SuperNova
Results
3
comments of
SuperNova
`ChannelWiseDivergence` is a response-based kd for semantic segmentation. Please use kl_div for cls head, it should work. Besides, if you want to do distillation on Reg heads, you can try...
dafl的训练似乎十分不稳定，同样的超参数设置，不同的随机数种子，结果差的很远 不设置随机数种子，精度能达到88以上，设置种子为12345以后，精度出现了下降 以上结果皆是运行在8卡机器，每个卡上的batch_size为128，总1024，损失的权重保持此仓库的默认设置。
Same question. Where is the pre-inverted dataset from?