are no layers like relu or exp following it, so how can we ensure the output are non-negative, especially in the first few iterations of training?

mmyolo/mmyolo/models/dense_heads/rtmdet_head.py

Lines 148 to 151 in 8c4d9dc

	bias_cls = bias_init_with_prob(0.01)
	for rtm_cls, rtm_reg in zip(self.rtm_cls, self.rtm_reg):
	normal_init(rtm_cls, std=0.01, bias=bias_cls)
	normal_init(rtm_reg, std=0.01)

mmyolo/mmyolo/models/dense_heads/rtmdet_head.py

Lines 172 to 185 in 8c4d9dc

	for idx, x in enumerate(feats):
	cls_feat = x
	reg_feat = x
	for cls_layer in self.cls_convs[idx]:
	cls_feat = cls_layer(cls_feat)
	cls_score = self.rtm_cls[idx](cls_feat)
	for reg_layer in self.reg_convs[idx]:
	reg_feat = reg_layer(reg_feat)
	reg_dist = self.rtm_reg[idx](reg_feat)
	cls_scores.append(cls_score)
	bbox_preds.append(reg_dist)