ryouchinsa
sam2.1_tiny.onnx to trt engine Problems encountered
hi, After I successfully converted two onnx models using the conversion code provided by the project, I need to use tensorrt to accelerate. In the process of onnx->engine, the encoder model can be converted to engine normally, but the decoder model has encountered problems, which are as follows:
[02/15/2025-15:37:40] [E] Error[4]: [graph.cpp::nvinfer1::builder::Node::symbolicExecute::535] Error Code 4: Internal Error (/OneHot: an IIOneHotLayer cannot be used to compute a shape tensor) [02/15/2025-15:37:40] [E] [TRT] ModelImporter.cpp:768: While parsing node number 146 [Tile -> "/Tile_output_0"]: [02/15/2025-15:37:40] [E] [TRT] ModelImporter.cpp:769: --- Begin node --- [02/15/2025-15:37:40] [E] [TRT] ModelImporter.cpp:770: input: "/Unsqueeze_11_output_0" input: "/Reshape_4_output_0" output: "/Tile_output_0" name: "/Tile" op_type: "Tile"
[02/15/2025-15:37:40] [E] [TRT] ModelImporter.cpp:771: --- End node --- [02/15/2025-15:37:40] [E] [TRT] ModelImporter.cpp:774: ERROR: ModelImporter.cpp:195 In function parseGraph: [6] Invalid Node - /Tile [graph.cpp::nvinfer1::builder::Node::symbolicExecute::535] Error Code 4: Internal Error (/OneHot: an IIOneHotLayer cannot be used to compute a shape tensor) [02/15/2025-15:37:40] [E] Failed to parse onnx file [02/15/2025-15:37:40] [I] Finished parsing network model. Parse time: 0.0610398 [02/15/2025-15:37:40] [E] Parsing model failed [02/15/2025-15:37:40] [E] Failed to create engine from model or file. [02/15/2025-15:37:40] [E] Engine set up failed &&&& FAILED TensorRT.trtexec [TensorRT v8600] # trtexec.exe --onnx=./sam2.1_tiny.onnx --saveEngine=./sam2.1_tiny.engine
@CvBokchoy I'm not sure, but one effective way to deal with operations like OneHot or Tile that manipulate shape tensors is to remove the need for dynamic shapes in your model.
- Try to convert the model using dynamic shapes
- Fix sam2 torch code before converting onnx
or check it out this
@CvBokchoy我不确定，但处理诸如
OneHot或Tile操纵形状张量的操作的一种有效方法是消除模型中对动态形状的需求。
- 尝试使用动态形状转换模型
- 在转换 onnx 之前修复 sam2 torch 代码
或者查看这个
Thank you for your reply, which has been of great help to me. I hope to communicate with you in my future study and work. Please leave your email address, My email address is [email protected] and my name is Li Jialin
