About export model

[Egorundel]

Hello, I have a question related to exporting a model. Ultimately, I'm interested in the TensorRT (.trt or .engine) format, but I'm also interested in ONNX (.onnx). Then from ONNX I will be able to make the TensorRT engine myself using trtexec from NVIDIA.

My question is, is it possible to export a trained model to ONNX?

[NVigne-cloud]

Hi, I am also interested. Have you tried ?

[Egorundel]

@NVigne-cloud Haven't tried :(

Apparently, we will not wait for the solution of our question

[SangbumChoi]

Currently I'm trying to convert into tensorRT. converting to ONNX seems OK but like always I think tensorRT conversion need some extra work.

[SangbumChoi]

it seems that 2d RoPE

EVA/EVA-02/det/detectron2/modeling/backbone/utils.py

Line 309 in b947f3f

class VisionRotaryEmbeddingFast(nn.Module):

makes trouble with following error
[graphShapeAnalyzer.cpp::analyzeShapes::1872] Error Code 4: Miscellaneous (IElementWiseLayer /backbone/net/blocks.2/attn/rope_glb_1/Mul: broadcast dimensions must be conformable)

[xinlin-xiao]

any update? converting to ONNX is OK but converting to tensorRT had returned a erros:
`[12/18/2023-15:05:48] [E] [TRT] [graph.cpp::symbolicExecute::611] Error Code 4: Internal Error (/ScatterND: an IScatterLayer cannot be used to compute a shape tensor)
[12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:726: While parsing node number 90 [Pad -> "/Pad_output_0"]:
[12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:727: --- Begin node ---
[12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:728: input: "/Div_output_0"
input: "/Cast_2_output_0"
input: "/Constant_36_output_0"
output: "/Pad_output_0"
name: "/Pad"
op_type: "Pad"
attribute {
name: "mode"
s: "constant"
type: STRING
}

[12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:729: --- End node ---
[12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:732: ERROR: ModelImporter.cpp:185 In function parseGraph:
[6] Invalid Node - /Pad
[graph.cpp::symbolicExecute::611] Error Code 4: Internal Error (/ScatterND: an IScatterLayer cannot be used to compute a shape tensor)
`

[woshidengweimo]

is there anyone tells me how to connvert "pth" to "onnx"

[LordonCN]

any update? converting to ONNX is OK but converting to tensorRT had returned a erros: `[12/18/2023-15:05:48] [E] [TRT] [graph.cpp::symbolicExecute::611] Error Code 4: Internal Error (/ScatterND: an IScatterLayer cannot be used to compute a shape tensor) [12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:726: While parsing node number 90 [Pad -> "/Pad_output_0"]: [12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:727: --- Begin node --- [12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:728: input: "/Div_output_0" input: "/Cast_2_output_0" input: "/Constant_36_output_0" output: "/Pad_output_0" name: "/Pad" op_type: "Pad" attribute { name: "mode" s: "constant" type: STRING }

[12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:729: --- End node --- [12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:732: ERROR: ModelImporter.cpp:185 In function parseGraph: [6] Invalid Node - /Pad [graph.cpp::symbolicExecute::611] Error Code 4: Internal Error (/ScatterND: an IScatterLayer cannot be used to compute a shape tensor) `

hello, I meet the same error, have you solved it?

[xinlin-xiao]

Just pip install onnxslim==0.1.9.1 And OnnxSlim-main# onnxslim model-input-no-mask.onnx slimmed_model-input-no-mask.onnx optimization --skip_fusion_patterns FusionGelu  can being solved it but onnx2trt will return a new errors,I can not solved the new errors

…

---Original--- From: ***@***.***> Date: Thu, Jan 4, 2024 11:32 AM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [baaivision/EVA] About export model (Issue #117) any update? converting to ONNX is OK but converting to tensorRT had returned a erros: `[12/18/2023-15:05:48] [E] [TRT] [graph.cpp::symbolicExecute::611] Error Code 4: Internal Error (/ScatterND: an IScatterLayer cannot be used to compute a shape tensor) [12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:726: While parsing node number 90 [Pad -> "/Pad_output_0"]: [12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:727: --- Begin node --- [12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:728: input: "/Div_output_0" input: "/Cast_2_output_0" input: "/Constant_36_output_0" output: "/Pad_output_0" name: "/Pad" op_type: "Pad" attribute { name: "mode" s: "constant" type: STRING } [12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:729: --- End node --- [12/18/2023-15:05:48] [E] [TRT] ModelImporter.cpp:732: ERROR: ModelImporter.cpp:185 In function parseGraph: [6] Invalid Node - /Pad [graph.cpp::symbolicExecute::611] Error Code 4: Internal Error (/ScatterND: an IScatterLayer cannot be used to compute a shape tensor) ` hello, I meet the same error, have you solved it? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>