Allow pycaffe to update solver parameters, e.g., learning rate and momentum

[PiranjaF]

Many SGD schedules specify that the learning rate should be annealed when validation accuracy has somewhat converged. Similarly, momentum should be around 0.5 in the beginning (http://ufldl.stanford.edu/tutorial/supervised/OptimizationStochasticGradientDescent/) and then to 0.9 when learning is stabilized.

Can these parameters be changed during training from the Python interface? Having this as an option seems like an obvious enhancement.

[samuelgoodell]

I could really use this in my code as well.

[shelhamer]

This is certainly worth doing and only requires minor coding to make the hyperparams for SGDSolver mutable and exposed through boost::python.

As a hack this can be done by running the solver through gdb -- not that I'm really suggesting that.

[talda]

@shelhamer a much simpler hack would be to train with an initial set of parameters and than change the parameters and fine tune on each step of parameter change.

[pgrossmann]

@talda Would you mind sharing an example of how to change any of the hyperparameters (e.g., the learning rate) using PyCaffe? I tried inspecting the solver objects, but I couldn't find the variable that stores that parameter. Thank you!

[ravi-june]

For solution, see http://stackoverflow.com/questions/31823898/changing-the-solver-parameters-in-caffe-through-pycaffe

[mitar]

+1 on allowing one to change the learning rate and other parameters.

[hala3]

i want to finetune a pretrained model which trained with pycaffe
are there any tutorial can help me please ??