conv.py

from rknn.api import RKNN
import os

# for suffix in ["n", "s", "m", "l", "x"]:
for suffix in ["n", "s", "m", "l", "x"]:
    # for soc in ["rk3562","rk3566", "rk3568", "rk3588"]:
    for soc in ["rk3588"]:
        # for QUANTIZATION in [True, False]:
        for QUANTIZATION in [True]:
            INPUT_MODEL = 'yolov8{}.onnx'.format(suffix)
            WIDTH = 320
            HEIGHT = 320
            OUTPUT_MODEL_BASENAME = 'yolov8{}'.format(suffix)
            # QUANTIZATION = False
            DATASET = './datasets/coco20/dataset_coco20.txt'

            # Config
            MEAN_VALUES = [[0, 0, 0]]
            STD_VALUES = [[255, 255, 255]]
            QUANT_IMG_RGB2BGR = True
            QUANTIZED_DTYPE = "asymmetric_quantized-8"
            QUANTIZED_ALGORITHM = "normal"
            QUANTIZED_METHOD = "channel"
            FLOAT_DTYPE = "float16"
            OPTIMIZATION_LEVEL = 2
            TARGET_PLATFORM = soc
            CUSTOM_STRING = None
            REMOVE_WEIGHT = None
            COMPRESS_WEIGHT = False
            SINGLE_CORE_MODE = False
            MODEL_PRUNNING = False
            OP_TARGET = None
            DYNAMIC_INPUT = None

            if QUANTIZATION:
                quant_suff = "-i8"
            else:
                quant_suff = ""
            
            OUTPUT_MODEL_FILE = "./output/{}/{}-{}x{}{}-{}.rknn".format(soc, OUTPUT_MODEL_BASENAME, WIDTH, HEIGHT, quant_suff, soc)
            os.makedirs("./output/{}".format(soc), exist_ok=True)

            rknn = RKNN()
            rknn.config(mean_values=MEAN_VALUES,
                        std_values=STD_VALUES,
                        quant_img_RGB2BGR=QUANT_IMG_RGB2BGR,
                        quantized_dtype=QUANTIZED_DTYPE,
                        quantized_algorithm=QUANTIZED_ALGORITHM,
                        quantized_method=QUANTIZED_METHOD,
                        float_dtype=FLOAT_DTYPE,
                        optimization_level=OPTIMIZATION_LEVEL,
                        target_platform=TARGET_PLATFORM,
                        custom_string=CUSTOM_STRING,
                        remove_weight=REMOVE_WEIGHT,
                        compress_weight=COMPRESS_WEIGHT,
                        single_core_mode=SINGLE_CORE_MODE,
                        model_pruning=MODEL_PRUNNING,
                        op_target=OP_TARGET,
                        dynamic_input=DYNAMIC_INPUT)

            # if rknn.load_pytorch("./input/" + INPUT_MODEL, [[HEIGHT, WIDTH, 3]]) != 0:
            if rknn.load_onnx("./input/" + INPUT_MODEL) != 0:
                print('Error loading model.')
                exit()

            if rknn.build(do_quantization=QUANTIZATION, dataset=DATASET) != 0:
                print('Error building model.')
                exit()

            if rknn.export_rknn(OUTPUT_MODEL_FILE) != 0:
                print('Error exporting rknn model.')
                exit()