webassembly003 ggml GGML Tensor Library part-4 实现在浏览器端训练神经网络

code

#include "ggml.h"

#include 
#include 
#include 
#include 
#include 
#include "ggml.h"
#include 
#include 
#include 
#include 

#define MAX_NARGS 2

#if defined(__GNUC__)
#pragma GCC diagnostic ignored "-Wdouble-promotion"
#endif

//
// logging
//
#define GGML_DEBUG 0
#if (GGML_DEBUG >= 1)
#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
#else
#define GGML_PRINT_DEBUG(...)
#endif

#if (GGML_DEBUG >= 5)
#define GGML_PRINT_DEBUG_5(...) printf(__VA_ARGS__)
#else
#define GGML_PRINT_DEBUG_5(...)
#endif

#if (GGML_DEBUG >= 10)
#define GGML_PRINT_DEBUG_10(...) printf(__VA_ARGS__)
#else
#define GGML_PRINT_DEBUG_10(...)
#endif

#define GGML_PRINT(...) printf(__VA_ARGS__)


float frand(void) {
    return (float)rand()/(float)RAND_MAX;
}

int irand(int n) {
    return rand()%n;
}

void get_random_dims(int64_t * dims, int ndims) {
    dims[0] = dims[1] = dims[2] = dims[3] = 1;

    for (int i = 0; i < ndims; i++) {
        dims[i] = 1 + irand(4);
    }
}

void get_random_dims_minmax(int64_t * dims, int ndims, int min, int max) {
    dims[0] = dims[1] = dims[2] = dims[3] = 1;

    for (int i = 0; i < ndims; i++) {
        dims[i] = min + irand(max-min);
    }
}


struct ggml_tensor * get_random_tensor(
        struct ggml_context * ctx0,
        int ndims,
        int64_t ne[],
        float fmin,
        float fmax) {
    struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_F32, ndims, ne);

    switch (ndims) {
        case 1:
            for (int i0 = 0; i0 < ne[0]; i0++) {
                ((float *)result->data)[i0] = frand()*(fmax - fmin) + fmin;
            }
            break;
        case 2:
            for (int i1 = 0; i1 < ne[1]; i1++) {
                for (int i0 = 0; i0 < ne[0]; i0++) {
                    ((float *)result->data)[i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
                }
            }
            break;
        case 3:
            for (int i2 = 0; i2 < ne[2]; i2++) {
                for (int i1 = 0; i1 < ne[1]; i1++) {
                    for (int i0 = 0; i0 < ne[0]; i0++) {
                        ((float *)result->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
                    }
                }
            }
            break;
        case 4:
            for (int i3 = 0; i3 < ne[3]; i3++) {
                for (int i2 = 0; i2 < ne[2]; i2++) {
                    for (int i1 = 0; i1 < ne[1]; i1++) {
                        for (int i0 = 0; i0 < ne[0]; i0++) {
                            ((float *)result->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
                        }
                    }
                }
            }
            break;
        default:
            assert(false);
    };

    return result;
}

float get_element(const struct ggml_tensor * t, int idx) {
    return ((float *)t->data)[idx];
}

void set_element(struct ggml_tensor * t, int idx, float value) {
    ((float *)t->data)[idx] = value;
}



float run(int val){
    return 0.1f;
}


struct ggml_init_params params = {
        /* .mem_size   = */ 1024*1024*10, // 设置内存大小为1GB
        /* .mem_buffer = */ NULL, // 设置内存缓冲区为空
        /* .no_alloc   = */ false, // 允许内存分配
};
struct ggml_context * ctx = ggml_init(params); // 创建一个ggml_context结构体对象

int64_t ne1[4] = {2, 512, 1, 1};
int64_t ne2[4] = {512, 256, 1, 1};
int64_t ne3[4] = {1, 256, 1, 1};
int64_t ne4[4] = {256, 3, 1, 1};
int64_t ne5[4] = {3, 1, 1, 1};

int64_t mid_w[4] = {256, 256, 1, 1};
int64_t mid_b[4] = {256, 1, 1, 1};

struct res {ggml_cgraph ge;ggml_tensor * e;ggml_tensor * output;ggml_tensor * intput;ggml_tensor * target;};

 res hf_getcompute_graph(){

    struct ggml_tensor * X = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 2);//    struct ggml_tensor * X = get_random_tensor(ctx, 2, ne1, -1, +1); // 创建一个2维随机张量a
    struct ggml_tensor * target = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3);

    struct ggml_tensor * proj = get_random_tensor(ctx, 2, ne1, -1, +1);
    struct ggml_tensor * fc1_weight = get_random_tensor(ctx, 2, ne2, -1, +1);
    struct ggml_tensor * fc1_bias = get_random_tensor(ctx, 2, ne3, -1, +1);
    struct ggml_tensor * fcout_weight = get_random_tensor(ctx, 2, ne4, -1, +1);
    struct ggml_tensor * fcout_bias = get_random_tensor(ctx, 2, ne5, -1, +1);

    struct ggml_tensor * fc3_weight = get_random_tensor(ctx, 2, mid_w, -1, +1);
    struct ggml_tensor * fc3_bias = get_random_tensor(ctx, 2, mid_b, -1, +1);
    struct ggml_tensor * fc4_weight = get_random_tensor(ctx, 2, mid_w, -1, +1);
    struct ggml_tensor * fc4_bias = get_random_tensor(ctx, 2, mid_b, -1, +1);
    struct ggml_tensor * fc5_weight = get_random_tensor(ctx, 2, mid_w, -1, +1);
    struct ggml_tensor * fc5_bias = get_random_tensor(ctx, 2, mid_b, -1, +1);

    ggml_set_param(ctx, fcout_weight); ggml_set_param(ctx, fcout_bias); ggml_set_param(ctx, fc3_weight); ggml_set_param(ctx, fc3_bias);
    ggml_set_param(ctx, fc4_weight); ggml_set_param(ctx, fc4_bias); ggml_set_param(ctx, fc5_weight); ggml_set_param(ctx, fc5_bias);

    ggml_tensor * proj_of_x =  ggml_mul_mat(ctx, proj,X); // 512*1
    ggml_tensor * fc1 = ggml_add(ctx, ggml_mul_mat(ctx, proj_of_x,fc1_weight),  fc1_bias); // 1*256
    ggml_tensor * fc1_ = ggml_view_2d(ctx,fc1,256,1,4*256,0);//ggml_transpose(ctx,fc1);  // 256*1

    // ggml_tensor * fc2 =  ggml_mul_mat(ctx,fc3_weight,fc1_); // 256*1 ggml_tensor * fc2 =  ggml_mul_mat(ctx,fc1_,fc3_weight); // 256*1 &  256*256 =>1*256
    ggml_tensor * fc3 = ggml_add(ctx, ggml_mul_mat(ctx, fc3_weight, ggml_relu(ctx, fc1_)), fc3_bias);
    ggml_tensor * fc4 = ggml_add(ctx, ggml_mul_mat(ctx, fc4_weight, ggml_relu(ctx, fc3)), fc4_bias);
    ggml_tensor * fc5 = ggml_add(ctx, ggml_mul_mat(ctx, fc5_weight, ggml_relu(ctx, fc4)), fc5_bias);

    ggml_tensor * fcout = ggml_add(ctx, ggml_mul_mat(ctx, fcout_weight, ggml_relu(ctx, fc5)), fcout_bias);// 3*1
    ggml_tensor * fcout_ = ggml_sub(ctx,fcout,target);

    struct ggml_tensor * e  = ggml_sum(ctx, ggml_sqr(ctx, fcout_) ); // 计算张量d的平方和e

    struct ggml_cgraph ge = ggml_build_forward(e); // 构建计算图ge
    return {ge, e, fcout, X,target};

}


void hf_out(ggml_tensor * output){
    printf("f = %f\n", ggml_get_f32_1d(output, 0));//    cout<< "res "<<((float *)(fc2->data))[0];
    printf("f = %f\n", ggml_get_f32_1d(output, 1));
    printf("f = %f\n", ggml_get_f32_1d(output, 2));
}

void hf_free(){
    ggml_free(ctx); // 释放上下文内存
}

void hf_set_data_random(ggml_tensor * input,ggml_tensor * target){
    std::random_device rd;  // 获取随机数种子
    std::mt19937 gen(rd()); // 使用随机数种子初始化随机数生成器
    std::uniform_real_distribution dis(0.0, 1.0);  // 定义均匀分布函数，范围为[0.0, 1.0)
    std::vector digit(512); // 创建大小为512的vector
    for (int i = 0; i < digit.size(); i++) {
        digit[i] = dis(gen);  // 使用分布函数生成随机数，并赋值给vector的每个元素
    }
    memcpy(input->data, digit.data(), ggml_nbytes(input));
    digit = {100.3f,200.5f,100.1f};
    memcpy(target->data, digit.data(), ggml_nbytes(target));

}

void hf_set_data(ggml_tensor * input,ggml_tensor * target,vector& v_in ,vector& v_out){
    memcpy(input->data, v_in.data(), ggml_nbytes(input));
    memcpy(target->data, v_out.data(), ggml_nbytes(target));
}

res a = hf_getcompute_graph();


#ifdef __cplusplus
extern "C" {
#endif

struct rgb{ float r; float g; float b;};

float hf_play_test(int istrain) {


    ggml_cgraph ge = a.ge;
    ggml_tensor * e = a.e;
    ggml_tensor * output = a.output;
    ggml_tensor * input = a.intput;
    ggml_tensor * target = a.target;

    hf_set_data_random(input,target);

    if(istrain == 1){
        struct ggml_opt_params opt_params = ggml_opt_default_params(GGML_OPT_ADAM); // 获取默认的优化参数

        ggml_opt(ctx, opt_params, e); // 通过指定的优化参数优化张量e

        ggml_graph_reset(&ge); // 重置计算图

        ggml_graph_compute_with_ctx(ctx, &ge, /*n_threads*/ 1); // 使用指定上下文计算计算图
    }else{
        ggml_graph_reset(&ge); // 重置计算图
        ggml_graph_compute_with_ctx(ctx, &ge, /*n_threads*/ 1); // 使用指定上下文计算计算图
    }


    hf_out(output);
    float fe = ggml_get_f32_1d(e, 0); // 获取张量e的第一个元素值
    printf("%s: e = %.4f\n", __func__, fe); // 输出e的值
    return fe;

}



rgb hf_play_set_val(int istrain,vector& v_in ,vector& v_out) {

    //res a = hf_getcompute_graph();
    ggml_cgraph ge = a.ge;
    ggml_tensor * e = a.e;
    ggml_tensor * output = a.output;
    ggml_tensor * input = a.intput;
    ggml_tensor * target = a.target;

    hf_set_data(input,target,v_in,v_out);

    if(istrain == 1){
        struct ggml_opt_params opt_params = ggml_opt_default_params(GGML_OPT_ADAM); // 获取默认的优化参数

        ggml_opt(ctx, opt_params, e); // 通过指定的优化参数优化张量e

        ggml_graph_reset(&ge); // 重置计算图

        ggml_graph_compute_with_ctx(ctx, &ge, /*n_threads*/ 1); // 使用指定上下文计算计算图
    }else{
        ggml_graph_reset(&ge); // 重置计算图
        ggml_graph_compute_with_ctx(ctx, &ge, /*n_threads*/ 1); // 使用指定上下文计算计算图
    }


    hf_out(output);
    float fe = ggml_get_f32_1d(e, 0); // 获取张量e的第一个元素值
    printf("%s: e = %.4f\n", __func__, fe); // 输出e的值

    return {ggml_get_f32_1d(output, 0),ggml_get_f32_1d(output, 1),ggml_get_f32_1d(output, 2)};

}

#ifdef __cplusplus
}
#endif

int main(void) {
//    hf_play_test(0);
//    hf_play_test(1);

    vector v_in; vector v_out;

    v_in={1.0f,1.0f}; v_out = {100.3f,200.5f,100.1f};

    hf_play_set_val(0,v_in,v_out);

    hf_play_set_val(1,v_in,v_out);

    hf_play_set_val(0,v_in,v_out);

    return 0;
}

浏览器端训练神经网络

• emcc -I~/***/76/ ~/***/76/ggml.c ~/leetcode/76/main.cpp -o ~/leetcode/76/hf.js -s EXPORTED_FUNCTIONS='["_hf_play_test","_hf_play_set_val","_malloc","_free"]' -s EXPORTED_RUNTIME_METHODS='["ccall"]' -s ALLOW_MEMORY_GROWTH=1

• 解决
• emcc -I~/***/76/ ~/***/76/ggml.c ~/leetcode/76/main.cpp -o ~/leetcode/76/hf.html -s EXPORTED_FUNCTIONS='["_hf_play_set_val","_hf_play_test","_malloc","_free"]' -s EXPORTED_RUNTIME_METHODS='["ccall"]' -s STACK_SIZE=955360

CG

• 更多网络的训练方法: 神经网络的极限训练方法：gradient checkpointing
• ggml_tensor * fc2 = ggml_mul_mat(ctx,fc1_,fc3_weight); // 256*1 & 256*256 =>1*256
• hf_play_set_val’ has C-linkage specified, but returns user-defined type ‘vector’ which is incompatible with C [-Wreturn-type-c-linkage]
• https://github.com/cython/cython/issues/1839
• https://stackoverflow.com/questions/22901697/error-in-c-code-linkage-warning-c4190-type-has-c-linkage-specified-but-retu