Closed Chen-Song closed 3 years ago
The prototype turns out to be not helping or damaging the results.
The momentum encoding seems to be a bit lower than reported in the paper, tested on VOC Split 1 and get nAP=61+. But the code might not be necessary with the one uploaded here (shouldn't have uploaded those ..).
In your code:
_C.MODEL.ROI_BOX_HEAD.CONTRASTIVE_BRANCH.PROTOTYPE.PATH = '' self.prototype_path = cfg.MODEL.ROI_BOX_HEAD.CONTRASTIVE_BRANCH.PROTOTYPE.PATH prototype_tensor = torch.load(self.prototype_path) # [num_classes+1, 1024]
Is PROTOTYPE.PATH empty?
In your code: _C.MODEL.ROI_BOX_HEAD.CONTRASTIVE_BRANCH.PROTOTYPE.PATH = '' self.prototype_path = cfg.MODEL.ROI_BOX_HEAD.CONTRASTIVE_BRANCH.PROTOTYPE.PATH prototype_tensor = torch.load(self.prototype_path) # [num_classes+1, 1024]
Is PROTOTYPE.PATH empty?
In your code, there are some variants based on MoCo ("eg, ContrastiveROIHeadsWithStorage") or Prototype ("eg, ContrastiveROIHeadsWithPrototype"). What is the final result (nAP) of these two variants? Will it be much lower than the optimal result reported in the paper?