I met a problem in transforming yolov4-608 to INT8 format.

[EthanGuan]

verbose

[TensorRT] VERBOSE: Layer: 161_convolutional Weights: 1044480 HostPersistent: 0 DevicePersistent: 0
[TensorRT] VERBOSE: Layer: (Unnamed Layer* 506) [PluginV2IOExt] Weights: 0 HostPersistent: 16 DevicePersistent: 0
[TensorRT] VERBOSE: Layer: (Unnamed Layer* 507) [PluginV2IOExt] Weights: 0 HostPersistent: 16 DevicePersistent: 0
[TensorRT] VERBOSE: Layer: (Unnamed Layer* 508) [PluginV2IOExt] Weights: 0 HostPersistent: 16 DevicePersistent: 0
[TensorRT] VERBOSE: Total Host Persistent Memory: 63456
[TensorRT] VERBOSE: Total Device Persistent Memory: 13814784
[TensorRT] VERBOSE: Total Weight Memory: 250931200
[TensorRT] VERBOSE: Builder timing cache: created 85 entries, 1044 hit(s)
[TensorRT] VERBOSE: Engine generation completed in 70.9201 seconds.
[TensorRT] VERBOSE: Calculating Maxima
[TensorRT] INFO: Starting Calibration.
[TensorRT] INFO: Post Processing Calibration data in 3.744e-06 seconds.
Traceback (most recent call last):
File "onnx_to_tensorrt.py", line 195, in
main()
File "onnx_to_tensorrt.py", line 184, in main
args.model, args.category_num, args.int8, args.dla_core, args.verbose)
File "onnx_to_tensorrt.py", line 154, in build_engine
engine = builder.build_engine(network, config)
IndexError: _Map_base::at

[jkjung-avt]

This seems to be an issue of ONNX. Reference: onnx/onnx#2458

If you generate the ONNX from a PyTorch model, I think you could try this fix: onnx/onnx#2417 (comment). For example,

    torch.onnx.export(model, dummy_input, “yolov4-608.onnx”, export_params=True, keep_initializers_as_inputs=True, verbose=True)

[EthanGuan]

The model is downloaded from the original darknet project, and I use your protobuf install script, onnx 1.4.1.

Perhaps it is a package conflict? I installed pytorch on xavier from [(https://forums.developer.nvidia.com/t/pytorch-for-jetson-version-1-8-0-now-available/72048)]

Thank you for your quick reply.

[jkjung-avt]

Check if your onnx 1.4.1 has been replaced by a newer version (during the process of installing pytorch or something else). Otherwise, I have not met this problem myself. It would be difficult for me to check this issue further...

[EthanGuan]

Hi, JK

I still cannot handle this issue. I uninstalled pytorch, and make sure when I run onnx_to_tensorrt.py, the ONNX version is 1.4.1.

But the DLA FP16 works for me. On yolov4-608x608, it has 4FPS.

[jkjung-avt]

IndexError: _Map_base::at

I cannot reproduce/debug this issue, so I don't have further suggestions now.

[jkjung-avt]

Closing this issue as "unreproducible".