这里 in_channels=256(P层是256的)。步骤如下
1) x = make_roi_keypoint_feature_extractor(cfg, in_channels)(features, proposals)
2) kp_logits= make_roi_keypoint_predictor(cfg, self.feature_extractor.out_channels)(x)
self.feature_extractor.out_channels=512
3) post_processor = make_roi_keypoint_post_processor(cfg)
4) loss_evaluator = make_roi_keypoint_loss_evaluator(cfg)
5) 训练:loss_kp = self.loss_evaluator(proposals, kp_logits)
也就是loss_kp =make_roi_keypoint_loss_evaluator(cfg)(proposals, kp_logits)
return x, proposals, dict(loss_kp=loss_kp)
测试:
result = make_roi_keypoint_post_processor(cfg)(kp_logits, proposals)
6) loss_kp = make_roi_keypoint_loss_evaluator(cfg)(proposals, kp_logits)
1) https://blog.csdn.net/m0_37644085/article/details/88725903这里