Model preparation for Hailo

Translating

First, the model needs to be parsed from a .pth format into onnx, run convert_pth_to_onnx.py with the pytorch model.

Parsing

Use the translated .onnx model file to parse it to a .har (Hailo archive) file. Go to /hailo

Run:


                . parse_model.sh "path_to_.onnx_model"

Optimization

To optimize the model we first need to create the calibration set so the model will know what data types to expect as input and calibrates the weights correctly when optimizing.

Go to hailo/2.create_calibration_set_Final
Run:


                # Create a directory with preprocessed faces resized to 224x224 as the model expects
                

                python3 preprocess_images.py
                
 

                # create the calibration set as a .npy array file
                

                python3 prepare_calibration_set_FLOAT32.py

Go to hailo/3.optimization_Final
Run:


                # Optimize the model
                

                . optimize_model.sh --hw-arch 'hardware architecture' --calib-set-path 'calibration set' --output-har-path 'output file name' 'input file
                

                

                # For help run
                

                . optimize_model.sh

Compilation

Go to hailo/4.compilation
Run:


                # Compile the model
                

                . compile_model.sh --hw-arch 'hardware architecture' --output-dir output 'input_file'
                

                

                # For help run
                

                . compile_model.sh

Usage

Copy the compiled .hef model to the main detector file directory and specify the path to the model in the code