yiyangcai / crnn-pytorch2tensorrt-via-onnx Goto Github PK

Hi, thanks for your work!
I have some problems. When I ran 'pytorch2onnx.py', it worked! But I failed when running onnx2tensorort.py.
[TensorRT] ERROR: (Unnamed Layer* 144) [Slice]: slice size must be positive, size = [0,0,0]
[TensorRT] ERROR: (Unnamed Layer* 145) [Slice]: slice size must be positive, size = [0,0,0]

[TensorRT] ERROR: (Unnamed Layer* 316) [Slice]: slice size must be positive, size = [0,0,0]
[TensorRT] ERROR: (Unnamed Layer* 317) [Slice]: slice size must be positive, size = [0,0,0]

I've tried to print shape of the output. It's shown as : the final shape:(26, 1, 37)

I've tested on different versions of pytorch: 1.2, 1.3.1 ,1.4, they all failed.

Looking forward to your reply!

the generated onnx input width which is image lenght is fixed? Then how to handle different lenght charactors input prediction when using onnx or tensorrt inference?

Thank you for this wonderful project.
But I met a problem in python api of tensorrt to inference the model which is trained on pytorch.
I found the speed is slower in tensorrt than pytorch.
Like this project, I have 7226 classes of char, so the speed is 2.9ms/pic (32*100) in tensorrt (if dynamic input, the speed is 9ms/pic) . But when I test the same picture in pytorch(pytorch is dynamic input), it show only 3ms/pic.
Have you test the speed in pytorch?