0

我正在尝试将 python 代码重写为mnist_clientc++。由于我是 tensorflow 和 TF 服务的新手，我遇到了一些困难。我浏览了教程和 C++ 客户端示例 ( inception_client)。Pythonmnist_client工作没有任何问题，但是当我运行我的 c++ 客户端时，它给了我arg[0] is not a matrix 

gRPC call return code: 3: In[0] is not a matrix 
 [[Node: MatMul = MatMul[T=DT_FLOAT, _output_shapes=[[?,10]], transpose_a=false, transpose_b=false, _device="/job:localhost/replica:0/task:0/cpu:0"](_arg_x_0_0, Variable/read)]]

我按照教程中的方法训练了模型，并检查了我读取的 minst 数据是否正常。

由此： tensorflow Invalid argument: In[0] is not a matrix，我知道MatMul至少需要 2-dim 数据。但是，我浏览了inception_client和 python的 c++ 代码，mnist_client并且都将图像数据读入了 1-dim char 数组......我在这里错过了什么？

代码inception_client：https ://github.com/tensorflow/serving/blob/master/tensorflow_serving/example/inception_client.cc

任何帮助将非常感激。:)

class ServingClient{
public:
ServingClient(std::shared_ptr<Channel> channel) : stub_(PredictionService::NewStub(channel)){}

tensorflow::string callPredict( const tensorflow::string &model_name,
                                const tensorflow::string &model_signature,
                                const int num_tests){
PredictRequest request;
PredictResponse response;
ClientContext context;
int image_size;
int image_offset = 16;
int label_offset = 8;

request.mutable_model_spec()->set_name(model_name);
request.mutable_model_spec()->set_signature_name(model_signature);

google::protobuf::Map<tensorflow::string, tensorflow::TensorProto> &inputs = *request.mutable_inputs();

std::fstream imageFile("t10k-images-idx3-ubyte", std::ios::binary | std::ios::in);
std::fstream labelFile("t10k-labels-idx1-ubyte", std::ios::binary | std::ios::in);

labelFile.seekp(0);
imageFile.seekp(0);

uint32_t magic_number_images;
uint32_t nImages;
uint32_t magic_number_labels;
uint32_t rowsI =0;
uint32_t rowsL =0;
uint32_t colsI = 0;
uint32_t colsL = 0;


imageFile.read((char *)&magic_number_images, sizeof(magic_number_images));
imageFile.read((char *)&nImages, sizeof(nImages));
imageFile.read((char *)(&rowsI), sizeof(rowsI));
imageFile.read((char *)&colsI, sizeof(colsI));

image_size = ReverseInt(rowsI) * ReverseInt(colsI);

labelFile.read((char *)&magic_number_labels, sizeof(magic_number_labels));
labelFile.read((char *)&rowsL, sizeof(rowsL));

for(int i=0; i<num_tests; i++){
    tensorflow::TensorProto proto;

    labelFile.seekp(label_offset);
    imageFile.seekp(image_offset);

    //read mnist image
    char *img = new char[image_size]();
    char label = 0;
    imageFile.read((char *)img, image_size);

    image_offset += image_size;
    //read label
    labelFile.read(&label, 1);
    label_offset++;

    //predict
    proto.set_dtype(tensorflow::DataType::DT_STRING);
    proto.add_string_val(img, image_size);
    proto.mutable_tensor_shape()->add_dim()->set_size(1);
    inputs["images"] = proto;

    Status status = stub_->Predict(&context, request, &response);
    delete[] img;

    if(status.ok()){
    std::cout << "status OK." << std::endl;
    OutMap &map_outputs = *response.mutable_outputs();
    OutMap::iterator iter;
    int output_index = 0;

    for(iter = map_outputs.begin(); iter != map_outputs.end(); ++iter){
        tensorflow::TensorProto &result_tensor_proto = iter->second;
        tensorflow::Tensor tensor;
        //check if response converted succesfully 
        bool converted = tensor.FromProto(result_tensor_proto);
        if (converted) {
            std::cout << "the result tensor[" << output_index << "] is:" << std::endl
                        << tensor.SummarizeValue(10) << std::endl;
         } 
         else {
            std::cout << "the result tensor[" << output_index
                        << "] convert failed." << std::endl;
        }
        ++output_index;
                }
        }
    else{
        std::cout << "gRPC call return code: " << status.error_code() << ": "
            << status.error_message() << std::endl;
            }
        }
imageFile.close();
labelFile.close();
}

private:
    std::unique_ptr<PredictionService::Stub> stub_;

};

编辑 1：我认为问题一定出在模型的创建方式以及客户端发送的数据的维度上。我使用了提供的 python 程序来训练和导出设置尺寸的模型：

feature_configs = {'x': tf.FixedLenFeature(shape=[784], dtype=tf.float32),}
tf_example = tf.parse_example(serialized_tf_example, feature_configs)
x = tf.identity(tf_example['x'], name='x')  # use tf.identity() to assign name
y_ = tf.placeholder('float', shape=[None, 10])
w = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10]))
4

1 回答 1

0

正如预期的那样，修复是显而易见的。所要做的就是添加另一个维度：

   proto.mutable_tensor_shape()->add_dim()->set_size(image_size);

得到[image_size,1]形状。

于 2017-08-03T08:28:38.400 回答