MlpTrainXOR-Made-Primal.c

Go to the documentation of this file.

#define ae2f_NEED_CLASS 0
 
#include <assert.h>
#include <ae2f/Ann/Act.h>
 
#include <ae2f/Ann/Mlp.h>
#include <ae2f/Ann/Slp.h>
#include <stdio.h>
#include <math.h>
#include <stdlib.h>
#include <time.h>
 
static void Act(ae2f_float_t* r, ae2f_float_t x) {
    *(r) = 1.0 / (1.0 + exp(-x));
}
 
static void ActDeriv(ae2f_float_t* r, ae2f_float_t output) {
    *r = output * (1.0 - output);
}
 
static void LossDeriv(ae2f_float_t* r, const ae2f_float_t* output, const ae2f_float_t* target, size_t i, size_t c) {
    *r = ((output[i] - target[i]) / c);
}
 
const ae2f_float_t inp[4][2] = {
    {0, 0},
    {0, 1},
    {1, 0},
    {1, 1}
};
 
const ae2f_float_t goal_xor[4] = {0, 1, 1, 0};
 
ae2f_float_t output[1] = { 0 };
 
#define MLP_DEPTH 4
#define MLP_IN 2
#define MLP_HID 3
#define MLP_OUT 1
#define MLP_OUT_GREATEST MLP_HID // The greatest number of outputs from any single layer
 
ae2f_AnnMlp_t mlp;
ae2f_AnnSlp_t slplast;
 
const size_t mlp_szv[] = { MLP_IN, MLP_HID, 1, MLP_OUT };
 
// --- Corrected Memory Layout ---
// The MLP functions expect a flat, padded memory layout for weights and biases.
// Each layer's weights are expected to start at a fixed stride.
#define MLP_WEIGHT_STRIDE (MLP_OUT_GREATEST * MLP_OUT_GREATEST)
#define MLP_BIAS_STRIDE (MLP_OUT_GREATEST)
 
ae2f_float_t mlp_weights[(MLP_DEPTH - 1) * MLP_WEIGHT_STRIDE] = {0};
ae2f_float_t mlp_bias[(MLP_DEPTH - 1) * MLP_BIAS_STRIDE] = {0};
 
ae2f_float_t mlp_outstream[(MLP_DEPTH - 1) * MLP_OUT_GREATEST] = {0};
ae2f_float_t mlp_deltastream[(MLP_DEPTH - 1) * MLP_OUT_GREATEST] = {0};
// --- End Corrected Memory Layout ---
 
ae2f_AnnAct_t* mlp_acts[MLP_DEPTH - 1] = { Act, Act, Act };
ae2f_AnnAct_t* mlp_actderivs[MLP_DEPTH - 1] = { ActDeriv, ActDeriv, ActDeriv };
 
size_t i, j, k;
 
union TEST_STACK {
    ae2f_AnnMlpPredictStream_t  m_predictsteam;
    ae2f_AnnSlpFetchDelta_t     m_fetch;
    ae2f_AnnMlpFollow_t     m_propagate;
    ae2f_AnnMlpTrain_t      m_train;
} __test_stack;
 
int main() {
    puts("MlpTrainXOR-Primal start");
 
    puts("Configuring mlp");
    mlp.m_outc = MLP_OUT_GREATEST;
    mlp.m_depth = MLP_DEPTH;
 
    puts("Configuring last slp");
    slplast.m_inc = MLP_HID;
    slplast.m_outc = MLP_OUT;
 
    srand(0);
 
    puts("Initializing weights randomly with correct memory layout");
    // Layer 0: 2 inputs -> 3 neurons
    size_t weight_base_l0 = 0 * MLP_WEIGHT_STRIDE;
    size_t bias_base_l0 = 0 * MLP_BIAS_STRIDE;
    for (i = 0; i < mlp_szv[1]; i++) { // 3 output neurons
        for (k = 0; k < mlp_szv[0]; k++) { // 2 input weights
            mlp_weights[weight_base_l0 + i * mlp_szv[0] + k] = ((double)rand() / RAND_MAX) - 0.5;
        }
        mlp_bias[bias_base_l0 + i] = ((double)rand() / RAND_MAX) - 0.5;
    }
 
    // Layer 1: 3 inputs -> 1 neuron
    size_t weight_base_l1 = 1 * MLP_WEIGHT_STRIDE;
    size_t bias_base_l1 = 1 * MLP_BIAS_STRIDE;
    for (i = 0; i < mlp_szv[2]; i++) { // 1 output neuron
        for (k = 0; k < mlp_szv[1]; k++) { // 3 input weights
            mlp_weights[weight_base_l1 + i * mlp_szv[1] + k] = ((double)rand() / RAND_MAX) - 0.5;
        }
        mlp_bias[bias_base_l1 + i] = ((double)rand() / RAND_MAX) - 0.5;
    }
 
    puts("See first output (before training)");
    for(i = 0; i < 4; ++i) {
        __ae2f_AnnMlpPredictStream_imp(
                __test_stack.m_predictsteam, mlp, inp[i], output, mlp_szv,
                mlp_weights, mlp_bias, mlp_outstream, mlp_acts
                );
        printf("Initial Output for [%d, %d]: %f (goal: %d)\n"
                , (int)inp[i][0], (int)inp[i][1], output[0], (int)goal_xor[i]);
    }
 
    puts("Training...");
    for(j = 0; j < 99000; ++j) {
        for(i = 0; i < 4; ++i) {
#if 0
            __ae2f_AnnMlpPredictStream_imp(
                    __test_stack.m_predictsteam, mlp, inp[i], output, mlp_szv,
                    mlp_weights, mlp_bias, mlp_outstream, mlp_acts
                    );
 
            __ae2f_AnnSlpFetchDelta_imp(
                    __test_stack.m_fetch, slplast, output, &goal_xor[i],
                    ActDeriv, LossDeriv, &mlp_deltastream[MLP_OUT_GREATEST]
                    );
 
            __ae2f_AnnMlpFollow_imp(
                    __test_stack.m_propagate, mlp, inp[i]
                    , &mlp_deltastream[MLP_OUT_GREATEST],
                    mlp_szv, mlp_outstream, mlp_deltastream,
                    mlp_weights, mlp_bias,
                    0.6, 0.5,
                    mlp_actderivs
                    );
#else
            __ae2f_AnnMlpTrainPrimal_imp(
                    -1, ae2f_NONE
                    , __test_stack.m_train, mlp
                    , inp[i], output,  &goal_xor[i]
                    , mlp_szv, mlp_outstream
                    , mlp_deltastream, mlp_weights, mlp_bias
                    , 0.2, 0.3
                    , mlp_acts, mlp_actderivs, LossDeriv
                    );
#endif
        }
    }
    puts("Training complete.");
 
    puts("See last output after training");
    for(i = 0; i < 4; ++i) {
        __ae2f_AnnMlpPredictStream_imp(
                __test_stack.m_predictsteam, mlp, inp[i], output, mlp_szv,
                mlp_weights, mlp_bias, mlp_outstream, mlp_acts
                );
        printf("Final Output for [%d, %d]: %f (goal: %d)\n"
                , (int)inp[i][0], (int)inp[i][1], output[0], (int)goal_xor[i]
              );
    }
 
    return 0;
}