How use " Ultra-Light-Fast-Generic-Face-Detector-1M" on RKNN?

[VistanaGit]

I need to use Ultra-lightweight face detection model for a face recognition project on a dev board. For this application, it requires to convert the model to rknn model and run the converted model on rknn. If there is any solution , could you please help me to use Ultra-lightweight face detection mode on RKNN?, Thanks

[KorolLich]

@VistanaGit Hi! You can try something like this:

import numpy as np
from rknn.api import RKNN

MODEL_PATH = "models/onnx/version-RFB-320_new.onnx"
MODEL_NAME = MODEL_PATH.split("/")[-1]

# Initialize RKNN object
rknn: RKNN = RKNN(verbose=False)

if rknn.config(target_platform='rk3588'): #dynamic_input=DYNAMIC_INPUT
    print("RKNN config failed")
    exit(1)

# Load ONNX model
# input_size_list = [[1, 3, 240, 320]]
ret = rknn.load_onnx(model=MODEL_PATH) #, inputs=['modelInput'], input_size_list=input_size_list

# Check for successful model loading
if ret != 0:
    print("Loading ONNX model failed")
    exit(1)

ret = rknn.build(do_quantization=False, dataset="./dataset.txt")

if ret != 0:
    print("Building RKNN model failed")
    exit(1)

if rknn.init_runtime():
    print("RKNN initialization failed")
    exit(1)

rknn_model_path = MODEL_PATH.replace(".onnx", ".rknn")

ret = rknn.export_rknn(rknn_model_path)
if ret != 0:
    print("Export failed")
    rknn.release()
    exit(ret)

input_data = np.random.rand(1, 240, 320, 3).astype(np.float32)
outputs = rknn.inference(inputs=[input_data])
rknn.release()

It is working for me with rknn-toolkit2==2.2.0. Make sure that you have the same versions for rknn-toolkit2, rknn-toolkit-lite and RKNPU driver on your board