search

dusty-nv / jetson-inference

Hello AI World guide to deploying deep-learning inference networks and deep vision primitives with TensorRT and NVIDIA Jetson.
https://developer.nvidia.com/embedded/twodaystoademo
MIT License
7.59k stars 2.94k forks source link

Jetpack 4.3 generated mobilenet.uff file working differently in Jetpack 4.6 #1349

Closed aviontics closed 1 year ago

aviontics commented 2 years ago

I generated a modilenet .uff file in jetpack 4.3 for my custom mobilenet model and when ever I switched to a different TX2 with same jetpack 4.3 a new Tensorrt engine file was generated for that new TX2. All work same and fine in terms of detections.

However when I repeat the steps in Jetson NX with Jetpack 4.6, the Jetson Inference generates a new Tensorrt engine file _ssd_mobilenet_v2coco.uff.1.1.8001.GPU.FP16.engine from the old _ssd_mobilenet_v2coco.uff file . However the detection is different and seems to act differently from the custom mobilenet TensorRT engine generated in TX2 Jetpack 4.3. This may have something to do with versions of Tensorrt or Tensorflow.

Question is, which is it? Is it the TensorRT version or the Tensorflow version that is causing the issue in Jetpack 4.6.

Jetpack has these versions: tensorboard-plugin-wit 1.8.1 tensorflow 2.6.2+nv21.12 tensorflow-estimator 2.6.0 tensorrt 8.0.1.6

Jetpack 4.3 Version tensorrt 6.0.1.10 uff 0.6.5

Notice that tensorflow is not present in the Jetpack 4.3 TX2.

By deduction I would think its the TensorRT version difference that affected the way the mobilenet.uff is used to generate the TensorRT engine file.

Am I correct? If so, how do I fix this issue?

dusty-nv commented 2 years ago

Hi @aviontics, are you saying that your same mobilenet.uff behaves differently on JetPack 4.3 vs JetPack 4.6?

aviontics commented 2 years ago

Hi @aviontics, are you saying that your same mobilenet.uff behaves differently on JetPack 4.3 vs JetPack 4.6?

Yes, exactly. The inference engine is generated from the .uff file from Jetpack 4.3, without any clear error and the video streams fine but the detection is different.

I do not understand how this is possible.

dusty-nv commented 2 years ago

Hmm I am not sure either, sorry. It sounds related to newer version of TensorRT?

FWIW I don't use UFF much, primarily I use ONNX. The only UFF model I use is that ssd-mobilenet-coco one, and I haven't changed it in some time. Going forward I am always using ONNX.

aviontics commented 2 years ago

Is it possible to use an older TensorRT in Jetpack 4.6 or I have to downgrade to older Jetpack like 4.4.

dusty-nv commented 2 years ago

You would need to downgrade to an older version of JetPack.