Decrease in accuracy of TensorRt model (pytorch->onnx->TensorRT) on jetson nano

Question

I am inferencing TensorRt model on jetson nano. It is converted from pytorch through onnx. At the time of inference, the accuracy of TensorRt model has decreased drastically. It is an object detection type model.

link for code: https://github.com/NVIDIA/TensorRT/issues/467

Please add more information. How did you infer the accuracy has decreased? Also do add your code. — mibrahimy, Mar 31 '20 at 10:43
I have checked its score and compare it with the original model. — Akshay Sharma, Mar 31 '20 at 10:53
If you could share your code, we would know better what can be the issue. — mibrahimy, Mar 31 '20 at 11:06
hi @IbrahimYousuf if more explanation is required than tell me I will share that as well. — Akshay Sharma, Apr 01 '20 at 03:46
You are using the same postprocessing in TensorRT and PyTorch? — mibrahimy, Apr 01 '20 at 06:21
I don't see it in the code? Please also share the output values from TensorRT and PyTorch. — mibrahimy, Apr 01 '20 at 07:32
Thanks I have figure out the problem. I don't why but type of input effect the accuracy of model. — Akshay Sharma, Apr 01 '20 at 08:45

Akshay Sharma · Accepted Answer · 2020-04-02T04:16:43.873

0

I have figure out the problem. I don't know why but type of input effect the accuracy of model.

I have replace

img_raw = cv2.imread(image_path, cv2.IMREAD_COLOR)

by

img_raw = imageio.imread(image_path)

I will update the answer if I get to know the reason behind it.

edited Apr 02 '20 at 04:16

answered Apr 01 '20 at 08:47

Akshay Sharma

11
8

What type did you change it to? And please specify the code line. – mibrahimy Apr 01 '20 at 09:26
@IbrahimYousuf I have replace img_raw = cv2.imread(image_path, cv2.IMREAD_COLOR) by img_raw = imageio.imread(image_path) – Akshay Sharma Apr 02 '20 at 04:06

Decrease in accuracy of TensorRt model (pytorch->onnx->TensorRT) on jetson nano

1 Answers1