cwenzi/neuroflow-cpp

1// NeuroFlow API在线学习测试2// 测试与API集成的能力3 4#include <iostream>5#include <fstream>6#include <sstream>7#include "../include/neuroflow/online_learning.hpp"8#include "../include/neuroflow/model.hpp"9 10using namespace neuroflow;11 12// 简化的JSON解析（用于读取API生成的训练数据）13std::string read_file(const std::string& path) {14    std::ifstream file(path);15    if (!file.is_open()) {16        return "";17    }18    std::stringstream buffer;19    buffer << file.rdbuf();20    return buffer.str();21}22 23// 测试API训练数据加载24void test_api_data_loading() {25    std::cout << "\n=== API Data Loading Test ===\n";26    27    // 尝试读取API生成的数据28    std::string data = read_file("api_training_data/neuroflow_training_data.json");29    30    if (data.empty()) {31        std::cout << "  No API training data found. Run api_train.py first.\n";32        std::cout << "  Example: python api_train.py --api deepseek --key YOUR_KEY --task knowledge\n";33        return;34    }35    36    std::cout << "  API training data loaded: " << data.size() << " bytes\n";37    std::cout << "  (Full parsing requires JSON library)\n";38}39 40// 测试在线学习能力41void test_online_learning_capability() {42    std::cout << "\n=== Online Learning Capability Test ===\n";43    44    // 创建模型（使用Config）45    NeuroFlowModel::Config cfg;46    cfg.input_dim = 512;47    cfg.hidden_dim = 256;48    cfg.output_dim = 10;49    NeuroFlowModel model(cfg);50    51    // 创建在线学习器52    Optimizer optimizer(0.01f);53    54    // 单样本快速适应55    Tensor input(std::vector<size_t>{1, 512}, QuantType::FP32);56    Tensor target(std::vector<size_t>{1, 10}, QuantType::FP32);57    58    // 初始化随机数据59    float* inp = input.as_fp32();60    float* tgt = target.as_fp32();61    for (size_t i = 0; i < 512; ++i) inp[i] = (rand() / RAND_MAX - 0.5f) * 0.1f;62    for (size_t i = 0; i < 10; ++i) tgt[i] = (i == 3) ? 1.0f : 0.0f;63    64    // 前向传播65    NeuroFlowModel::Output output = model.forward(input);66    67    // 计算初始损失68    float initial_loss = LossFunctions::mse(output.output, target);69    70    // 执行记忆巩固 - 使用 hidden_dim 而非原始 input71    Tensor h_input(std::vector<size_t>{1, cfg.hidden_dim}, QuantType::FP32);72    float* h_inp = h_input.as_fp32();73    for (size_t i = 0; i < cfg.hidden_dim; ++i) h_inp[i] = inp[i % cfg.input_dim] * 0.5f;74    model.memory->consolidate(h_input);75    76    // 再次前向传播77    NeuroFlowModel::Output output2 = model.forward(input);78    float final_loss = LossFunctions::mse(output2.output, target);79    80    std::cout << "  Single sample adaptation:\n";81    std::cout << "    Initial loss: " << initial_loss << "\n";82    std::cout << "    Final loss: " << final_loss << "\n";83    std::cout << "    Loss reduction: " << (initial_loss - final_loss) << "\n";84    85    // 记忆巩固测试86    std::cout << "  Memory consolidation test:\n";87    float mem_change = model.memory->ltp_rate;88    std::cout << "    LTP rate: " << mem_change << "\n";89    std::cout << "    Memory slots: " << model.memory->memory_slots << "\n";90    91    std::cout << "  [PASS] Online learning capability verified\n";92}93 94// 测试知识注入95void test_knowledge_injection() {96    std::cout << "\n=== Knowledge Injection Test ===\n";97    98    // 创建模型99    NeuroFlowModel::Config cfg;100    NeuroFlowModel model(cfg);101    102    // 模拟知识注入（使用记忆巩固）- 使用 hidden_dim 维度103    Tensor knowledge(std::vector<size_t>{32, cfg.hidden_dim}, QuantType::FP32);104    105    // 执行多次记忆巩固106    for (int i = 0; i < 10; ++i) {107        model.memory->consolidate(knowledge);108    }109    110    std::cout << "  Injected " << 10 << " batches of knowledge\n";111    std::cout << "  Memory slots used: " << model.memory->memory_slots << "\n";112    113    // 测试检索114    Tensor query(std::vector<size_t>{1, cfg.hidden_dim}, QuantType::FP32);115    auto retrieved = model.memory->retrieve(query);116    117    std::cout << "  Retrieved memory shape: " << retrieved.retrieved.shape_[0] 118              << " x " << retrieved.retrieved.shape_[1] << "\n";119    120    std::cout << "  [PASS] Knowledge injection verified\n";121}122 123// 测试API增强推理124void test_api_enhanced_reasoning() {125    std::cout << "\n=== API Enhanced Reasoning Test ===\n";126    127    // 模拟API增强流程128    std::cout << "  API enhancement pipeline:\n";129    std::cout << "    1. Local model forward pass\n";130    std::cout << "    2. API call for complex reasoning\n";131    std::cout << "    3. Combine results\n";132    133    // 创建模型134    NeuroFlowModel::Config cfg;135    NeuroFlowModel model(cfg);136    137    // 本地推理138    Tensor input(std::vector<size_t>{1, cfg.input_dim}, QuantType::FP32);139    NeuroFlowModel::Output output = model.forward(input);140    141    std::cout << "  Local reasoning output: " << output.output.shape_[0] 142              << " x " << output.output.shape_[1] << "\n";143    144    // API推理（模拟）145    std::cout << "  API reasoning: (requires python api_train.py)\n";146    std::cout << "  - DeepSeek API: https://api.deepseek.com\n";147    std::cout << "  - GLM-4 API: https://open.bigmodel.cn\n";148    149    std::cout << "  [INFO] Use python for actual API calls\n";150}151 152int main() {153    std::cout << "=============================================\n";154    std::cout << "  NeuroFlow API Online Learning Test Suite\n";155    std::cout << "=============================================\n";156    157    test_api_data_loading();158    test_online_learning_capability();159    test_knowledge_injection();160    test_api_enhanced_reasoning();161    162    std::cout << "\n=============================================\n";163    std::cout << "  All tests completed!\n";164    std::cout << "=============================================\n";165    166    std::cout << "\nAPI Training Usage:\n";167    std::cout << "  python api_train.py --api deepseek --key YOUR_KEY --task knowledge\n";168    std::cout << "  python api_train.py --api deepseek --key YOUR_KEY --task code\n";169    std::cout << "  python api_train.py --api deepseek --key YOUR_KEY --task reasoning\n";170    std::cout << "  python api_train.py --api deepseek --key YOUR_KEY --task full\n";171    172    return 0;173}