I am using Jetson Xavier with 4.2 Jetpack, cuda version is 10.0, tensorrt 5.1.6.1.
According to tlt introduction, I trained a SSD resnet18 model which located in NGC. But encountered some problems.
1, First problem: tlt-convert failed
nvidia@nvidia-desktop:~/samba/temp/tlt-converter-trt5.1$ ./tlt-converter -k $KEY -d 3,384,1248 -o NMS -e ssd-trt.engine ssd_resnet18_epoch_180.etlt
[ERROR] UffParser: Validator error: FirstDimTile_4: Unsupported operation _BatchTilePlugin_TRT
[ERROR] Failed to parse uff model
[ERROR] Network must have at least one output
[ERROR] Unable to create engine
Segmentation fault (core dumped)
It tells that _BatchTilePlugin_TRT not defined.
Then I compiled the TensorRT OSS( GitHub - NVIDIA/TensorRT: TensorRT is a C++ library for high performance inference on NVIDIA GPUs and deep learning accelerators. ) with branch of 5.1. compile successfully, replace the libnvinfer_plugin libs in /usr/lib/aarch64-linux-gnu.
Later I run tlt-converter successfully and got the engine file.
nvidia@nvidia-desktop:~/samba/temp/tlt-converter-trt5.1$ ./tlt-converter -k $KEY -d 3,384,1248 -o NMS -e ssd-trt.engine ssd_resnet18_epoch_180.etlt
[INFO] UFFParser: parsing Input
[INFO] UFFParser: Applying order forwarding to: Input
[INFO] UFFParser: parsing conv1/kernel
[INFO] UFFParser: Applying order forwarding to: conv1/kernel
[INFO] UFFParser: parsing conv1/convolution
[INFO] UFFParser: Applying order forwarding to: conv1/convolution
[INFO] UFFParser: parsing conv1/bias
[INFO] UFFParser: Applying order forwarding to: conv1/bias
[INFO] UFFParser: parsing conv1/BiasAdd
[INFO] UFFParser: Applying order forwarding to: conv1/BiasAdd
[INFO] UFFParser: parsing bn_conv1/moving_variance
[INFO] UFFParser: Applying order forwarding to: bn_conv1/moving_variance
[INFO] UFFParser: parsing bn_conv1/Reshape_1/shape
[INFO] UFFParser: Applying order forwarding to: bn_conv1/Reshape_1/shape
[INFO] UFFParser: parsing bn_conv1/Reshape_1
[INFO] UFFParser: Applying order forwarding to: bn_conv1/Reshape_1
[INFO] UFFParser: parsing bn_conv1/batchnorm/add/y
[INFO] UFFParser: Applying order forwarding to: bn_conv1/batchnorm/add/y
[INFO] UFFParser: parsing bn_conv1/batchnorm/add
……
[INFO] Block size 46080
[INFO] Block size 15360
[INFO] Total Activation Memory: 1431005696
[INFO] Detected 1 input and 2 output network tensors.
[INFO] Data initialization and engine generation completed in 0.14853 seconds.
2，But I found another problem in deserializing the engine file.
nvidia@nvidia-desktop:~/samba/temp/resnet10_SSD-TensorRT/build/bin$ ./resnet_ssd
loading network profile from cache…
createInference
getPluginCreator could not find plugin BatchTilePlugin_TRT version 1 namespace
Cannot deserialize plugin BatchTilePlugin_TRT
getPluginCreator could not find plugin BatchTilePlugin_TRT version 1 namespace
Cannot deserialize plugin BatchTilePlugin_TRT
getPluginCreator could not find plugin BatchTilePlugin_TRT version 1 namespace
Cannot deserialize plugin BatchTilePlugin_TRT
getPluginCreator could not find plugin BatchTilePlugin_TRT version 1 namespace
Cannot deserialize plugin BatchTilePlugin_TRT
getPluginCreator could not find plugin BatchTilePlugin_TRT version 1 namespace
Cannot deserialize plugin BatchTilePlugin_TRT
getPluginCreator could not find plugin BatchTilePlugin_TRT version 1 namespace
Cannot deserialize plugin BatchTilePlugin_TRT
getPluginCreator could not find plugin NMS_TRT version 1 namespace
Cannot deserialize plugin NMS_TRT
Segmentation fault (core dumped)
It seems that BatchTilePlugin_TRT not work at all. But I compile and replace the libnvinfer_plugin libs which should include it. Is there any wrong in these procedures? Or any more step should do?