Chris B

hmm . . still struggling . . here are the networks: template <template <typename> class BN, typename SUBNET> using block = BN<dlib::con<256, 3, 3, 1, 1, dlib::relu<BN<dlib::con<256, 3, 3, 1, 1, SUBNET>>>>>; template < template <template<typename>class, typename> class block, template<typename>class BN, typename SUBNET> using residual = dlib::add_prev1<block<BN, dlib::tag1<SUBNET>>>; template <typename SUBNET> using res = dlib::relu<residual<block, dlib::bn_con, SUBNET>>; using Tower = dlib::repeat<20,...

thank you for creating dlib . . i am naively (newb) trying to follow the creation of googles alphago network. a python implementation may be found here: https://github.com/gcp/leela-zero/blob/master/training/tf/tfprocess.py the network is a 20 or 40 block res-net feeding two output heads (policy and value) the softmax policy loss must be summed with the MSE loss of the value head + an L2 term to create a total trainable loss i define the base tower with : using tower = dlib::repeat<20, res, dlib::relu<dlib::bn_con<dlib::con<256,...