Nvinfer Internal data stream error

Accelerated Computing Intelligent Video Analytics DeepStream SDK
#1

I created a custom bbox parser for my custom onnx model which I’m using with nvinfer plugin.

Everything works fine when I’m using it as primary detector but when I set the same to secondary it just stops after giving some warnings.

• Hardware Platform: JETSON NANO
• DeepStream Version : 5.0
• JetPack Version : 4.4DP
• TensorRT Version : 7.0

Pipeline is PREROLLING ...
Opening in BLOCKING MODE
NvMMLiteOpen : Block : BlockType = 261
NVMEDIA: Reading vendor.tegra.display-size : status: 6
NvMMLiteBlockCreate : Block : BlockType = 261
Pipeline is PREROLLED ...
Setting pipeline to PLAYING ...
New clock: GstSystemClock
0:02:35.326956106  8953   0x55632acde0 WARN                 nvinfer gstnvinfer.cpp:1240:convert_batch_and_push_to_input_thread:<nvinfer1> error: NvBufSurfTransform failed with error -2 while converting buffer
ERROR: from element /GstPipeline:pipeline0/GstNvInfer:nvinfer1: NvBufSurfTransform failed with error -2 while converting buffer
Additional debug info:
/dvs/git/dirty/git-master_linux/deepstream/sdk/src/gst-plugins/gst-nvinfer/gstnvinfer.cpp(1240): convert_batch_and_push_to_input_thread (): /GstPipeline:pipeline0/GstNvInfer:nvinfer1
Execution ended after 0:00:04.095490590
Setting pipeline to PAUSED ...
Setting pipeline to READY ...
0:02:35.329451363  8953   0x55632acde0 WARN                 nvinfer gstnvinfer.cpp:1946:gst_nvinfer_output_loop:<nvinfer0> error: Internal data stream error.
0:02:35.330817900  8953   0x55632acde0 WARN                 nvinfer gstnvinfer.cpp:1946:gst_nvinfer_output_loop:<nvinfer0> error: streaming stopped, reason error (-5)
Setting pipeline to NULL ...
Freeing pipeline ...
#2

Hi,

Looks like gst_nvinfer_output_loop is returning GST_FLOW_ERROR that seems to happen when “pushing buffer to downstream element”. There is probably something wrong with the buffer. I would further debug the issue by checking the buffer in nvinfer (sources/gst-plugins/gst-nvinfer/gstnvinfer.cpp).

#3

Could you share your nvinfer config file for further check?
BTW, what’s the resolution of your input and have you do resize in streammux?

#4

Resolution in streammux 1280x720
Config files for pgie and sgie.
pgie.txt (4.0 KB)
sgie.txt (3.5 KB)

Pipeline description:
gst-launch-1.0 uridecodebin uri=file:///home/Username/pipline/sample_day.mp4 ! m.sink_0 nvstreammux name=m batch-size=1 width=1280 height=720 ! nvvideoconvert ! nvinfer config-file-path=/home/Username/pipline/redaction_with_deepstream/configs/pgie_config_fd_lpd.txt ! nvvideoconvert ! nvinfer config-file-path=/home/Username/Downloads/opt/nvidia/deepstream/deepstream-5.0/sources/apps/sample_apps/testocr/dstest1_pgie_config_cpy.txt ! nvvideoconvert ! nvdsosd ! nvvideoconvert ! fpsdisplaysink name=fpssink text-overlay=true video-sink=xvimagesink sync=0

Also, I couldn’t reproduce this error on any other model.

#5

What’s the input dims of your pgie/sgie model? I cannot find it in your config files.

#6

For pgie I’m using sample model in redaction sample from Nvidia https://github.com/NVIDIA-AI-IOT/redaction_with_deepstream/tree/master/fd_lpd_model

and for sgie, input dims are 3x200x200

#8

Could you run the engine using trtexec?

#9

Yes. It ran normally using trtexec.

#10

OK, we have similar issue that the scale factor cannot be larger than 16, and we will fix it on next release.
currently can you modify the detected-min-h/detected-min-w of pgie both to lager than 15 since your network resolution is 3 * 200 * 200.

1 Like
#11

That explains it. Thank you.