Kevin Chen comments

Results 88 comments of


                                            Kevin Chen

Problem converting ONNX model to TensorRT 7.2.1(CTPN.onnx) (Dynamic ReverseSequence support)

Are you able to share the ONNX model?

Problem converting ONNX model to TensorRT 7.2.1(CTPN.onnx) (Dynamic ReverseSequence support)

I have reproduced the error with my own model, so looking into a possible fix now. If you can still share your model it would be helpful as an additional...

Problem converting ONNX model to TensorRT 7.2.1(CTPN.onnx) (Dynamic ReverseSequence support)

Updated the title. The root cause is that we do not support dynamic dimensions for the ReverseSequence operator. Marking as enhancement to fix in a later release.

Can provide a bazel build file?

No plans to include a bazel build file in the near future.

While parsing node number 7 [Loop]: ERROR: ModelImporter.cpp:92 In function parseGraph:

Thanks for the report. We'll take a look at this issue

In function importTopK: [8] Assertion failed: inputs.at(1).is_weights()

Hi all, TensorRT unfortunately only supports constant values for `k` in its TopK layer. Are the K values in the models actually dynamic? Perhaps they can be constant-folded using the...

Dose trt8.2 has native API to torch.roll?

You can try to use https://github.com/NVIDIA/Torch-TensorRT to see if the operator is supported. Does your model export into ONNX?

Make -j Error - nvonnxparser_static.dir/ModelImporter.cpp.o

The 21.05 container has an older version of TensorRT (7.2) that's not compatible with the code on the master branch (that's meant for TRT 8.0). You can use a later...

Invalid Node - model.0.conv.bias_DequantizeLinear

Are you able to share the model? Dequantize nodes should have floating point input types (activation types in TRT), and this error suggests that there is a type mismatch.

Question about converting onnx quantized model to tensorrt

Hi @Nedae, we only support QuantizeLinear and DequantizeLinear operators in TensorRT 7.2. We are planning to extend support and provide better tooling for prequantized networks in the next TensorRT release.