tensorflow: [TFLite] TfLiteInterpreterInvoke has memory leak and increasing duration time

Issue type

Bug

Have you reproduced the bug with TensorFlow Nightly?

No

Source

source

TensorFlow version

tf 2.15.0

Custom code

Yes

OS platform and distribution

Android/iOS mobile platform

Mobile device

No response

Python version

No response

Bazel version

No response

GCC/compiler version

No response

CUDA/cuDNN version

No response

GPU model and memory

No response

Current behavior?

I found memory and duration time increasing when I use tflite C API code in below. The duration time increase 10 msec in just 10 minitues.

TfLiteStatus init(){
    // create model 
    mTFLModel = TfLiteModelCreate(dst_buf, dlc_size);

    mTFLInterpreter = TfLiteInterpreterCreate(mTFLModel, NULL);
    // Allocate tensor buffers. 
    mStatus = TfLiteInterpreterAllocateTensors(mTFLInterpreter);
    // output buffer
    mdnnOutputDataBuffer = new float[OutputSize.width * OutputSize.height];
    memset(mdnnOutputDataBuffer, 0, sizeof(float) * OutputSize.width * OutputSize.height );
    return mStatus;
}

void release(){
    if (mdnnOutputDataBuffer != NULL) delete[] mdnnOutputDataBuffer;
    if (mTFLInterpreter != NULL) {
        TfLiteInterpreterDelete(mTFLInterpreter);
        mTFLInterpreter = NULL;
    }
    if (mTFLModel != NULL) {
        TfLiteModelDelete(mTFLModel);
        mTFLModel = NULL;
    }
}

TfLiteStatus main(const float* input, uint8_t* intValues) {

        TfLiteTensor* input_tensor = TfLiteInterpreterGetInputTensor(mTFLInterpreter,0);
        //
        mStatus = TfLiteTensorCopyFromBuffer(input_tensor, input, TfLiteTensorByteSize(input_tensor));
        //
        mStatus = TfLiteInterpreterInvoke(mTFLInterpreter);

        const TfLiteTensor* output_tensor =
                TfLiteInterpreterGetOutputTensor(mTFLInterpreter, 0);
        
        TfLiteTensorCopyToBuffer(output_tensor, mdnnOutputDataBuffer,
                                 TfLiteTensorByteSize(output_tensor));

}

Now I know is … ・Not depending on tf version(I tried both 2.9.0 and 2.16.1) ・Not depenging on iOS/android(both happen) ・Low resolution input/output model slightly better ・Not depending on device(I tried sevaral devices)

I checked similar mention, but those problem are about NNapi or GPU delegate option. I don’t use delegate option and doesen’t work for me.

Standalone code to reproduce the issue

…

Relevant log output

No response

About this issue

  • Original URL
  • State: closed
  • Created 2 years ago
  • Comments: 28 (3 by maintainers)

Most upvoted comments

meet the same problem while using metal delegate with iOS

Hi @kkk-jpeg ,

You can upload your model to your google drive and share a link here or you can mail me at sawantkumar@google.com.