TensorFlow C API加载.pb模型遇localhost容器不存在错误，求C实现变量初始化方案

错误信息 #

Failed precondition: Error while reading resource variable value_dense_2/bias from Container: localhost. This could mean that the variable was uninitialized. Not found: Container localhost does not exist. (Could not find resource: localhost/value_dense_2/bias) [[{{node nnue_model/value_dense_2/BiasAdd/ReadVariableOp}}]] [[StatefulPartitionedCall/_35]]

问题本质 #

该错误源于模型内的变量未完成初始化，对应Python中通过tf.global_variables_initializer()初始化全局变量并运行操作的解决方案，在C语言中需通过TensorFlow C API完成等价操作。

C语言实现方案 #

在TensorFlow C API中，需完成以下步骤初始化变量：

• 加载模型图并创建会话后，查找全局变量初始化操作节点（通常节点名为init或global_variables_initializer）；
• 创建并运行该初始化操作的会话任务。

代码示例 #

#include <tensorflow/c/c_api.h>
#include <stdio.h>
#include <stdlib.h>

// 辅助函数：读取二进制模型文件
TF_Buffer* ReadBinaryFile(const char* filename, TF_Status* status) {
    FILE* f = fopen(filename, "rb");
    if (f == NULL) {
        TF_SetStatus(status, TF_NOT_FOUND, "Failed to open model file");
        return NULL;
    }
    fseek(f, 0, SEEK_END);
    long fsize = ftell(f);
    fseek(f, 0, SEEK_SET);
    
    void* data = malloc(fsize);
    fread(data, 1, fsize, f);
    fclose(f);
    
    TF_Buffer* buf = TF_NewBuffer();
    buf->data = data;
    buf->length = fsize;
    buf->data_deallocator = free;
    return buf;
}

int main() {
    // 初始化TensorFlow资源
    TF_Graph* graph = TF_NewGraph();
    TF_Status* status = TF_NewStatus();
    TF_SessionOptions* sess_opts = TF_NewSessionOptions();

    // 加载.pb模型文件（替换为你的模型路径）
    TF_Buffer* model_buf = ReadBinaryFile("your_model.pb", status);
    if (TF_GetCode(status) != TF_OK) {
        fprintf(stderr, "%s\n", TF_Message(status));
        return 1;
    }
    TF_ImportGraphDefOptions* import_opts = TF_NewImportGraphDefOptions();
    TF_GraphImportGraphDef(graph, model_buf, import_opts, status);
    TF_DeleteImportGraphDefOptions(import_opts);
    TF_DeleteBuffer(model_buf);

    // 创建会话
    TF_Session* session = TF_NewSession(graph, sess_opts, status);
    TF_DeleteSessionOptions(sess_opts);
    if (TF_GetCode(status) != TF_OK) {
        fprintf(stderr, "Failed to create session: %s\n", TF_Message(status));
        return 1;
    }

    // 获取全局变量初始化操作节点
    TF_Operation* init_op = TF_GraphOperationByName(graph, "init");
    if (init_op == NULL) {
        init_op = TF_GraphOperationByName(graph, "global_variables_initializer");
    }
    if (init_op == NULL) {
        fprintf(stderr, "Failed to find initialization operation node\n");
        goto cleanup;
    }

    // 运行初始化操作
    TF_Run(session, NULL, NULL, NULL, 0, &init_op, NULL, 0, NULL, 0, NULL, status);
    if (TF_GetCode(status) != TF_OK) {
        fprintf(stderr, "Failed to run init operation: %s\n", TF_Message(status));
        goto cleanup;
    }

    // 此处添加模型推理代码
    // ...

cleanup:
    // 释放资源
    TF_DeleteSession(session, status);
    TF_DeleteGraph(graph);
    TF_DeleteStatus(status);
    return 0;
}

注意事项 #

• 若模型导出时未包含初始化节点，需在Python环境中先运行变量初始化，再导出为.pb文件；
• 可通过TensorBoard查看模型图结构，确认初始化操作的实际节点名称。

内容的提问来源于stack exchange，提问作者Gustasvs