Better sampler_t interface - Githubissues

accosmin / nano

C++ library [machine learning & numerical optimization] - superseeded by libnano

MIT License

1 stars 0 forks source link

Better sampler_t interface #62

Closed accosmin closed 8 years ago

accosmin commented 8 years ago

Confusing sampling:

the stochastic trainer uses trainer_data_t::set_batch
the minibatch trainer uses two custom functions (setup_minibatch & reset_minibatch) to keep the minibatch fixed for a short number of batch iterations

This should be solved by having a stack of sampling commands: sampler_t(task).push(batch size = 0, {annotated|all})

Usage:

batch: nothing (use the training sampler as given)
minibatch: [sampler_t::push(minibatch_size, fixed) ... train ... sampler_t::pop() ... evaluate]+
stochastic: [sampler_t::push(minibatch_size, random) ... train ... sampler_t::pop() ... evaluate]+