How to add lasso (L1-norm) reisudial in Ceres Solver

Question

I want to add a regular term in least non-linear square function. How to do in Ceres Solver?

Answer 1

Pure L1-norm residual is non-differentiable and you cannot use it with ceres or for that matter any solver which depends on derivatives. However, Huber loss, or the SoftL1 loss included with Ceres are smooth approximations to L1 loss and you should be able to use them.

http://ceres-solver.org/nnls_modeling.html#instances