/knowledgeDistillation

keras implementation of paper FRSKD: Refine Myself by Teaching Myself

Primary LanguagePython

biFPN

upSamp: Upsample(resize) + 1x1 DCB(align)
downSamp: maxpooling + 1x1 DCB
same-level path: 3x3 DCB
fuse: weight sum ('fast' mode)
通道对齐: 1x1 conv-bn-relu

depthConvBlock:
  ------ entry -----
  depth-wise conv: filters_in
  point-wise conv: filters_out
  BN
  ReLU
  ------ entry -----
  ------repeats by depth-------
  depth-wise conv: filters_out
  point-wise conv: filters_out
  BN
  ReLU
  ------repeats by depth-------  


efficientDet的biFPN:
upSamp: resize
downSamp: maxpooling
fuse: weight sum ('fast' mode)
通道对齐: 1x1 conv-bn-swish