Yao-DD
Yao-DD
Sorry for my carelessness, the square root should be taken here. Indeed, the score of the maximun peak always is approximately equal to 1, so that whether taking the square...
You can load checkpoints in jupyter as follows: from nest import register,modules model = modules.s3n(mode ='resnet50', num_classes = 200, task_input_size = 448, base_ratio = 0.09, radius = 0.09, radius_inv =...
You can comment out lines 114-122 in 'Nest/src/nest/modules.py' and reinstall Nest.
Test_branch1_top1,test_branch2_top1,test_branch3_top1,test_branch4_top1 match the top-1 accuracies of aggragation of all features, the origin image branch, the discriminative branch and the complementary branch.
Thank you~ I fixed the problem according to your suggestion.
We define self.map_origin just for convenience when computing class response maps. So self.map_origin copies data of self.raw_classifier. self.raw_classifier is added to the graph in the forward when "agg_origin = self.raw_classifier(self.avg(feature_raw).view(-1,...
Could you please share your operation environment？
One of the most important requirements is using 2 Tesla P100. Other kinds of GPU, such as V100 or K80 will lead to slightly lower accuracy.
test_branch1_top1