Quantization 8bit for yolov4

[Kartikeya]

Hi,

I am trying to convert fp32 yolo model(trained on custom classes) into an int8 low precision quantized model. However upon conversion I am unable to see any bounding boxes(unlike fp32/fp16) when I try to do inference even though .xml and .bin files are generated. I have tried Default and Accuracy aware training both. I am able to achieve the conversion through command line tools. DL Workbench does not convert to int8 IR also.
How can I tackle the above problem?

[Sebastian_M_Intel]

Hello Kartikeya, 

Thank you for posting on the Intel® communities.  

Based on what you are reporting this seems to be related to the OpenVINO™ toolkit, we will move this thread to the proper sub-forum for better assistance and kindly wait for a response.  

Best regards, 

Regards, 

Sebastian M  

Intel Customer Support Technician  

[IntelSupport]

Hi,

Thanks for reaching out.

Yolov4 is not validated with OpenVINO. Only yolov1,yolov2 and yolov3 are validated. The information is available in the following page :

https://docs.openvinotoolkit.org/latest/openvino_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_YOLO_From_Tensorflow.html#yolov3-to-ir

The following page contains the list of topologies that have been validated for 8-bit inference feature.

https://docs.openvinotoolkit.org/latest/openvino_docs_IE_DG_Int8Inference.html#low_precision_8_bit_integer_inference_workflow.

There are three Yolo models (TensorFlow Yolov3,Caffe Yolo v1 tiny and Caffe Yolov3 ) in the above list.

I suggest you use any of these models for post training quantization

Here is an additional reference regarding the workflow of converting a model from FP32 to INT8:

https://docs.openvinotoolkit.org/latest/workbench_docs_Workbench_DG_Int_8_Quantization.html

Regards,

Aznie

[IntelSupport]

Hi Kartikeya,

This thread will no longer be monitored since we have provided a solution. If you need any additional information from Intel, please submit a new question.

Regards,

Aznie