TensorFlow 2 C++部署saved_model.pb预测时出现越界错误求助

我在C++中使用训练好的saved_model.pb，以文本为输入获取预测类别ID和分数时，出现如下错误：
W tensorflow/core/framework/op_kernel.cc:1767] OP_REQUIRES failed at strided_slice_op.cc:108 : Invalid argument: slice index 0 of dimension 0 out of bounds.

环境：TensorFlow 2.5，模型由Python训练得到。

saved_model_cli输出的SignatureDef信息：

MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:

signature_def['classification']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['inputs'] tensor_info:
        dtype: DT_STRING
        shape: (-1)
        name: Placeholder:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['classes'] tensor_info:
        dtype: DT_STRING
        shape: (-1, 74)
        name: head/Tile:0
    outputs['scores'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 74)
        name: head/predictions/probabilities:0
  Method name is: tensorflow/serving/classify

signature_def['predict']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['content'] tensor_info:
        dtype: DT_STRING
        shape: (-1)
        name: Placeholder:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['all_class_ids'] tensor_info:
        dtype: DT_INT32
        shape: (-1, 74)
        name: head/predictions/Tile:0
    outputs['all_classes'] tensor_info:
        dtype: DT_STRING
        shape: (-1, 74)
        name: head/predictions/Tile_1:0
    outputs['class_ids'] tensor_info:
        dtype: DT_INT64
        shape: (-1, 1)
        name: head/predictions/ExpandDims:0
    outputs['classes'] tensor_info:
        dtype: DT_STRING
        shape: (-1, 1)
        name: head/predictions/hash_table_Lookup/LookupTableFindV2:0
    outputs['logits'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 74)
        name: add:0
    outputs['probabilities'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 74)
        name: head/predictions/probabilities:0
  Method name is: tensorflow/serving/predict

signature_def['serving_default']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['inputs'] tensor_info:
        dtype: DT_STRING
        shape: (-1)
        name: Placeholder:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['classes'] tensor_info:
        dtype: DT_STRING
        shape: (-1, 74)
        name: head/Tile:0
    outputs['scores'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 74)
        name: head/predictions/probabilities:0
  Method name is: tensorflow/serving/classify

我的C++代码：

// run the model
std::string test(const char *modelPath, const char *text) {
    tensorflow::SavedModelBundle model;
    tensorflow::Status status = tensorflow::LoadSavedModel(
            tensorflow::SessionOptions(),
            tensorflow::RunOptions(),
            modelPath,
            {"serve"},
            &model);

    TF_CHECK_OK(status);

    // Provide input data.
    tensorflow::Tensor tensor(tensorflow::DT_STRING, tensorflow::TensorShape());
    tensor.scalar<tensorflow::tstring>()() = tensorflow::tstring(text);

    // Link the data with some tags so tensorflow know where to put those data entries.
    std::vector<std::pair<std::string, tensorflow::Tensor>> feedInputs = {{"Placeholder:0", tensor}};
    std::vector<std::string> fetches = {"head/Tile:0", "head/predictions/probabilities:0"};

    // We need to store the results somewhere.
    std::vector<tensorflow::Tensor> outputs;

    // Let's run the model...
    status = model.GetSession()->Run(feedInputs, fetches, {}, &outputs);
    TF_CHECK_OK(status);

    for (const auto& output : outputs) {
        // TODO::
    }

    return "";
}

我已尝试修改C代码中的输入输出名称，但似乎当前使用的名称是正确的。我对TensorFlow C API不太熟悉，希望得到帮助。

错误原因 #

输入张量维度不匹配：从saved_model_cli的输出可知，输入Placeholder:0的shape是(-1)（一维批量输入），但代码中创建的是0维标量张量（TensorShape()为空），模型期望的是至少一维的张量（哪怕是单样本，shape也应为[1]），导致维度越界错误。

修改步骤 #

1. 调整输入张量的形状为一维，匹配模型的批量输入要求
2. 优先使用SignatureDef定义的输入输出名称（而非直接用张量名），提升代码健壮性
3. 如果需要直接获取类别ID，推荐使用predict签名，它直接输出class_ids和probabilities，无需额外处理

修改后的代码 #

// run the model
std::string test(const char *modelPath, const char *text) {
    tensorflow::SavedModelBundle model;
    tensorflow::Status status = tensorflow::LoadSavedModel(
            tensorflow::SessionOptions(),
            tensorflow::RunOptions(),
            modelPath,
            {"serve"},
            &model);

    TF_CHECK_OK(status);

    // 1. 创建一维张量，shape为[1]，对应单样本批量输入
    tensorflow::Tensor tensor(tensorflow::DT_STRING, tensorflow::TensorShape({1}));
    tensor.flat<tensorflow::tstring>()(0) = tensorflow::tstring(text);

    // 2. 使用predict签名的输入名称（更健壮，避免张量名变化）
    std::vector<std::pair<std::string, tensorflow::Tensor>> feedInputs = {{"content", tensor}};
    // 3. 获取predict签名的输出：类别ID和概率分数
    std::vector<std::string> fetches = {"class_ids", "probabilities"};

    std::vector<tensorflow::Tensor> outputs;
    status = model.GetSession()->Run(feedInputs, fetches, {}, &outputs);
    TF_CHECK_OK(status);

    // 解析输出结果
    if (outputs.size() >= 2) {
        // 类别ID：shape (-1,1)，取第一个样本的ID
        int64_t class_id = outputs[0].flat<tensorflow::int64>()(0);
        // 概率分数：shape (-1,74)，取对应类别的分数
        float score = outputs[1].flat<float>()(class_id);
        
        // 示例输出
        std::cout << "类别ID: " << class_id << ", 分数: " << score << std::endl;
    }

    return "";
}

补充说明 #

• 如果坚持使用serving_default或classification签名，只需将输入名称改为inputs，输入张量形状保持[1]即可
• 若要处理多样本输入，只需调整张量形状为[N]（N为样本数量），并填充对应文本即可

内容的提问来源于stack exchange，提问作者Sina KH