TDA4VL-Q1: DLR Model Infer Slowly

After the ONNX model converted to DLR model by using "examples/osrt_python/tvm_dlr/tvm_compilation_onnx_example.py", I infer it on TDA4VL slowly by c++ code, about 10000ms per frame, and the time of inferring by python DLR API same as above.

The key point is that inferring the ONNX model by python onnxruntime directly will be so fast, about 12ms per frame.

ONNX model download url: http://software-dl.ti.com/jacinto7/esd/modelzoo/09_02_00/models/vision/detection/coco/edgeai-yolox/yolox-s-ti-lite_39p1_57p9.onnx

PS: I deleted NMS part after downloaded it !

Python DLR code :

import dlr
import numpy as np

# Load model.
# /path/to/model is a directory containing the compiled model artifacts (.so, .params, .json)
model = dlr.DLRModel('/opt/yolox-s-ti-lite_39p1_57p9_nonms.onnx', 'cpu', 0)

# Prepare some input data.
x = np.random.rand(1, 3, 640, 640)

# Run inference.
y = model.run(x)

The onnxruntime code is "examples/osrt_python/ort/onnxrt_ep.py", and the run shell is "python3 onnxrt_ep.py -m yolox-s-ti-lite_39p1_57p9.onnx"