General Information

The following example shows how to evaluate AI Models for RZ/V2H. The created inference will be able to run on the DRP-AI (INT8). The inference output may not be accurate because the training needs to be done to further calibrate the AI Model.

For this demo we are using the Darknet YoloV2 VOC.

Supported Devices (INT8 DRP-AI IP):

• RZV2H

Requirements:

• RZ/V2H EVK
• RZ/V2H AI_SDK
• TVM Translator ( Follow RZV2H Installation ) here
• Install TVM Translator here ( Follow RZ/V2H Docker Installation )

TVM Translation

Step 1) Start Docker Container

mkdir data
docker run -it --name drp-ai_tvm_v2h_container_${USER} -v $(pwd)/data:/drp-ai_tvm/data drp-ai_tvm_v2h_image_${USER}

Step 2) Preparation

add the following to access the TVM Compile Scripts.

Code Block
# Added the following paths to use the TVM Scripts PYTHONPATH=/drp-ai_tvm/tvm/python:/drp-ai_tvm/tutorials/ # Create Symbolic links for the following bash scripts # These are requried to run TVM translator ln -s /drp-ai_tvm/tutorials/run_*

Step 3) Download Darknet yolov2 voc inference

wget https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/yolov2-voc.cfg
wget https://pjreddie.com/media/files/yolov2-voc.weights

Step 4) Get the yolo.ini

This configuration file to run the provided conversion scripts

cp /drp-ai_tvm/how-to/sample_app/docs/object_detection/yolo/yolo.ini .

Step 5) Convert Darknet Inference model to Pytorch file

python3 ../scripts/convert_to_pytorch.py yolov2

Step 6) Convert

Pytorch file to

ONNX file

python3 ../scripts/convert_to_onnx.py yolov2

Step 7) Make a copy of the TVM Script compile_onnx_model.

cp /drp-ai_tvm/tutorials/compile_onnx_model_quant.py /drp-ai_tvm/tutorials/compile_onnx_model_yolov2.py

Step 8) Modify the python script

The script compile_onnx_model_yolov2.py  pre-processing section of the

script below needs need to be changed accordingly

config.shape_in     = [1, 480, 640, 3]
mean    = [0.485, 0.456, 0.406]
stdev   = [0.229, 0.224, 0.225]

config.shape_in     = [1, 480, 640, 3]
mean    = [0.0, 0.0, 0.0]
stdev   = [1.0, 1.0, 1.0]

# Configure TVM DRP-AI Preprocessor Camera/Image input shape
sed -i 's/^.*config.shape_in.*$/ config.shape_in = [1, 480, 640, 3]/' /drp-ai_tvm/tutorials/compile_onnx_model_yolov2.py

# Change the default TVM DRP-AI Preprocessor mean and stdev to YoloV2
sed -i 's/mean\s*=\s*\[[0-9., ]*\]/mean = [0.0, 0.0, 0.0]/' /drp-ai_tvm/tutorials/compile_onnx_model_yolov2.py
sed -i 's/stdev\s*=\s*\[[0-9., ]*\]/stdev = [1.0, 1.0, 1.0]/' /drp-ai_tvm/tutorials/compile_onnx_model_yolov2.py

Step 9) Translate

Using the modified script created in step 7 and 8 translate the onnx model to DRP-AI TVM Model.

python3 compile_onnx_model_yolov2.py \
./d-yolov2.onnx \
-t $SDK \
-d $TRANSLATOR \
-c $QUANTIZER \
-i input1 \
-s 1,3,416,416 \
-o yolov2_onnx \
-v 100

Minimum Required arguments:

• ./d-yolov2.onnx - This is the ONNX input file
• -t is path to Yocto SDK
• -d path to the DRP AI Translator Tool
• -c Path to DRP-AI Quantization Tool 
• -s Define the Inference input shape
• -o Output directory
• -v Specifies the version of TVM