RTX 4060 and TensorRT

[EO]

Hi, I would like to know if ultimateALPR supports RTX 4060 for TensorRT?

My setup is:

TensorRT Version: 8.6.1.6 (TensorRT 8.6 GA for Linux x86_64 and CUDA 11.0, 11.1, 11.2, 11.3, 11.4, 11.5, 11.6, 11.7 and 11.8)

OS: Ubuntu 22.04

NVIDIA GPU: RTX 4060

NVIDIA Driver Version: 535.183.01

CUDA Version: 12.2

CUDNN Version: 8.9.0

I manually installed CUDA 11.4.

I meet all dependencies of "ldd libultimatePluginTensorRT.so".

When running "trt_optimizer", all models were created correctly.

However, when running recognizer or benchmark with "certain" images, I have several errors:

.- No detection
.- Incorrect detection
.- Wrong confidence values

And when I run it with my app, it crashes randomly.

These errors only happen if I enable TensorRT.

I attach logs.

If necessary, I can send the test image, which is quite clear and obvious for detection (I can't share it publicly).

Thanks for the help.

[Mamadou DIOP]

Hi,

Support for TensorRT was added in v3.13 and is new. It may have issues on certain GPUs.

Reports about TensorRT will be managed internally with very low priority. So, don't expect to have any fix in the coming weeks.

This said, using RTX 4060 you can update your Tensorflow as explained at https://github.com/DoubangoTelecom/ultimateALPR-SDK/blob/master/samples/c++/README.md#migration-tf2 to get GPGPU acceleration using CUDA.

TensorRT is x2 to x3 times faster than Tensorflow+CUDA (https://github.com/DoubangoTelecom/ultimateALPR-SDK/tree/master/samples/c%2B%2B/benchmark#peformance-numbers) but both are fast enough for any use-case.

My advise: Find a public image that you can share here that reproduce the issue using benchmark app and I'll check.

If the issue happens with "certain" images only, then it's probably something related to the NMS algorithm.

--
You received this message because you are subscribed to the Google Groups "doubango-ai" group.
To unsubscribe from this group and stop receiving emails from it, send an email to doubango-ai...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/doubango-ai/5cd65cd7-2ec3-4ee8-a834-9d66be83e712n%40googlegroups.com.

[Mamadou DIOP]

Using image at https://platesmania.com/br/foto25907684 I had the issue on RTX3060

Checking out https://github.com/DoubangoTelecom/ultimateALPR-SDK/commit/ecfc7e394329d7b57bb1119509d83c629b5303a2 fixed the issue

    - With command [1] my frame rate is 210

    - With command [2] my frame rate is 136

So, I highly recommend using parallel mode

[1] LD_LIBRARY_PATH=../../../binaries/linux/x86_64:$LD_LIBRARY_PATH ./benchmark     --positive ../../../assets/images/25907684.jpg     --negative ../../../assets/images/london_traffic.jpg     --assets ../../../assets     --charset latin     --trt_enabled true     --loops 100     --rate 1.0     --parallel true

[2] LD_LIBRARY_PATH=../../../binaries/linux/x86_64:$LD_LIBRARY_PATH ./benchmark     --positive ../../../assets/images/25907684.jpg     --negative ../../../assets/images/london_traffic.jpg     --assets ../../../assets     --charset latin     --trt_enabled true     --loops 100     --rate 1.0     --parallel false

On 10/7/2024 3:36 AM, Erick Ospino wrote:

Sorry, here is the HD image.

Not only is TensorRT incredibly faster in my setup, it is also more "accurate". I have images that were only recognized in the online demo and now they are recognized by my setup as well.

On Sun, Oct 6, 2024 at 9:19 PM Erick Ospino <conk...@gmail.com> wrote:

Hi Mamadou, thanks for your answer.

I look for images on https://platesmania.com/ and found somes that allowed me to reproduce the error in recognizer but not in benchmark. I realized that if I disable parallel it is much easier to find an image that replicates the issue with benchmark.

I hope this helps.

--

Se despide agradecido,
Erick Ospino

--

Se despide agradecido,
Erick Ospino