Open super-dainiu opened 1 year ago
Hi super-dainiu, thank you for your interest.
POET uses Gurobi for solving the MILP formulation. Gurobi can be replaced by cvxpy/coin-or or any open-source ILP solvers. We generate the graph and the corresponding dependencies ourselves. The network is defined using Pytorch-compatible methods.
To further highlight the benefit, POFO supports chain (linear) model graphs while POET is more generic and supports any arbitrary model graphs such as ResNets/BERTs etc.
Dear Authors, I am amazed that the performance of POET outperforms POFO so much. Since PyTorch does not have a traced backward graph, I am curious about which framework you are using.