Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Problems getting custom model to run on edge_tpu #104

Open
Perugius opened this issue Mar 21, 2024 · 3 comments
Open

Problems getting custom model to run on edge_tpu #104

Perugius opened this issue Mar 21, 2024 · 3 comments

Comments

@Perugius
Copy link

Perugius commented Mar 21, 2024

Description

I am trying to get my simple model to run on the edge tpu.
Graph of model
image

Graph of model compiled for edge-tpu
image

I have followed the edge-tpu runtime guide, but when I flash the script it doesn't work, there is no output similar to when the OP resolver doesn't correctly add all the ops. I also included the correct path to the edge-tpu model in the CMake file.

Is there a problem with my code, is it the model itself, or am I just encountering a bug?

#include <cstdio>
#include <chrono>
#include "libs/tensorflow/utils.h"
#include "examples/tflm_hello_world/hello_world_model.h"
#include "libs/base/led.h"
#include "libs/base/timer.h"
#include <climits>
#include <ctime>
#include <stdint.h>
#include "libs/tpu/edgetpu_manager.h"
#include "libs/tpu/edgetpu_op.h"
#include "third_party/freertos_kernel/include/FreeRTOS.h"
#include "third_party/freertos_kernel/include/task.h"
#include "third_party/tflite-micro/tensorflow/lite/micro/all_ops_resolver.h"
#include "third_party/tflite-micro/tensorflow/lite/micro/micro_mutable_op_resolver.h"
#include "third_party/tflite-micro/tensorflow/lite/micro/micro_error_reporter.h"
#include "third_party/tflite-micro/tensorflow/lite/micro/micro_interpreter.h"
#include "libs/base/filesystem.h"
// Runs a tiny TFLM model on the M7 core, NOT on the Edge TPU, which simply
// outputs sine wave values over time in the serial console.
//
// For more information about this model, see:
// https://github.com/tensorflow/tflite-micro/tree/main/tensorflow/lite/micro/examples
//
// To build and flash from coralmicro root:
//    bash build.sh
//    python3 scripts/flashtool.py -e tflm_hello_world

const int test_data[] = { 
4,4,9,3,6,8,0,8,7,4,8,2,0,6,6,7,1,2,6,1,9,0,6,8,6,8,8,7,1,3,5,8,9,9,8,6,4,6,1,7,3,2,6,5,5,8,7,6,6,7,3,9,2,5,2,1,5,1,3,8,7,6,9,6,2,2,7,4,0,9,2,4,5,8,7,5,8,7,0,8,8,3,0,1,3,0,2,7,9,6,5,4,3,9,2,0,1,8,0,3,5,2,5,9,4,4,7,4,2,6,7,5,7,3,7,6,8,9,2,7,1,3,9,7,1,0,3,8,0,5,3,2,9,2,3,2,1,1,2,1,0,1,8,7,4,0,4,9,1,5,4,3,8,0,7,5,3,2,0,2,3,1,6,2,5,7,0,9,3,3,2,2,8,6,5,6,2,6,1,2,4,7,8,8,3,3,8,9,0,7,5,6,9,1,5,0,6,1,3,0,8,3,9,2,9,9,7,7,7,1,9,1,5,3,8,0,5,1,9,2,0,0,6,9,4,2,9,2,7,2,1,1,3,4,0,5,8,8,7,3,
};

// const float test_data[] = { 
// 0.22768896476796524,0.2047959025298774,0.21628123544932484,0.2037094521185783,0.2047959025298774,0.193388173211237,0.17049511097314915,0.18190284029178955,0.1590097780537017,0.1590097780537017,0.1590097780537017,0.16940866056185008,0.13619431941642093,0.09040819494024523,0.0,0.3295048890268508,0.5228930622380878,0.7723110352320348,0.6476020487350613,0.5114077293186404,0.39702002172900824,0.4199130839670961,0.3741269594909204,0.3741269594909204,0.35239795126493867,0.2952040974701226,0.28379636815148224,0.2494179729939469,0.2151171814372187,0.21628123544932484,0.2390966940866056,0.26198975632469346,0.28379636815148224,0.32841843861555176,0.3409902219462983,0.4210771379792022,0.5,0.5904081949402452,0.7047959025298773,0.8054477727766568,0.9313208132857365,1.0,0.8741269594909203,0.7379326400744994,0.7150395778364116,0.6578457240415956,0.3752910135030265,0.12470898649697346,0.0,0.011485332919447462,0.03321434114542915,0.06751513270215738,0.07900046562160484,0.06751513270215738,0.09149464535154431,0.06867918671426354,0.056107403383516996,0.056107403383516996,0.07900046562160484,0.10189352785969269,0.8333965844402277,0.8333965844402277,0.8473434535104364,0.8195445920303606,0.8333965844402277,0.8329222011385199,0.8333965844402277,0.8329222011385199,0.8333965844402277,0.8329222011385199,0.8467741935483871,0.8467741935483871,0.8752371916508539,0.8752371916508539,0.9443074003795067,0.9860531309297913,1.0,0.7637571157495257,0.27751423149905124,0.04127134724857685,0.0,0.19392789373814043,0.22229601518026565,0.15265654648956356,0.12485768500948767,0.11091081593927894,0.11091081593927894,0.09743833017077799,0.08358633776091082,0.11091081593927894,0.13870967741935483,0.1666034155597723,0.18055028462998102,0.1944022770398482,0.18055028462998102,0.22229601518026565,0.3057874762808349,0.4306451612903226,0.47239089184060723,0.444022770398482,0.6663187855787476,0.8195445920303606,0.9164136622390892,0.9860531309297913,0.9994307400379506,0.9722011385199241,0.930360531309298,0.902561669829222,0.8891840607210626,0.9031309297912713,0.902561669829222,0.8752371916508539,0.8886148007590133,0.8746679316888045,0.8886148007590133,0.8886148007590133,0.8746679316888045,0.8607210626185958,0.8473434535104364,0.8467741935483871,0.7136757350250091,0.695498353055996,0.7157496645114066,0.7067219714529706,0.9636452360619739,0.9818226180309869,0.9839575454434549,0.9657801634744418,0.9546175430035378,0.7954129559594973,0.7339270464804196,0.7409418079785287,0.7045870440405027,0.679455898499451,0.7954129559594973,0.34982310601439554,0.18854458948395755,0.24094180797852874,0.14316213248749543,0.0,0.15225082347200194,0.2318531169940222,0.22276442600951568,0.4365011589605953,0.5453824569964622,0.5272050750274491,0.5544711479809686,0.4637062339880444,0.4476027815054288,0.4385140905209223,0.3931316335244602,0.4022203245089667,0.42039770647797975,0.4294863974624863,0.4385140905209223,0.4476027815054288,0.5705136025375137,0.6543247529583994,0.6976332804684641,0.4476027815054288,0.7681468830059778,0.8861778699524216,0.9909113090154935,0.9364401610345249,0.8589727949249726,0.9022203245089667,1.0,0.9113090154934732,0.7863242649749909,0.6976332804684641,0.6613395144565085,0.6885445894839576,0.6613395144565085,0.7136757350250091,0.8044406490179334,0.7772355739904844,0.6976332804684641,0.679455898499451,0.6976332804684641,0.6976332804684641,0.03571894784935709,0.0494093074864565,0.05774753606161478,0.05774753606161478,0.12463285686312904,0.1337543241302787,0.1497454474251028,0.1406076626852033,0.13146987794530382,0.08057568043861367,0.07600678806866393,0.06608576463677306,0.015175249657333072,0.0,0.10488871483584622,0.17175771816461066,0.29492200248025585,0.40813262841851056,0.4841557339599243,0.5373180601788395,0.5570458847333725,0.5844755564258207,0.5927648325827296,0.5904803863977547,0.5707525618432218,0.5973337249526793,0.5904803863977547,0.5098883884863912,0.5107368970693819,0.5410384439657986,0.5844755564258207,0.6527804973565694,0.6962339272893414,0.7333888127406827,0.750815873637491,0.780546308987664,0.8002578160694471,0.8337086352065792,0.7728118269042491,0.7045232034462502,0.8314078715488545,0.9303407088310163,1.0,0.943182559885125,0.7393773252398669,0.5510410547614385,0.3069806148423732,0.29182168265778996,0.28349977155538153,0.274361986815482,0.20063964493179295,0.16720514326741073,0.14436068141766203,0.15431433979505255,0.19607075256184323,0.1983551987468181,0.16035180471248614,0.14892957378761176,0.14207623523268717,0.14517655505515306,
// };

namespace coralmicro {
namespace{

uint64_t startTime;
uint64_t endTime;
uint64_t inferenceTime;

int inference_count = 0;
const int kInferencesPerCycle = 1000;

// An area of memory to use for input, output, and intermediate arrays.
constexpr int kTensorArenaSize = 8 * 1024 * 1024;
STATIC_TENSOR_ARENA_IN_SDRAM(tensor_arena, kTensorArenaSize);


[[noreturn]] void Main() {

  coralmicro::TimerInit();
  LedSet(Led::kStatus, true);

  //-----------------------------TURN ON EDGETPU---------------------------
  auto tpu_context = EdgeTpuManager::GetSingleton()->OpenDevice();
  if (!tpu_context) {
    printf("ERROR: Failed to get EdgeTpu context\r\n");
  }
  //-----------------------------END TURN ON EDGETPU---------------------------

  //---------------------------------LOAD MODEL-----------------------------
  constexpr char kModelPath[] = "/models/model_V4_QUANTIZED_edgetpu.tflite";

  std::vector<uint8_t> model;
  if (!LfsReadFile(kModelPath, &model)) {
    printf("ERROR: Failed to load %s\r\n", kModelPath);
    vTaskSuspend(nullptr);
  }
  //---------------------------END LOAD MODEL--------------------------------

  //---------------------------OP RESOLVER---------------------------------
  tflite::MicroErrorReporter error_reporter;
  tflite::MicroMutableOpResolver<10> resolver;
  resolver.AddQuantize();
  resolver.AddDequantize();
  resolver.AddDepthwiseConv2D();
  resolver.AddConv2D();
  resolver.AddExpandDims();
  resolver.AddReshape();
  resolver.AddMean();
  resolver.AddSoftmax();
  resolver.AddFullyConnected();
  resolver.AddLogistic();
  resolver.AddCustom(kCustomOp, RegisterCustomOp());
  //-------------------------END OP RESOLVER--------------------------------

  //-------------------------INTERPRETER INIT------------------------------

  tflite::MicroInterpreter interpreter(tflite::GetModel(model.data()),
                                     resolver, tensor_arena,
                                     kTensorArenaSize, &error_reporter);

  if (interpreter.AllocateTensors() != kTfLiteOk) {
    printf("ERROR: AllocateTensors() failed\r\n");
    vTaskSuspend(nullptr);
  }

  auto* input_tensor = interpreter.input_tensor(0);
  auto* output_tensor = interpreter.output_tensor(0);
  //----------------------END INTERPRETER INIT-----------------------------

  //------------------------START INFERENCE LOOP---------------------------
  while (true){

    vTaskDelay(pdMS_TO_TICKS(1000));
    //std::memcpy(tflite::GetTensorData<uint8_t>(input_tensor), test_data, sizeof(test_data));
    for (int i = 0; i < 240; i++) {
      input_tensor->data.uint8[i] = test_data[i];
      printf("\r\n progress: %d", i);
      //printf("\r\n %d", input_tensor->data.uint8[i]);
    }

    //-------------------------------INVOKE INTERPRETER-------------------------------
    startTime = coralmicro::TimerMicros();
    if (interpreter.Invoke() != kTfLiteOk) {
      printf("ERROR: Invoke() failed\r\n");
    }
    endTime = coralmicro::TimerMicros();
    inferenceTime = endTime-startTime;
    printf("inference time: %u \r\n ", (unsigned int)(inferenceTime));

    //-------------------------------END INVOKE INTERPRETER-------------------------------


    printf("inference time: %u \r\n ", (unsigned int)(inferenceTime));
    vTaskDelay(pdMS_TO_TICKS(10000));
  }
  //---------------------------END INFERENCE LOOP-------------------------------------
}
} //namespace 
} //namespace coralmicro

extern "C" void app_main(void* param) {
  (void)param;
  coralmicro::Main();
}

Click to expand!

Issue Type

Support

Operating System

Linux, Ubuntu

Coral Device

Dev Board Micro

Other Devices

No response

Programming Language

C++

Relevant Log Output

No response

@Perugius
Copy link
Author

I suspect now that the problem is the input shape of the model, since I trained it on (4, 60), but the input array I am using is flat, which normally wouldn't be a problem, but it might cause problems on the edge-tpu. So I will try again with a new model trained with a flat input array.

@alexskip
Copy link

Hi @Perugius did you get your custom model running? My board runs pretrained models fine, but I am experiencing similar issues with absolutely trivial architecture (1 input, 1 dense) - no errors, the board just hangs on invoke.

@Perugius
Copy link
Author

Hey, I never got any type of custom model to run on the edge tpu sadly. I tried contacting the coral team multiple ways but I think at this point the whole project is deprecated

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

No branches or pull requests

2 participants