Nvdspreprocess + nvinfer(unet) ，model output error

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Please provide complete information as applicable to your setup.

• Hardware Platform (dGPU)
• DeepStream Version (7.0)
• TensorRT Version(8.6.1)
• NVIDIA GPU Driver Version (545.29.02)
• Issue Type( questions)
• How to reproduce the issue ? (run in
Given the configuration files preprocess_config.txt, nvinfer.txt, run preprocess_unet.py in nvcr.io/nvidia/deepstream:7.0-samples-multiarch. I will give the specific code file below.)
• Requirement details( deepstream-segmentation&deepstream-preprocess mix)

My problem is to perform segmentation on part of the image, so I added nvdspreprocess element before nvinfer. Refer to Tao deploy. I set input-tensor-from-meta=1 and so on, but nvinfer has no output. If input-tensor-from-meta =0, nvinfer (argmax_1) has output.

Below is my running environment, running script, model and configuration file. I look forward to your reply, thank you.

container:
nvcr.io/nvidia/deepstream:7.0-samples-multiarch

script（The .py file cannot be uploaded, so it is named. txt）
preprocess_unet.txt (16.7 KB)

model(Uploading is not supported)
same as deepstream_tao_apps/models/unet/unet_resnet18.onnx

nvdspreprocess config file
config_preprocess.txt (1019 Bytes)

nvinfer config file
dstest1_pgie_configcopy.txt (784 Bytes)

3

Did your project work well before updating to the DeepStream 7.0? Could you attach your video source?

4

Thank you for your quick reply！

No, deepstream6.4 has the same problem.
File source: file:///opt/nvidia/deepstream/deepstream/samples/streams/sample_720p.mp4

log
input-tensor-from-meta=0
log_input-tensor-from-meta_equal0.log (444.1 KB)

input-tensor-from-meta=1
log_input-tensor-from-meta_equal1.log (444.1 KB)

5

How do you check if there is an output of nvinfer?

6 
  masks = pyds.get_segmentation_masks(segmeta)
  masks = np.array(masks, copy=True, order='C')
  # map the obtained masks to colors of 2 classes.
  frame_image = map_mask_as_display_bgr(masks)
  cv2.imwrite("test" + "/" + str(frame_number) + ".jpg", frame_image)

When input-tensor-from-meta=1, i get nothing in the above code contained in seg_src_pad_buffer_probe，but input-tensor-from-meta=1, i get mask.

7

OK. If you use the preprocess, the metadata is attached to the batch_meta directly. You can refer to the l_user=batch_meta.batch_user_meta_list to get the metadata.

8

Thanks for your reply, I will try it.

My problem is that using nvdspreprocess for pre-processing and setting input-tensor-from-meta=1, nvinfer will not pre-process the data but directly use the data output by nvdspreprocess. I have set all nvinfer data preprocessing items in nvdspreprocess, why does nvinfer not output mask?

9

I have attached the reason. Please change your code: l_user = frame_meta.frame_user_meta_list to the l_user=batch_meta.batch_user_meta_list I attached.

10

Thank you! Have a nice day@yuweiw

11

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.