Function ggml_opt_fit

ggml/src/ggml-opt.cpp:999–1079

}

void ggml_opt_fit(
        ggml_backend_sched_t            backend_sched,
        ggml_context                  * ctx_compute,
        ggml_tensor                   * inputs,
        ggml_tensor                   * outputs,
        ggml_opt_dataset_t              dataset,
        enum ggml_opt_loss_type         loss_type,
        enum ggml_opt_optimizer_type    optimizer,
        ggml_opt_get_optimizer_params   get_opt_pars,
        int64_t                         nepoch,
        int64_t                         nbatch_logical,
        float                           val_split,
        bool                            silent) {
    ggml_time_init();
    const int64_t t_start_us = ggml_time_us();

    const int64_t ndata           = ggml_opt_dataset_data(dataset)->ne[1];
    const int64_t nbatch_physical = inputs->ne[1];
    GGML_ASSERT(ndata          % nbatch_logical  == 0);
    GGML_ASSERT(nbatch_logical % nbatch_physical == 0);

    const int64_t opt_period       = nbatch_logical / nbatch_physical;
    const int64_t nbatches_logical = ndata / nbatch_logical;

    GGML_ASSERT(val_split >= 0.0f);
    GGML_ASSERT(val_split <  1.0f);
    const int64_t ibatch_split = int64_t(((1.0f - val_split) * nbatches_logical)) * opt_period; // train <-> val split index (physical)
    const int64_t idata_split  = ibatch_split * nbatch_physical;

    int64_t epoch = 1;

    ggml_opt_params params = ggml_opt_default_params(backend_sched, loss_type);
    params.ctx_compute     = ctx_compute;
    params.inputs          = inputs;
    params.outputs         = outputs;
    params.opt_period      = opt_period;
    params.get_opt_pars    = get_opt_pars;
    params.get_opt_pars_ud = &epoch;
    params.optimizer       = optimizer;
    ggml_opt_context_t opt_ctx = ggml_opt_init(params);

    // Shuffling the data is generally useful but there is only a point if not all data is used in a single batch.
    if (nbatch_logical < ndata) {
        ggml_opt_dataset_shuffle(opt_ctx, dataset, -1); // Shuffle all data (train + validation).
    }

    ggml_opt_result_t result_train = ggml_opt_result_init();
    ggml_opt_result_t result_val   = ggml_opt_result_init();

    ggml_opt_epoch_callback epoch_callback = silent ? nullptr : ggml_opt_epoch_callback_progress_bar;

    for (; epoch <= nepoch; ++epoch) {
        if (nbatch_logical < idata_split) {
            ggml_opt_dataset_shuffle(opt_ctx, dataset, idata_split);
        }

        ggml_opt_result_reset(result_train);
        ggml_opt_result_reset(result_val);