I read several papers, where they propose to implement residual blocks of ResNet as follows $$ u^{k+1} = u^k - \tau K^T \sigma(K u^k), $$ where $u^{k}$ denotes output on k-th layer, $\tau$ is artificial time-step and K is convolution matrix. The ResNet with this type of blocks should be more numerically stable as with more common block: $u^{k+1} = u^k - \tau K_2 \sigma(K_1 u^k)$.

How does one share weights and implement $K^T$ operation using keras/tf?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.