Bus Error when running model

[phil-fill]

Description

I have a project where i want to run the blaze face short range (https://github.com/google/mediapipe/blob/master/mediapipe/examples/coral/models/face-detector-quantized_edgetpu.tflite) compiled for the edge tpu on my Raspberry pi 4 with Debian bookworm plus the coral TPU Accelerator with c++. I already tried it with the pycoral api and it is runing on my edgetpu (~400 fps) So i assume the model should be fine.

I made a code similar to the code suggested when running a model with edgetpu and tensorflowlite (https://coral.ai/docs/edgetpu/tflite-cpp/#set-up-the-tf-lite-interpreter-with-libedgetpu). This is my code:

#include<stdio.h>
#include<iostream>
#include<tensorflow/lite/interpreter.h>
#include<tensorflow/lite/kernels/register.h>
#include<ctime>
#include<cstdlib>
#include<vector>
#include<tensorflow/lite/c/common.h>
#include<tensorflow/lite/model.h>
#include<memory>
#include<tensorflow/lite/tools/gen_op_registration.h>
#include<opencv2/opencv.hpp>
#include<vector>
#include<chrono>
#include<tflite/public/edgetpu.h>

using namespace edgetpu;
using namespace tflite;
int main()
        {

        //load image
        std::string image_path = "/home/jean/Screenshot.png";
        cv::Mat img = cv::imread(image_path);
        if (img.empty()){
        std::cout << "Failed to load image!";
        }

        //transform to required model input
        int new_width = 128;
        int new_height = 128;
        cv::resize(img, img, cv::Size( new_width, new_height));
        
        //load model 
        const std::string model_path = "face-detector-quantized_edgetpu.tflite";
        std::unique_ptr<tflite::FlatBufferModel> model =
                tflite::FlatBufferModel::BuildFromFile(model_path.c_str());
        std::shared_ptr<edgetpu::EdgeTpuContext> edgetpu_context =
                edgetpu::EdgeTpuManager::GetSingleton()->OpenDevice();
         edgetpu::EdgeTpuContext* edgetpu_context_ptr = edgetpu_context.get();
        tflite::ops::builtin::BuiltinOpResolver resolver;
        resolver.AddCustom(edgetpu::kCustomOp, edgetpu::RegisterCustomOp());
        std::unique_ptr<tflite::Interpreter> interpreter = std::make_unique<tflite::Interpreter>();
        tflite::InterpreterBuilder builder(*model, resolver);
        std::cout << "i was here 1" << std::endl;
        interpreter->SetExternalContext(kTfLiteEdgeTpuContext, reinterpret_cast<TfLiteExternalContext*>(edgetpu_context_ptr));
        std::cout << "i was here2" << std::endl;
        interpreter->SetNumThreads(1);
        std::cout << "i was here3" << std::endl;
        interpreter->AllocateTensors();
        std::cout << "i was here4" << std::endl;
        int* inputImg_ptr = img.ptr<int>(0);
        
        //get output
        TfLiteTensor* input_tensor = interpreter->tensor(interpreter->inputs()[0]);
        TfLiteTensor* output_box = interpreter->tensor(interpreter->outputs()[0]);
        TfLiteTensor* output_score = interpreter->tensor(interpreter->outputs()[1]);
        memcpy(input_tensor->data.f, img.ptr<int>(0),128 * 128 * 3 * sizeof(int));
        interpreter->Invoke();
        }

this is the output

i was here 1
i was here2
Bus error

So as soon i am calling the set threads, there is going something wrong. A problem could be that i crosscompiled the libtensorflowlite.so the newest TF (2.15) with Bazel fro arm (https://www.tensorflow.org/lite/guide/build_arm) and for the libedgetpu i crosscompiled it for aarch64 with Docker and Bazel (https://github.com/google-coral/libedgetpu). The problem is that there is a default version for TF for the libedgetpu in the workspace.btl file (TF 2.5). I dont know if that is the problem but i tried first to crosscompile libedgetpu.so with the newest TF by adjusting the commit and sha256 in the workspace.bzl file and i tried to crosscompile libtensorflowlite.so for a TF 2.5. Both did not work. Do you have any hints for me?

Best wishes

Click to expand!

Issue Type

Build/Install

Operating System

Ubuntu

Coral Device

USB Accelerator

Other Devices

Rapsberry Pi 4

Programming Language

C++

Relevant Log Output

No response