Engine file is rebuilt on every pipeline start

[a.kantsemal]

• Jetson Orin Nano
• DS 7.0.0
• JP 6.0.0 v2

Warning from NvDsInferContextImpl::generateBackendContext() <nvdsinfer_context_impl.cpp:2188> [UID = 
1]: deserialize backend context from engine from file :/path/to/d0_avlp_544_960_f32.engine failed, try rebuild

Even though the .engine file was built and saved to the path successfully.

What could be the troubleshooting steps? Thanks.

[a.kantsemal]

cc @kesong

[miguel.taylor]

Hi,

nvinfer saves the engine after generating it, but the default location is /opt/, and the default user doesn’t have permission to write there. You can try running the application with sudo or modify the permissions of the destination folder.

[junshengy]

@a.kantsemal

deserialize backend context from engine

This may be caused by the mismatch between your engine file and the batch-size/network-mode in the gie configuration file.

Refer this documentation Gst-nvinfer — DeepStream documentation 6.4 documentation

[a.kantsemal]

Thanks for your responses guys!

@miguel.taylor

default user doesn’t have permission to write there.

The folder that I am using has all the permissions (775).

@junshengy

This may be caused by the mismatch between your engine file and the batch-size/network-mode in the gie configuration file.

The cfg is not changed between the runs. So the first run creates an .engine file, and the second run without any modifications already must be rebuilt.

Anything else I can check?

[junshengy]

The engine file created by deepstream is usually named xxxx._bx_gpux_int8.engine. such as resnet18_trafficcamnet_pruned.onnx_b1_gpu0_int8.engine

Do you modify the batch-size property of nvinfer in the program dynamiclly ? If the batch-size not match with engine file, it will be rebuilt.

[a.kantsemal]

@junshengy

xxxx._bx_gpux_int8.engine

I keep it default, but does it matter?

@junshengy

Do you modify the batch-size property of nvinfer in the program dynamiclly ?

No, they are set in the configs

[junshengy]

@a.kantsemal

I keep it default, but does it matter?

This is true. If the model-engine-file does not match the batch-size/network-mode/precision(int8/fp32..), it will be rebuilt every time it is run.

[miguel.taylor]

After the engine is saved, make sure to update the config file to match the saved engine’s name.

[cihn]

I resolved this issue by running the following command:
sudo chown -R nvidia:nvidia /opt/nvidia/deepstream/deepstream

[a.kantsemal]

Thanks, issue resolved!