Is there any implementation of yolov8 in cpp using tensorflow lite?

[Thunderzen]

Search before asking

• I have searched the YOLOv8 issues and discussions and found no similar questions.

Question

I have the .tflite and the relevant libraries in cpp but unsure on how to perform inference.

Additional

No response

[glenn-jocher]

@Thunderzen hello! Currently, we don't provide a direct C++ implementation for YOLOv8 using TensorFlow Lite. However, you can perform inference with your .tflite model in C++ by using the TensorFlow Lite C++ API. Here's a basic outline of the steps you'd typically follow:

1. Load the TensorFlow Lite model:

   std::unique_ptr<tflite::FlatBufferModel> model = tflite::FlatBufferModel::BuildFromFile("your_model.tflite");

2. Build the interpreter:

   tflite::ops::builtin::BuiltinOpResolver resolver;
   std::unique_ptr<tflite::Interpreter> interpreter;
   tflite::InterpreterBuilder(*model, resolver)(&interpreter);

3. Allocate tensors and perform inference:

   interpreter->AllocateTensors();
   // Set input data
   float* input = interpreter->typed_input_tensor<float>(0);
   // Fill 'input' with your input data
   interpreter->Invoke();
   // Get output data
   float* output = interpreter->typed_output_tensor<float>(0);

For a complete example and more details, you might want to check out the TensorFlow Lite C++ API documentation. Hope this helps!

[Thunderzen]

@glenn-jocher ,

Hi, is there documentation on the post-processing after performing interpreter->invoke in c++? For example, code to extract the information like bounding boxes, confidences and class scores?

[glenn-jocher]

Hello! Currently, we don't have specific documentation for post-processing YOLOv8 outputs in C++ after using TensorFlow Lite's interpreter->Invoke(). However, typically, you'll need to access the output tensor, which contains the detection results, and then apply the appropriate logic to extract bounding boxes, confidence scores, and class IDs.

Here's a brief example of how you might start accessing the output tensor:

float* output = interpreter->typed_output_tensor<float>(0);
// Output processing code here

The exact details depend on the output format of your model. You might need to reshape the tensor and apply non-max suppression to filter overlapping boxes. For a more detailed guide, you might find TensorFlow's C++ API documentation helpful, or consider exploring community forums for specific examples related to YOLOv8. Hope this helps!