convert to onnx, the inference time cost increase

[VincentMy]

`import onnx
import onnxruntime as ort
import numpy as np
import cv2
import time

model_path = 'weights/end2end.onnx'

onnx_model = onnx.load(model_path)
onnx.checker.check_model(onnx_model)

image = cv2.imread('images/demo.jpg').astype('float32')
image = cv2.resize(image, (640, 640), interpolation=cv2.INTER_AREA).reshape(3, 640, 640)
image = np.expand_dims(image, axis=0)

t = time.time()
ort_sess = ort.InferenceSession(model_path, providers=['CPUExecutionProvider'])
tcost = time.time() - t
print(f"tcost = {tcost}")

outputs = ort_sess.run(None, {'input': image})
print('batch size: {}'.format(len(outputs[0])))
print('Num of detection: {}'.format(len(outputs[0][0])))

for bbox, label in zip(outputs[0][0], outputs[1][0]):
print('label: {}, bbox: {}'.format(label, bbox))`

i conver yolox to onnx with detection_onnxruntime_dynamic.py, but when inference image with onnxruntime, no speed increase.
Besides ort.InferenceSession(model_path, providers=['CUDAExecutionProvider']) cost 2.4s, ort.InferenceSession(model_path, providers=['CPUExecutionProvider']) cost0.18s instead. could anyone can explain it?