SIGSEGV on using custom network in DeepStream 2.0 for Tesla

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Hi,

My DeepStream setup is clearly OK, because I am able to run the resnet10 example with customer parser (default provided in libnvparsebbox.so), but I want to run my own network.

I have tested that trtexec has no problems running the network with my provided 2 outputs layer names. But, DeepStream seems to die suddenly (SIGSEGV). GST_DEBUG level of 6 is also not able to show anything different between working sample case, and our network.

Kindly suggest how to go about it.

Last few lines:
0x555cfccc4c00 TRACE structure gststructure.c:154:gst_structure_new_id_empty_with_size: created structure 0x55d6c6777580
0x555cfccc4c00 DEBUG query gstquery.c:675:gst_query_new_custom: creating new query 0x55d6c6402400 custom
0x555cfccc4c00 DEBUG GST_PADS gstpad.c:4174:gst_pad_peer_query:nvinfer0:sink peer query 0x55d6c6402400 (custom)
0x555cfccc4c00 INFO GST_PADS gstpad.c:4230:gst_pad_peer_query:nvinfer0:sink pad has no peer
Caught SIGSEGV

Regards,
Anup

2

Hi,

Could you remove the cache file and try it again?

rm .cache/*

Thanks.

3

Hi,

I ran both after cleaning cache each time. 250MB files diffed, but not a single line of difference! Do you have any detector working with one of the popular open source networks which use custom parser that you can share, on DS2?

Regards,
Anup

4

Hi,

You can uncomment the custom parser parameter in the config_infer_resnet.txt config.

For example:

...
parse-func=0
output-bbox-name=conv2d_bbox
output-blob-names=conv2d_cov
parse-bbox-func-name=parse_bbox_custom_resnet
parse-bbox-lib-name=/path/to/this/directory/libnvparsebbox.so

It can show you how to use a custom parser with DeepStream 2.0.
Thanks.