Not sure if it has been covered already, but is there any information available on which changes/improvements have been made in 1.5 comparing to 1.0?
BTW, big thanks to NVidia developers for providing source code of dsexample GST plugin and nvgstiva-app application.
Well, at least I found one difference - the models are not compatible. The sample models from Deepsteam 1.0 were crashing nvgstiva-app (segfault in libnvinfer.so.4.0.4 library) when used with Deepstream 1.5. This is rather strange, since I think the new TensorRT should support all layers from the previous releases?
Could you let me know which layers can’t be supported?
We didn’t try our own models yet. What I did was:
- Installed Deepsteam SDK 1.0 (libraries + sample models).
- Installed Deepsteam SDK 1.5 (only libraries)
- Observed that nvgstiva-app is crashing (segfault in libnvinfer.so.4.0.4 library) when used with sample models from SDK 1.0.
- Replaced sample models with sample models from SDK 1.5.
- Observed that nvgstiva-app is no longer crashing when used with sample models.
That brought me to conclusion that SDK 1.5 is not compatible with sample models included in SDK 1.0.
I’m not sure what is the cause of this issue, but you should be able to reproduce it by just following the steps mentioned above.
Can somebody give few insights about what’s new in DeepStream 1.5?