Cuda failure: status=1 in cuResData at line 316

[jeharul]

Please provide complete information as applicable to your setup.

• Hardware Platform (GPU) T4
• DeepStream Version 6.0
• TensorRT Version 8.0.1
• NVIDIA GPU Driver Version (valid for GPU only) 470.63.01

Hi Team,

We executed the Action detection model by running the following command through DeepStream SDK.

$ deepstream-3d-action-recognition -c deepstream_action_recognition_config.txt

Made all the necessary changes in the appropriate config and txt files as mentioned in the README file and got the following errors.

Please advise.

Thanks
Jeharul

[jeharul]

Hi @amycao ,

Could you please assign someone to look into this. We are actually blocked because of this error.

Regards
Jeharul

[jeharul]

Hi @amycao ,

Now we are not getting the Cuda Error as mentioned in our issue any more, but the below errors are still blocking the execution of the command.

Please help.

[jeharul]

Hi @amycao ,

Could you please redirect this to someone as we are completely blocked with some of our experiments.

[amycao]

Did you install nvidia driver?

[jeharul]

Hi @amycao . Driver already got installed. We dont have a display for time being with a T4 GPU in use. All we are doing for other pretrained models like Licent Plate Recognition and People Recognition is to save the inferred videos to the disk.

In case of Action Recognition, we dont see any configurations for saving the inferred videos to the disk.

[amycao]

You can refer to licence plate app to add “encode output to videos” code into Action Recognition app.

[jeharul]

Will take a look and let you know @amycao . Thanks

[jeharul]

Hi @amycao. I can’t find anything which specifies about “encode output to videos” in the deepstream_lpr_app.c file. Did you specify about a comment with that text or the logic related to that. It is because the License Plate recognition only takes mp4 but the action recognition is sending the mov file formats as inputs by default.

[amycao]

I mean logic, not comment.
gst_element_link_many (nvosd, queue9, nvvidconv1, capfilt, queue10,
nvh264enc, sink, NULL))
the bold part.
you still need other related code.

[jeharul]

I have been looking at it since yesterday. There are some differences in the number of queues and input file format (mp4 only preferred for license plate).

Anyways, thanks for the reply. I am already going in the proper direction then. Will try to make all the necessary changes and get back to you in case of any issues.

[jeharul]

Hi @amycao ,

Have made necessary changes to accommodate the nvh264enc and sink logic in the code. We are still getting the below error.

In the below code, when we replace the nvosd with streammux we are able to save the video (but without any inference).

gst_element_link_many (nvosd, queue9, nvvidconv1, capfilt, queue10,
nvh264enc, sink , NULL))

[amycao]

Please add GST_DEBUG=5 at the beginning of your command to get more logs for analysis.

In the below code, when we replace the nvosd with streammux we are able to save the video (but without any inference).

You do not have nvinfer components between them, that’s why you did not get the inference.

[jeharul]

Hi @amycao

We did not miss anything related to infer in our code, we just followed the License Plate model as is (since it was working without any issue). Please find below the code refferred.

g_object_set (G_OBJECT (sink), “location”, argv[argc-1],NULL);
gst_bin_add_many (GST_BIN (pipeline), nvvidconv1, nvh264enc, capfilt,
queue9, queue10, NULL);

if (!gst_element_link_many (nvosd, queue9, nvvidconv1, capfilt, queue10,
       nvh264enc, sink, NULL)) {
  g_printerr ("OSD and sink elements link failure.\n");
  return -1;
}

If you look above, gst_element_link_many (nvosd, queue9, nvvidconv1, capfilt, queue10,
** nvh264enc, sink, NULL)** doesn’t have anything to infer either but it is working fine.

[jeharul]

Hi @amycao . We found that the issue was with the way the pipeline was constructed. We will further look at the inferred video (saved on the disk) and get back to you in case of any queries. Thanks for your support.

[jeharul]

Hi @amycao.

We are now able to successfully save the inferred video of the action detection model on to our disk.

Thanks for following up.

[amycao]

Glad to know.

[jeharul]

Hi @amycao.

Pasting the code changes below. Hope it would one day help others.