Initial tensors(Initializer in Graph) reading by C++ are empty

Ask a Question

Question

When I use C + + to parse the *.onnx file by compiling the onnx.proto file in the repository, I can’t read the initial value though this value can be read in netron.app.

The code showing this problem is provided later. I use graph.initialzier() to read the initial tensors and print their names, types and some values. For example, in the model addition lstm, there is a none empty value W but I can’t read it(size() = 0).

#include <fstream>
#include <cassert>
#include <vector>
#include <cstdio>
#include <iostream>
#include <cstdlib>
#include <cstring>

#include "onnx.pb.h"

std::string onnxPath = "../model/addition-lstm/addition_lstm.onnx";

int main(int argc, char** argv)
{
    std::ifstream readStream(onnxPath, std::ios::ate | std::ios::in | std::ios::binary); // open file and move current position in file to the end
    if (readStream.is_open()) {
        printf("open successfully!\n");
    }
    std::streamsize size = readStream.tellg(); // get current position in file
    readStream.seekg(0, std::ios::beg);
    std::vector<char> buffer(size);
    readStream.read(buffer.data(), size);
    onnx::ModelProto model;
    model.ParseFromArray(buffer.data(), size);

    auto graph = model.graph();
    auto nodes = graph.node();

    std::cout << graph.input().size() << std::endl;

    std::cout << "initializer" << std::endl;
    for (auto init : graph.initializer()) {
        std::cout << init.name()
            << ", type: " << init.data_type()
            << ", INT32_SIZE:" << init.int32_data().size()
            << ", INT64_SIZE:" << init.int64_data().size()
            << ", FLOAT32_SIZE:" << init.float_data().size();

        std::cout << std::endl;
    }

    return 0;
}

The model can be downloaded from [addition_lstm_onnx](https://drive.google.com/file/d/1skYS4NRNHSJklS50jN5kjCWl3WlrDhiT/view?usp=sharing). In fact, The same problem occurs in all the models I have tried.

[SeHuibixiaoxing]

@jcwchen Thank you! Your suggestion succeeded in solving my problem. Wish you all the best！

Hi, would you please share the code snippet for reading initializer values?

Sure. It looks like:

auto initializer = graph->initializer();
for (auto tmp : initializer)
{
		if (tmp.data_type() == onnx::TensorProto::DataType::TensorProto_DataType_INT32) {
			auto intVec = ParseData<int32_t>(&tmp);
		}
		else if (tmp.data_type() == onnx::TensorProto::DataType::TensorProto_DataType_INT64) {
		
			auto llVec = ParseData<int64_t>(&tmp);
		}
               ... ...
}

The function ParseData is the one mentioned above.

[SeHuibixiaoxing]

@jcwchen Thank you! Your suggestion succeeded in solving my problem. Wish you all the best！