Act.h

/**
 * @file vAct.h
 * @author ae2f
 * @brief 
 * @version 0.1
 * @date 2025-02-13
 * 
 * @copyright Copyright (c) 2025
 * 
 */
 
#ifndef ae2fCL_AnnAct_h
#define ae2fCL_AnnAct_h
 
#include <stddef.h>
#include <ae2f/Float.h>
#include <ae2f/Guide.h>
 
/// @brief Customisable activasion function type.
typedef void ae2f_AnnActFFN_t (
        ae2f_float_t*               ret
        , ae2f_LP(count) const ae2f_float_t*    out
        , const size_t              index
        , const size_t              count
        );
 
#define ae2f_AnnActFFN_PASS(r, o, i, c)     *(r) = (o)[i]
#define ae2f_AnnActDerivFFN_PASS(r, o, i, c)    *(r) = 1.0
 
/// @brief
/// Specify the way of calculating loss.
/// 
/// This will be multiplied with learning rate, \n
/// and that value will be added to each weights field and bias.
/// @param out Predicted value
/// @param goal Expected value (wanted)
/// @param ret Calculated loss.
typedef void ae2f_AnnLossFFN_t (
        ae2f_float_t*               ret,
        ae2f_LP(count) const ae2f_float_t*  out,
        ae2f_LP(count) const ae2f_float_t*  goal,
        const size_t                index,
        const size_t                count
        );
 
#define ae2f_AnnLossDerivFFN_PASS(r, o, g, i, c)    *(r) = (o)[i] * (1.0 - (g)[i]);
 
typedef void ae2f_AnnActFwdMHATTN_t(
        ae2f_float_t*               ret,
        size_t                  prm_retidx,
        ae2f_LP(prm_len) const ae2f_float_t*    prm_inp,
        size_t                  prm_len
        );
 
typedef void ae2f_AnnActBwdMHATTN_t(
        ae2f_float_t*               ret,
        size_t                  prm_retidx,
        ae2f_LP(prm_len) const ae2f_float_t*    prm_grad_in,
        ae2f_LP(prm_len) const ae2f_float_t*    prm_softmax_out,
        size_t                  prm_len
        );
 
#endif