Looking for tlt-converter-8.21.zip for Isaac SDK (Jetson AGX Xavier Jetpack 4.51)

• Hardware: AGX Xavier
• Network Type Detectnet_v2
• TLT Version docker tag: nvcr.io/nvidia/isaac-ml-training:2020.1
• Training spec file
• How to reproduce the issue ? Notebook says to download tlt-converter from this link:
https://developer.download.nvidia.com/assets/TLT/Secure/tlt-converter-8.21.zip

We’re running the latest version of Isaac SDK, and it has this tutorial:
https://docs.nvidia.com/isaac/doc/tutorials/training_in_docker.html

Everything works and with the Jupyter notebook we are able to go through the training process to detect the tennis ball in simulation. For deployment on the robot with the AGX Xavier running Jetpack 4.51, the instructions in the notebook say to download tlt-converter-8.21.zip but the link is dead. On the Isaac SDK forum and on this forum there are many links to download TLT converter but all links are dead. Could someone at Nvidia please make that TLT converter 8.21 available again so that users of Isaac SDK can complete the tutorial ?

This is for running on actual robot hardware, not simulation. Please do not suggest moving to newer versions of TAO or whatever because the very latest Isaac SDK specifically supports only Jetpack 4.51 and that is what our robots are based on.

Thanks so much in advance.

In the Transfer Learning Toolkit 3.0 documentation

I found the download link for tlt-converter 7.1 for Jetpack 4.5:
cuda10.2_trt7.1_jp4.5-20210209T001136Z-001.zip

Unfortunately, it crashes the Xavier almost instantly, as soon as it reads the etlt file.

Here is the output from TLT converter running the notebook on the x86 development machine with docker:

!tlt-converter $USER_EXPERIMENT_DIR/experiment_dir_final/resnet18_detector.etlt
-k "object-detection-from-sim-pipeline"
-c $USER_EXPERIMENT_DIR/experiment_dir_final/calibration.bin
-o output_cov/Sigmoid,output_bbox/BiasAdd
-d 3,384,1248
-i nchw
-m 64
-t int8
-e $USER_EXPERIMENT_DIR/experiment_dir_final/resnet18_detector.trt
-b 4

[INFO] Reading Calibration Cache for calibrator: EntropyCalibration2
[INFO] Generated calibration scales using calibration cache. Make sure that calibration cache has latest scales.
[INFO] To regenerate calibration cache, please delete the existing one. TensorRT will generate a new calibration cache.
[INFO] Some tactics do not have sufficient workspace memory to run. Increasing workspace size may increase performance, please check verbose output.
[INFO] Detected 1 inputs and 2 output network tensors.

and the resulting model works when tested in simulation in the same notebook. So the etlt and key are correct.

But when transferring the ETLT file and BIN file to the AGX Xavier and running the very same command on the tlt-converter downloaded from the link above , it gets this far:


./tlt-converter resnet18_detector.etlt               
 -k \"object-detection-from-sim-pipeline\"   \
 -c calibration.bin \
 -o output_cov/Sigmoid,output_bbox/BiasAdd \
 -d 3,384,1248  \
 -i nchw  \
 -m 64 \
 -t int8  \
 -e resnet18_detector.trt \
 -b 4

[INFO] Reading Calibration Cache for calibrator: EntropyCalibration2
[INFO] Generated calibration scales using calibration cache. Make sure that calibration cache has latest scales.
[INFO] To regenerate calibration cache, please delete the existing one. TensorRT will generate a new calibration cache.
[INFO] 
[INFO] --------------- Layers running on DLA: 
[INFO] 
[INFO] --------------- Layers running on GPU: 
[INFO] conv1/convolution + activation_1/Relu, block_1a_conv_1/convolution + block_1a_relu_1/Relu, block_1a_conv_2/convolution, block_1a_conv_shortcut/convolution + add_1/add + block_1a_relu/Relu, block_1b_conv_1/convolution + block_1b_relu_1/Relu, block_1b_conv_2/convolution + add_2/add + block_1b_relu/Relu, block_2a_conv_1/convolution + block_2a_relu_1/Relu, block_2a_conv_2/convolution, block_2a_conv_shortcut/convolution + add_3/add + block_2a_relu/Relu, block_2b_conv_1/convolution + block_2b_relu_1/Relu, block_2b_conv_2/convolution + add_4/add + block_2b_relu/Relu, block_3a_conv_1/convolution + block_3a_relu_1/Relu, block_3a_conv_2/convolution, block_3a_conv_shortcut/convolution + add_5/add + block_3a_relu/Relu, block_3b_conv_1/convolution + block_3b_relu_1/Relu, block_3b_conv_2/convolution + add_6/add + block_3b_relu/Relu, block_4a_conv_1/convolution + block_4a_relu_1/Relu, block_4a_conv_2/convolution, block_4a_conv_shortcut/convolution + add_7/add + block_4a_relu/Relu, block_4b_conv_1/convolution + block_4b_relu_1/Relu, block_4b_conv_2/convolution + add_8/add + block_4b_relu/Relu, output_cov/convolution, output_cov/Sigmoid, output_bbox/convolution, 

that’s it… then the Xavier crashes so hard that the power light goes off. This is repeatable.

So it seems like I don’t have the correct version of tlt-converter and don’t know where to get it.

Please try TAO Converter | NVIDIA NGC
wget --content-disposition ‘https://api.ngc.nvidia.com/v2/resources/nvidia/tao/tao-converter/versions/v3.21.11_trt8.0_aarch64/files/tao-converter

Thank you, we tried, however it doesn’t work:

./tao-converter -h
./tao-converter: error while loading shared libraries: libnvinfer.so.8: cannot open shared object file: No such file or directory

This is on our robots that use your very latest Isaac SDK, which is based on Jetpack 4.51, which in turn uses Cuda 10.2

dpkg -l | grep nvinfer
ii libnvinfer-bin 7.1.3-1+cuda10.2 arm64 TensorRT binaries
ii libnvinfer-dev 7.1.3-1+cuda10.2 arm64 TensorRT development libraries and headers
ii libnvinfer-doc 7.1.3-1+cuda10.2 all TensorRT documentation
ii libnvinfer-plugin-dev 7.1.3-1+cuda10.2 arm64 TensorRT plugin libraries
ii libnvinfer-plugin7 7.1.3-1+cuda10.2 arm64 TensorRT plugin libraries
ii libnvinfer-samples 7.1.3-1+cuda10.2 all TensorRT samples
ii libnvinfer7 7.1.3-1+cuda10.2 arm64 TensorRT runtime libraries

The version of tlt-converter that we found in the link above is able to print its help function and is able to read the ETLT model but then crashes so badly that the AGX Xavier’s power light just goes off

Can you use v3.21.08_trt7.1_aarch64 instead?

Thanks for the pointer to the list of releases. And it is interesting that there is both a version with _aarch64 and agx. Unfortunately, both have the exact same issue as the tlt-converter:

carter@carter-001:~/Downloads$ ls
calibration.bin resnet18_detector.etlt tao-converter tlt-converter
carter@carter-001:~/Downloads$ ./tao-converter resnet18_detector.etlt -k "object-detection-from-sim-pipeline" -c calibration.bin -o output_cov/Sigmoid,output_bbox/BiasAdd -d 3,384,1248 -i nchw -m 64 -t int8 -e resnet18_detector.trt -b 4
[INFO] Reading Calibration Cache for calibrator: EntropyCalibration2
[INFO] Generated calibration scales using calibration cache. Make sure that calibration cache has latest scales.
[INFO] To regenerate calibration cache, please delete the existing one. TensorRT will generate a new calibration cache.
[INFO]
[INFO] --------------- Layers running on DLA:
[INFO]
[INFO] --------------- Layers running on GPU:
[INFO] conv1/convolution + activation_1/Relu, block_1a_conv_1/convolution + block_1a_relu_1/Relu, block_1a_conv_2/convolution, block_1a_conv_shortcut/convolution + add_1/add + block_1a_relu/Relu, block_1b_conv_1/convolution + block_1b_relu_1/Relu, block_1b_conv_2/convolution + add_2/add + block_1b_relu/Relu, block_2a_conv_1/convolution + block_2a_relu_1/Relu, block_2a_conv_2/convolution, block_2a_conv_shortcut/convolution + add_3/add + block_2a_relu/Relu, block_2b_conv_1/convolution + block_2b_relu_1/Relu, block_2b_conv_2/convolution + add_4/add + block_2b_relu/Relu, block_3a_conv_1/convolution + block_3a_relu_1/Relu, block_3a_conv_2/convolution, block_3a_conv_shortcut/convolution + add_5/add + block_3a_relu/Relu, block_3b_conv_1/convolution + block_3b_relu_1/Relu, block_3b_conv_2/convolution + add_6/add + block_3b_relu/Relu, block_4a_conv_1/convolution + block_4a_relu_1/Relu, block_4a_conv_2/convolution, block_4a_conv_shortcut/convolution + add_7/add + block_4a_relu/Relu, block_4b_conv_1/convolution + block_4b_relu_1/Relu, block_4b_conv_2/convolution + add_8/add + block_4b_relu/Relu, output_cov/convolution, output_cov/Sigmoid, output_bbox/convolution,
client_loop: send disconnect: Broken pipe

By the way, what does this mean:

[INFO] To regenerate calibration cache, please delete the existing one. TensorRT will generate a new calibration cache.

Should I delete calibration.bin ?

In case it is helpful, I did re-run the last few steps of the notebook to show that the model should be ok:

!ls -rlt $USER_EXPERIMENT_DIR/experiment_dir_retrain/weights

total 43960
-rw-r–r-- 1 lx root 45012680 Mar 15 20:20 resnet18_detector_pruned.tlt

! rm -rf "$USER_EXPERIMENT_DIR/experiment_dir_final"
!mkdir -p $USER_EXPERIMENT_DIR/experiment_dir_final
# Removing a pre-existing copy of the etlt if there has been any.
import os
output_file=os.path.join(os.environ['USER_EXPERIMENT_DIR'],
                         "experiment_dir_final/resnet18_detector.etlt")
if os.path.exists(output_file):
    os.system("rm {}".format(output_file))
!tlt-export detectnet_v2 \
            -m $USER_EXPERIMENT_DIR/experiment_dir_retrain/weights/resnet18_detector_pruned.tlt \
            -o $USER_EXPERIMENT_DIR/experiment_dir_final/resnet18_detector.etlt \
            -k $KEY`
Using TensorFlow backend.
NOTE: UFF has been tested with TensorFlow 1.14.0.
WARNING: The version of TensorFlow installed on this system is not guaranteed to work with UFF.
DEBUG [/usr/lib/python2.7/dist-packages/uff/converters/tensorflow/converter.py:96] Marking ['output_cov/Sigmoid', 'output_bbox/BiasAdd'] as outputs
[TensorRT] INFO: Some tactics do not have sufficient workspace memory to run. Increasing workspace size may increase performance, please check verbose output.
[TensorRT] INFO: Detected 1 inputs and 2 output network tensors.

print(‘Exported model:’)
print(‘------------’)
!ls -lh $USER_EXPERIMENT_DIR/experiment_dir_final

Exported model:

total 43M
-rw-r–r-- 1 lx root 43M Mar 18 13:57 resnet18_detector.etlt

!tlt-int8-tensorfile detectnet_v2 -e $SPECS_DIR/detectnet_v2_retrain_resnet18_kitti.txt
-m 10
-o $USER_EXPERIMENT_DIR/experiment_dir_final/calibration.tensor

Using TensorFlow backend.
2023-03-18 13:58:56,679 [INFO] iva.detectnet_v2.spec_handler.spec_loader: Merging specification from /workspace/specs/detectnet_v2_retrain_resnet18_kitti.txt
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
target/truncation is not updated to match the crop areaif the dataset contains target/truncation.
Writing calibration tensorfile: 100%|███████████| 10/10 [00:04<00:00, 2.32it/s]
Time taken to run iva.detectnet_v2.scripts.calibration_tensorfile:main: 0:00:06.858156.

!rm -rf $USER_EXPERIMENT_DIR/experiment_dir_final/resnet18_detector.etlt
!tlt-export detectnet_v2
-m $USER_EXPERIMENT_DIR/experiment_dir_retrain/weights/resnet18_detector_pruned.tlt
-o $USER_EXPERIMENT_DIR/experiment_dir_final/resnet18_detector.etlt
-k $KEY
–cal_data_file $USER_EXPERIMENT_DIR/experiment_dir_final/calibration.tensor
–data_type int8
–batches 10
–batch_size 4
–max_batch_size 4
–engine_file $USER_EXPERIMENT_DIR/experiment_dir_final/resnet18_detector.trt.int8
–cal_cache_file $USER_EXPERIMENT_DIR/experiment_dir_final/calibration.bin
–verbose

Using TensorFlow backend.
2023-03-18 13:59:18,789 [DEBUG] iva.common.export.base_exporter: Saving etlt model file at: /workspace/tlt-experiments/experiment_dir_final/resnet18_detector.etlt.
2023-03-18 13:59:20,047 [DEBUG] modulus.export._uff: Patching keras BatchNormalization…
2023-03-18 13:59:20,047 [DEBUG] modulus.export._uff: Patching keras Dropout…
2023-03-18 13:59:20,047 [DEBUG] modulus.export._uff: Patching UFF TensorFlow converter apply_fused_padding…
2023-03-18 13:59:20,696 [DEBUG] modulus.export._uff: Unpatching keras BatchNormalization layer…
2023-03-18 13:59:20,696 [DEBUG] modulus.export._uff: Unpatching keras Dropout layer…
NOTE: UFF has been tested with TensorFlow 1.14.0.
WARNING: The version of TensorFlow installed on this system is not guaranteed to work with UFF.
DEBUG [/usr/lib/python2.7/dist-packages/uff/converters/tensorflow/converter.py:96] Marking [‘output_cov/Sigmoid’, ‘output_bbox/BiasAdd’] as outputs
2023-03-18 13:59:21,884 [DEBUG] iva.common.export.base_exporter: Reading input dims from tensorfile.
2023-03-18 13:59:21,885 [DEBUG] modulus.export.data: Opening /workspace/tlt-experiments/experiment_dir_final/calibration.tensor with mode=r
2023-03-18 13:59:21,997 [DEBUG] iva.common.export.base_exporter: Input dims: (3, 368, 640)
2023-03-18 13:59:21,997 [DEBUG] modulus.export.data: Opening /workspace/tlt-experiments/experiment_dir_final/calibration.tensor with mode=r
2023-03-18 13:59:22,130 [DEBUG] modulus.export._tensorrt: read_calibration_cache - no-op
[TensorRT] INFO: Detected 1 inputs and 2 output network tensors.
[TensorRT] WARNING: Current optimization profile is: 0. Please ensure there are no enqueued operations pending in this context prior to switching profiles
[TensorRT] INFO: Starting Calibration with batch size 4.
DEPRECATED: This variant of get_batch is deprecated. Please use the single argument variant described in the documentation instead.
[TensorRT] INFO: Calibrated batch 0 in 0.0717277 seconds.
[TensorRT] INFO: Calibrated batch 1 in 0.0784402 seconds.
[TensorRT] INFO: Calibrated batch 2 in 0.0767087 seconds.
[TensorRT] INFO: Calibrated batch 3 in 0.0837051 seconds.
[TensorRT] INFO: Calibrated batch 4 in 0.0745355 seconds.
[TensorRT] INFO: Calibrated batch 5 in 0.0787708 seconds.
[TensorRT] INFO: Calibrated batch 6 in 0.0739311 seconds.
[TensorRT] INFO: Calibrated batch 7 in 0.0755091 seconds.
[TensorRT] INFO: Calibrated batch 8 in 0.0787474 seconds.
[TensorRT] INFO: Calibrated batch 9 in 0.0769921 seconds.
[TensorRT] INFO: Post Processing Calibration data in 1.134 seconds.
[TensorRT] INFO: Calibration completed in 11.0105 seconds.
2023-03-18 13:59:33,160 [DEBUG] modulus.export._tensorrt: read_calibration_cache - no-op
[TensorRT] INFO: Writing Calibration Cache for calibrator: TRT-7000-EntropyCalibration2
2023-03-18 13:59:33,160 [INFO] modulus.export._tensorrt: Saving calibration cache (size 4144) to /workspace/tlt-experiments/experiment_dir_final/calibration.bin
[TensorRT] INFO: Some tactics do not have sufficient workspace memory to run. Increasing workspace size may increase performance, please check verbose output.
[TensorRT] INFO: Detected 1 inputs and 2 output network tensors.

!tlt-converter $USER_EXPERIMENT_DIR/experiment_dir_final/resnet18_detector.etlt
-k "object-detection-from-sim-pipeline"
-c $USER_EXPERIMENT_DIR/experiment_dir_final/calibration.bin
-o output_cov/Sigmoid,output_bbox/BiasAdd
-d 3,384,1248
-i nchw
-m 64
-t int8
-e $USER_EXPERIMENT_DIR/experiment_dir_final/resnet18_detector.trt
-b 4

[INFO] Reading Calibration Cache for calibrator: EntropyCalibration2
[INFO] Generated calibration scales using calibration cache. Make sure that calibration cache has latest scales.
[INFO] To regenerate calibration cache, please delete the existing one. TensorRT will generate a new calibration cache.
[INFO] Some tactics do not have sufficient workspace memory to run. Increasing workspace size may increase performance, please check verbose output.
[INFO] Detected 1 inputs and 2 output network tensors.

So all the above was on the PC with the Notebook in the docker image. Then I copy this ETLT file and BIN file to the AGX Xavier as follows:

rcp *.etlt carter@192.168.1.101:/home/carter/Downloads
rcp *.bin carter@192.168.1.101:/home/carter/Downloads

And on the Carter’s AGX Xavier I run these commands (remotely with SSH):

carter@carter-001:~/Downloads$ ls -l
total 129600
-rw-r–r-- 1 carter carter 4144 Mar 18 17:20 calibration.bin
-rw-r–r-- 1 carter carter 87573224 Mar 18 17:20 calibration.tensor
-rw-r–r-- 1 carter carter 44874341 Mar 18 17:20 resnet18_detector.etlt
-rwxrwxr-x 1 carter carter 120936 Mar 18 15:45 tao-converter
-rwxrwxr-x 1 carter carter 125136 Mar 16 21:19 tlt-converter

./tao-converter resnet18_detector.etlt
-k "object-detection-from-sim-pipeline"
-c calibration.bin
-o output_cov/Sigmoid,output_bbox/BiasAdd
-d 3,384,1248
-i nchw
-m 64
-t int8
-e resnet18_detector.trt
-b 4

and then I just get this message before it crashes:

[INFO] Reading Calibration Cache for calibrator: EntropyCalibration2
[INFO] Generated calibration scales using calibration cache. Make sure that calibration cache has latest scales.
[INFO] To regenerate calibration cache, please delete the existing one. TensorRT will generate a new calibration cache.
[INFO]
[INFO] --------------- Layers running on DLA:
[INFO]
[INFO] --------------- Layers running on GPU:
[INFO] conv1/convolution + activation_1/Relu, block_1a_conv_1/convolution + block_1a_relu_1/Relu, block_1a_conv_2/convolution, block_1a_conv_shortcut/convolution + add_1/add + block_1a_relu/Relu, block_1b_conv_1/convolution + block_1b_relu_1/Relu, block_1b_conv_2/convolution + add_2/add + block_1b_relu/Relu, block_2a_conv_1/convolution + block_2a_relu_1/Relu, block_2a_conv_2/convolution, block_2a_conv_shortcut/convolution + add_3/add + block_2a_relu/Relu, block_2b_conv_1/convolution + block_2b_relu_1/Relu, block_2b_conv_2/convolution + add_4/add + block_2b_relu/Relu, block_3a_conv_1/convolution + block_3a_relu_1/Relu, block_3a_conv_2/convolution, block_3a_conv_shortcut/convolution + add_5/add + block_3a_relu/Relu, block_3b_conv_1/convolution + block_3b_relu_1/Relu, block_3b_conv_2/convolution + add_6/add + block_3b_relu/Relu, block_4a_conv_1/convolution + block_4a_relu_1/Relu, block_4a_conv_2/convolution, block_4a_conv_shortcut/convolution + add_7/add + block_4a_relu/Relu, block_4b_conv_1/convolution + block_4b_relu_1/Relu, block_4b_conv_2/convolution + add_8/add + block_4b_relu/Relu, output_cov/convolution, output_cov/Sigmoid, output_bbox/convolution,
[INFO] Some tactics do not have sufficient workspace memory to run. Increasing workspace size may increase performance, please check verbose output.

Sometimes I don’t even get that last message

OK one more piece of info. When I use a second SSH to the Carter and run the TOP command I see that the AGX Xavier crashes shortly after the process “nvmap-bz” appears

And for completeness I attached the BIN file and the ETLT file
calibration.bin (4.0 KB)
resnet18_detector.etlt (42.8 MB)

There is not info for crash.
Can you share full log?

Thank you so much for your quick response. Maybe there are different types of crashes on the Jetson processors, but this one causes the Jetson AGX Xavier to power down. So the SSH session terminates and there is no further info. After reboot there is no new file, and even if we had a monitor hooked up instead of SSH it would just go blank. So I am not sure how we can get any further info, but I shared both the BIN file and the ETLT file and the command line used successfully on the PC and used the same command line on the Xavier. Please note that the default password of the original tutorial notebook had the password in quotes, so in order for the Jetson to recognize the password I had to but backslashes in front of the quotes.

-k \"object-detection-from-sim-pipeline\" 

The key is needed to set to the same as the key during “export” or “train”.

More, to narrow down, can you run “tlt-converter” with fp32 mode or fp16 mode?
In these modes, -c calibration.bin is not needed.

Thanks again for your quick response. Just tried it with fp32 mode and fp16 mode, no difference.

But your responses indicated that this is unusual behavior and not commonly encountered. So I started thinking about what might be different, and tried it on a different AGX Xavier dev kit that is on my desk and has nothing but a fresh Jetpack installation and is not part of a fully functional Carter robot. It worked and generated a TRT file!

So I’ll take it from here, figuring what the difference is, or simply just run any required tao conversion on the desktop AGX Xavier and not the robot.

A couple of quick questions.

  1. Since there are no release notes, could you inquire what the difference between tao-converter v3.21.08_trt7.1_aarch64 and v3.21.11_trt7.1_agx is ? Which one should we use on AGX Xavier with Jetpack 4.51 ?

  2. Could you review the arguments ? This is the command line:

./tao-converter resnet18_detector.etlt
-k "object-detection-from-sim-pipeline"
-c ./calibration.bin
-o output_cov/Sigmoid,output_bbox/BiasAdd
-d 3,384,1248
-i nchw
-m 64
-t int8
-e ./resnet18_detector.trt
-b 4

[INFO] Reading Calibration Cache for calibrator: EntropyCalibration2
[INFO] Generated calibration scales using calibration cache. Make sure that calibration cache has latest scales.
[INFO] To regenerate calibration cache, please delete the existing one. TensorRT will generate a new calibration cache.
[INFO]
[INFO] --------------- Layers running on DLA:
[INFO]
[INFO] --------------- Layers running on GPU:
[INFO] conv1/convolution + activation_1/Relu, block_1a_conv_1/convolution + block_1a_relu_1/Relu, block_1a_conv_2/convolution, block_1a_conv_shortcut/convolution + add_1/add + block_1a_relu/Relu, block_1b_conv_1/convolution + block_1b_relu_1/Relu, block_1b_conv_2/convolution + add_2/add + block_1b_relu/Relu, block_2a_conv_1/convolution + block_2a_relu_1/Relu, block_2a_conv_2/convolution, block_2a_conv_shortcut/convolution + add_3/add + block_2a_relu/Relu, block_2b_conv_1/convolution + block_2b_relu_1/Relu, block_2b_conv_2/convolution + add_4/add + block_2b_relu/Relu, block_3a_conv_1/convolution + block_3a_relu_1/Relu, block_3a_conv_2/convolution, block_3a_conv_shortcut/convolution + add_5/add + block_3a_relu/Relu, block_3b_conv_1/convolution + block_3b_relu_1/Relu, block_3b_conv_2/convolution + add_6/add + block_3b_relu/Relu, block_4a_conv_1/convolution + block_4a_relu_1/Relu, block_4a_conv_2/convolution, block_4a_conv_shortcut/convolution + add_7/add + block_4a_relu/Relu, block_4b_conv_1/convolution + block_4b_relu_1/Relu, block_4b_conv_2/convolution + add_8/add + block_4b_relu/Relu, output_cov/convolution, output_cov/Sigmoid, output_bbox/convolution,
[INFO] Some tactics do not have sufficient workspace memory to run. Increasing workspace size may increase performance, please check verbose output.
[INFO] Detected 1 inputs and 2 output network tensors.

Are the values -m 64 and -b 4 optimal for AGX Xavier ?

  1. The TLT toolkit and TAO toolkit say that for efficiency to deploy to Jetson you should use the tlt-converter or tao-converter to create the TRT file for TensorRT. Isaac SDK kit seems to use the ETLT files directly and generates PLAN files on first use.

Is there a difference between these approaches ? Which is the correct or recommended method for robots with AGX Xavier running Jetpack 4.51 / Cuda 10.2 ? Right now there is no newer version of Isaac SDK so we have to use these versions of Jetpack/Cuda and it has TensorRT 7.1apparently.

Again, thanks so much, you’ve been very helpful, and your response pointing to the catalog of tao-converter 3.21 was the solution to the issue of where to find tlt-converter 8.21 mentioned in the notebook.

Glad to know it is working now. Great work!

Please use the aarch64 version.

It is hard to say. Usually we change these parameter when meet OOM issue.

Yes, will generate plan file(i.e., tensorrt engine) on first use. If use tao-converter or tlt-converter, it will generate tensorrt engine in advanced.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.