calibration failed

[andyluo7]

The command I use:
make calibrate RUN_ARGS="--benchmarks=resnet"

The output:

[2020-05-20 01:03:19,521 main.py:291 INFO] Using config files: measurements/Xavier/resnet/SingleStream/config.json,measurements/Xavier/resnet/MultiStream/config.json,measurements/Xavier/resnet/Offline/config.json
[2020-05-20 01:03:19,522 init.py:142 INFO] Parsing config file measurements/Xavier/resnet/SingleStream/config.json ...
[2020-05-20 01:03:19,527 init.py:142 INFO] Parsing config file measurements/Xavier/resnet/MultiStream/config.json ...
[2020-05-20 01:03:19,528 init.py:142 INFO] Parsing config file measurements/Xavier/resnet/Offline/config.json ...
[2020-05-20 01:03:19,529 main.py:295 INFO] Processing config "Xavier_resnet_SingleStream"
[2020-05-20 01:03:19,529 main.py:295 INFO] Processing config "Xavier_resnet_MultiStream"
[2020-05-20 01:03:19,529 main.py:295 INFO] Processing config "Xavier_resnet_Offline"
[2020-05-20 01:03:19,529 main.py:244 INFO] Generating calibration cache for Benchmark "resnet"
Traceback (most recent call last):
File "code/main.py", line 327, in
main()
File "code/main.py", line 323, in main
handle_calibrate(benchmark_name, benchmark_conf)
File "code/main.py", line 248, in handle_calibrate
b.calibrate()
File "/home/aluo/inference_results_v0.5/closed/NVIDIA/code/common/builder.py", line 134, in calibrate
self.initialize()
File "/home/aluo/inference_results_v0.5/closed/NVIDIA/code/resnet/tensorrt/ResNet50.py", line 64, in initialize
raise RuntimeError("ResNet50 onnx model parsing failed! Error: {:}".format(parser.get_error(0).desc()))
RuntimeError: ResNet50 onnx model parsing failed! Error: Assertion failed: !_importer_ctx.network()->hasImplicitBatchDimension() && "This version of the ONNX parser only supports TensorRT INetworkDefinitions with an explicit batch dimension. Please ensure the network was created using the EXPLICIT_BATCH NetworkDefinitionCreationFlag."
Makefile:309: recipe for target 'calibrate' failed
make: *** [calibrate] Error 1

[nvpohanh]

@andyluo7 Our submission codes for v0.5 only support TRT 6. We don't support TRT 7 or above.

[andyluo7]

@nvpohanh is there anyway to downgrade to TRT6 from TRT7 with cuda 10.2 on Ubuntu 18.04?
Is there any inference_result_v0.5 docker container which can run on Xavier?

[nvpohanh]

@andyluo7 We don't use docker on Xavier. TRT6 should be part of JetPack 4.3 DP. If you need to install TRT6 separately, I will ask internally if that's possible.

[andyluo7]

@nvpohanh Pls let me know if I can install TRT6 seperately

[psyhtest]

FWIW, once we successfully did the opposite - copied TRT6 libraries from a board with JetPack 4.3 to a board with JetPack 4.2 (with TRT5), and easily switched between the alternative TRT versions with the Collective Knowledge technology.