datasets/cityscapes.py:143 self.centroids = self.fine_centroids might be changed to self.centroids = copy.deepcopy(self.fine_centroids). As I see, self.centroids is a combination of self.fine_centroids and 'coarse_centroids', but if you just use = here, self.fine_centroids would be changed as self.centroids changing in the fowllowing codes: self.centroids[cid].extend(self.coarse_centroids[cid]), which locates at 163 in the same file.
So when disable_coarse() is called, self.centroids is not really changed, because self.fine_centroids is the same as self.centroids.
train.py:442: train_obj.disable_coarse(). In my test, even though we call this function, and it did execute, however, the dataset in DataLoader is not changed. If you print the id(self.imgs) inside function build_epoch() and __getitem__() separately after call disable_coarse(), you will find their id are different, which means self.imgs is changed in build_epoch(), but not used in __getitem__(). Maybe this is because of the implementation of DataLoader.
When epoch > max_cu_epoch, new generated self.imgs should be all come from 'gtFine_trainvaltest', but as these two problems, it won't work.
self.centroids = self.fine_centroidsmight be changed toself.centroids = copy.deepcopy(self.fine_centroids). As I see,self.centroidsis a combination ofself.fine_centroidsand 'coarse_centroids', but if you just use=here,self.fine_centroidswould be changed asself.centroidschanging in the fowllowing codes:self.centroids[cid].extend(self.coarse_centroids[cid]), which locates at 163 in the same file. So whendisable_coarse()is called,self.centroidsis not really changed, becauseself.fine_centroidsis the same asself.centroids.train_obj.disable_coarse(). In my test, even though we call this function, and it did execute, however, thedatasetin DataLoader is not changed. If you print theid(self.imgs)inside functionbuild_epoch()and__getitem__()separately after calldisable_coarse(), you will find their id are different, which meansself.imgsis changed inbuild_epoch(), but not used in__getitem__(). Maybe this is because of the implementation of DataLoader.epoch > max_cu_epoch, new generatedself.imgsshould be all come from 'gtFine_trainvaltest', but as these two problems, it won't work.