yolov4-csp-512.onnx to yolov4-csp-512.trt error

[c7934597]

[TensorRT] VERBOSE: ImporterContext.hpp:141: Registering layer: 173_convolutional_mish for ONNX node: 173_convolutional_mish [TensorRT] VERBOSE: ImporterContext.hpp:116: Registering tensor: 173_convolutional_mish for ONNX tensor: 173_convolutional_mish [TensorRT] VERBOSE: ModelImporter.cpp:179: 173_convolutional_mish [Mul] outputs: [173_convolutional_mish -> (1, 1024, 16, 16)], [TensorRT] VERBOSE: ModelImporter.cpp:103: Parsing node: 174_convolutional [Conv] [TensorRT] VERBOSE: ModelImporter.cpp:119: Searching for input: 173_convolutional_mish [TensorRT] VERBOSE: ModelImporter.cpp:119: Searching for input: 174_convolutional_conv_weights [TensorRT] VERBOSE: ModelImporter.cpp:119: Searching for input: 174_convolutional_conv_bias [TensorRT] VERBOSE: ModelImporter.cpp:125: 174_convolutional [Conv] inputs: [173_convolutional_mish -> (1, 1024, 16, 16)], [174_convolutional_conv_weights -> (18, 1024, 1, 1)], [174_convolutional_conv_bias -> (18)], [TensorRT] VERBOSE: builtin_op_importers.cpp:450: Convolution input dimensions: (1, 1024, 16, 16) [TensorRT] VERBOSE: ImporterContext.hpp:141: Registering layer: 174_convolutional for ONNX node: 174_convolutional [TensorRT] VERBOSE: builtin_op_importers.cpp:533: Using kernel: (1, 1), strides: (1, 1), prepadding: (0, 0), postpadding: (0, 0), dilations: (1, 1), numOutputs: 18 [TensorRT] VERBOSE: builtin_op_importers.cpp:534: Convolution output dimensions: (1, 18, 16, 16) [TensorRT] VERBOSE: ImporterContext.hpp:116: Registering tensor: 174_convolutional for ONNX tensor: 174_convolutional [TensorRT] VERBOSE: ModelImporter.cpp:179: 174_convolutional [Conv] outputs: [174_convolutional -> (1, 18, 16, 16)], 程式記憶體區段錯誤

---

問題如上，使用電腦跟Xavier Nx jetpack 4.4來轉都失敗，請問onnx_to_tensorrt.py程式，貌似裡面沒支援yolov4-csp-512嗎?

補充一下，我發現我的yolov4-csp-512這個是用ScaledYOLOv4的cfg訓出來的，有些地方不太一樣，我再試試用darknet的cfg訓練並轉轉看

[jkjung-avt]

https://github.com/jkjung-avt/tensorrt_demos/blob/899770162fecc21475db38d670b5b3467874046e/yolo/download_yolo.sh#L25-L27

I used these cfg and pre-trained weights files, and was able to convert it to a TensorRT engine on Jetson Xavier NX without problem. You could verify it.

[c7934597]

Thank you for that, I use the cfg of darknet, and then I already could convert it to a TensorRT engine.