RT-DETR PyTorch - ShapeInferenceError: Cannot export with --simplify for earlier versions of deepstream 6.0

[pullmyleg]

Hi, thanks for all the work with the repo @marcoslucianops . I noticed an issue when exporting a RT-DETR pytorch model for earlier versions of deepstream (6.0.1) with the --simplify tag.

Specifically, the error [ShapeInferenceError] Inferred shape and existing shape differ in rank: (3) vs (0) indicates a mismatch in the expected dimensions of the tensors involved in a multiplication ('Mul') operation.

It looks like the same issue here: https://github.com/onnx/onnx/issues/3565 and it also does successfully produce a .onnx file.

Questions:

1) Can this model support the simplify flag for older versions of deepstream such as 6.0.1? 2) Are you able to suggest a fix @marcoslucianops?

Error:

$ python3 export_rtdetr_pytorch.py -w output/rtdetr_r18vd_6x_coco/checkpoint0013.pth -c configs/rtdetr/rtdetr_r18vd_6x_coco.yml --simplify -s 640

Starting: output/rtdetr_r18vd_6x_coco/checkpoint0013.pth
Opening RT-DETR PyTorch model

Load PResNet18 state_dict

Exporting the model to ONNX
============= Diagnostic Run torch.onnx.export version 2.0.1+cu117 =============
verbose: False, log level: Level.ERROR
======================= 0 NONE 0 NOTE 0 WARNING 0 ERROR ========================

Simplifying the ONNX model
Traceback (most recent call last):
  File "export_rtdetr_pytorch.py", line 110, in <module>
    sys.exit(main(args))
  File "export_rtdetr_pytorch.py", line 83, in main
    model_onnx, _ = onnxsim.simplify(model_onnx)
  File "/home/inviol/.virtualenvs/RTDETR_pytorch/lib/python3.8/site-packages/onnxsim/onnx_simplifier.py", line 199, in simplify
    model_opt_bytes = C.simplify(
onnx.onnx_cpp2py_export.shape_inference.InferenceError: [ShapeInferenceError] (op_type:Mul, node name: /1/Mul): [ShapeInferenceError] Inferred shape and existing shape differ in rank: (3) vs (0) 

Working when there is no issue when exporting with --dynamic flag for later versions:

$ python3 export_rtdetr_pytorch.py -w output/rtdetr_r18vd_6x_coco/checkpoint0013.pth -c configs/rtdetr/rtdetr_r18vd_6x_coco.yml --dynamic

Starting: output/rtdetr_r18vd_6x_coco/checkpoint0013.pth
Opening RT-DETR PyTorch model

Load PResNet18 state_dict

Exporting the model to ONNX
============= Diagnostic Run torch.onnx.export version 2.0.1+cu117 =============
verbose: False, log level: Level.ERROR
======================= 0 NONE 0 NOTE 0 WARNING 0 ERROR ========================

Done: checkpoint0013.onnx

[IronmanVsThanos]

您好，感谢您为回购所做的所有工作@marcoslucianops。我在使用 --simplify 标签为早期版本的 deepstream (6.0.1) 导出 RT-DETR pytorch 模型时注意到一个问题。

具体来说，错误 [ShapeInferenceError] 推断形状和现有形状的等级不同：(3) 与 (0) 表示乘法（“Mul”）运算中涉及的张量的预期维度不匹配。

看起来这里有同样的问题：onnx/onnx#3565，它也成功生成了 .onnx 文件。

问题：

1. 该模型是否可以支持旧版本 Deepstream（例如 6.0.1）的简化标志？
2. 你能提出修复建议吗@marcoslucianops？

错误：

$ python3 export_rtdetr_pytorch.py -w output/rtdetr_r18vd_6x_coco/checkpoint0013.pth -c configs/rtdetr/rtdetr_r18vd_6x_coco.yml --simplify -s 640

Starting: output/rtdetr_r18vd_6x_coco/checkpoint0013.pth
Opening RT-DETR PyTorch model

Load PResNet18 state_dict

Exporting the model to ONNX
============= Diagnostic Run torch.onnx.export version 2.0.1+cu117 =============
verbose: False, log level: Level.ERROR
======================= 0 NONE 0 NOTE 0 WARNING 0 ERROR ========================

Simplifying the ONNX model
Traceback (most recent call last):
  File "export_rtdetr_pytorch.py", line 110, in <module>
    sys.exit(main(args))
  File "export_rtdetr_pytorch.py", line 83, in main
    model_onnx, _ = onnxsim.simplify(model_onnx)
  File "/home/inviol/.virtualenvs/RTDETR_pytorch/lib/python3.8/site-packages/onnxsim/onnx_simplifier.py", line 199, in simplify
    model_opt_bytes = C.simplify(
onnx.onnx_cpp2py_export.shape_inference.InferenceError: [ShapeInferenceError] (op_type:Mul, node name: /1/Mul): [ShapeInferenceError] Inferred shape and existing shape differ in rank: (3) vs (0) 

当使用 --dynamic 标志导出更高版本时没有问题时工作：

$ python3 export_rtdetr_pytorch.py -w output/rtdetr_r18vd_6x_coco/checkpoint0013.pth -c configs/rtdetr/rtdetr_r18vd_6x_coco.yml --dynamic

Starting: output/rtdetr_r18vd_6x_coco/checkpoint0013.pth
Opening RT-DETR PyTorch model

Load PResNet18 state_dict

Exporting the model to ONNX
============= Diagnostic Run torch.onnx.export version 2.0.1+cu117 =============
verbose: False, log level: Level.ERROR
======================= 0 NONE 0 NOTE 0 WARNING 0 ERROR ========================

Done: checkpoint0013.onnx

Hello i have the same question with U ， has you slove this problem？

[pullmyleg]

Hi @IronmanVsThanos, post export the model still converts runs on deepstream X86 succesfully but will not run on a jetson unit.

Will post more detail on this later. But the exported model with error still works.

[IronmanVsThanos]

你好@IronmanVsThanos，导出后模型仍可成功在 deepstream X86 上运行，但不会在 jetson 设备上运行。

稍后将发布更多详细信息。但导出的模型有错误仍然有效。

thank u for your replay，When I use the comand "python3 export_rtdetr_pytorch.py -w output/rtdetr_r18vd_6x_coco/checkpoint0013.pth -c configs/rtdetr/rtdetr_r18vd_6x_coco.yml --dynamic --simplify" , i can create the onnx model successfull,but it still not work on deepstream 6.0.1. and the following error occurs.

WARNING: [TRT]: onnx2trt_utils.cpp:366: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32. WARNING: [TRT]: onnx2trt_utils.cpp:392: One or more weights outside the range of INT32 was clamped ERROR: [TRT]: ModelImporter.cpp:773: While parsing node number 444 [GridSample -> "/0/decoder/decoder/layers.0/cross_attn/GridSample_output_0"]: ERROR: [TRT]: ModelImporter.cpp:774: --- Begin node --- ERROR: [TRT]: ModelImporter.cpp:775: input: "/0/decoder/decoder/layers.0/cross_attn/Reshape_5_output_0" input: "/0/decoder/decoder/layers.0/cross_attn/Reshape_6_output_0" output: "/0/decoder/decoder/layers.0/cross_attn/GridSample_output_0" name: "/0/decoder/decoder/layers.0/cross_attn/GridSample" op_type: "GridSample" attribute { name: "align_corners" i: 0 type: INT } attribute { name: "mode" s: "bilinear" type: STRING } attribute { name: "padding_mode" s: "zeros" type: STRING } ERROR: [TRT]: ModelImporter.cpp:776: --- End node --- ERROR: [TRT]: ModelImporter.cpp:779: ERROR: builtin_op_importers.cpp:4870 In function importFallbackPluginImporter: [8] Assertion failed: creator && "Plugin not found, are the plugin name, version, and namespace correct?" ERROR: Failed to parse onnx file ERROR: failed to build network since parsing model errors.

[pullmyleg]

Your export command looks incorrect. --dynamic (this is for later versions of deepstream) and --simplify (older versions) is not right I don't think. Try:

python3 export_rtdetr_pytorch.py -w output/rtdetr_r18vd_6x_coco/checkpoint0013.pth -c configs/rtdetr/rtdetr_r18vd_6x_coco.yml --simplify -s 640

[IronmanVsThanos]

Thank U for your reply！As you suggested, I used the command "python3 export_rtdetr_pytorch.py -w output/rtdetr_r18vd_6x_coco/checkpoint0013.pth -c configs/rtdetr/rtdetr_r18vd_6x_coco.yml --simplify -s 640" to export onnx on a server that trains rt-detr, but I get the following error：

`Opening RT-DETR PyTorch model

Load PResNet18 state_dict

Exporting the model to ONNX ============= Diagnostic Run torch.onnx.export version 2.0.1+cu117 ============= verbose: False, log level: Level.ERROR ======================= 0 NONE 0 NOTE 0 WARNING 0 ERROR ========================

Simplifying the ONNX model Traceback (most recent call last): File "export_rtdetr_pytorch.py", line 110, in sys.exit(main(args)) File "export_rtdetr_pytorch.py", line 83, in main modelonnx, = onnxsim.simplify(model_onnx) File "/home/inviol/.virtualenvs/RTDETR_pytorch/lib/python3.8/site-packages/onnxsim/onnx_simplifier.py", line 199, in simplify model_opt_bytes = C.simplify( onnx.onnx_cpp2py_export.shape_inference.InferenceError: [ShapeInferenceError] (op_type:Mul, node name: /1/Mul): [ShapeInferenceError] Inferred shape and existing shape differ in rank: (3) vs (0)`

[pullmyleg]

Yes, I get the same error but the model still runs in deepstream and converts (x86 only). It does not work on a Jetson unit.

[IronmanVsThanos]

Do you have the same environment for x86 and deepstream?

[IronmanVsThanos]

when u use the onnx model on jetson unit，is that the error like the follow？ WARNING: [TRT]: onnx2trt_utils.cpp:366: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32. WARNING: [TRT]: onnx2trt_utils.cpp:392: One or more weights outside the range of INT32 was clamped ERROR: [TRT]: ModelImporter.cpp:773: While parsing node number 444 [GridSample -> "/0/decoder/decoder/layers.0/cross_attn/GridSample_output_0"]: ERROR: [TRT]: ModelImporter.cpp:774: --- Begin node --- ERROR: [TRT]: ModelImporter.cpp:775: input: "/0/decoder/decoder/layers.0/cross_attn/Reshape_5_output_0" input: "/0/decoder/decoder/layers.0/cross_attn/Reshape_6_output_0" output: "/0/decoder/decoder/layers.0/cross_attn/GridSample_output_0" name: "/0/decoder/decoder/layers.0/cross_attn/GridSample" op_type: "GridSample" attribute { name: "align_corners" i: 0 type: INT } attribute { name: "mode" s: "bilinear" type: STRING } attribute { name: "padding_mode" s: "zeros" type: STRING } ERROR: [TRT]: ModelImporter.cpp:776: --- End node --- ERROR: [TRT]: ModelImporter.cpp:779: ERROR: builtin_op_importers.cpp:4870 In function importFallbackPluginImporter: [8] Assertion failed: creator && "Plugin not found, are the plugin name, version, and namespace correct?" ERROR: Failed to parse onnx file ERROR: failed to build network since parsing model errors.

by the way，have U tried the Ultrlytics rt-detr？

[pullmyleg]

Do you have the same environment for x86 and deepstream?

1. Yes - in x86 environment (deepstream 6.0.1) it works without any issues.
2. Have not tried Ultralytics, but they only support L & XL models. Have you tried Ultralytics?
3. I am going to try the Paddle version. Have you tried this?
4. Yes - Same error on jetson:

   
   ERROR: [TRT]: ModelImporter.cpp:774: --- Begin node ---
   ERROR: [TRT]: ModelImporter.cpp:775: input: "/0/decoder/decoder/layers.0/cross_attn/Reshape_5_output_0"
   input: "/0/decoder/decoder/layers.0/cross_attn/Reshape_6_output_0"
   output: "/0/decoder/decoder/layers.0/cross_attn/GridSample_output_0"
   name: "/0/decoder/decoder/layers.0/cross_attn/GridSample"
   op_type: "GridSample"
   attribute {
   name: "align_corners"
   i: 0
   type: INT
   }
   attribute {
   name: "mode"
   s: "bilinear"
   type: STRING
   }
   attribute {
   name: "padding_mode"
   s: "zeros"
   type: STRING
   } ```

[pullmyleg]

Smaller options is available: https://github.com/orgs/ultralytics/discussions/2545

[IronmanVsThanos]

1.Have not tried Ultralytics, but they only support L & XL models. Have you tried Ultralytics? I tried the L model with onnx opset version 12，but I get the following error：

(yolov8) hx@FitServer-R4200-V5:/mnt/sda1/Deep_learning/code/yolov8/ultralytics$ python3 export_rtdetr_ultralytics.py -w rtdetr-l.pt --simplify

Starting: rtdetr-l.pt Opening RT-DETR Ultralytics model

Ultralytics YOLOv8.0.155 Python-3.8.18 torch-1.9.1+cu111 CPU (Intel Xeon Gold 6148 2.40GHz) rt-detr-l summary: 494 layers, 32148140 parameters, 0 gradients

Creating labels.txt file

Exporting the model to ONNX Traceback (most recent call last): File "export_rtdetr_ultralytics.py", line 124, in sys.exit(main(args)) File "export_rtdetr_ultralytics.py", line 92, in main torch.onnx.export(model, onnx_input_im, onnx_output_file, verbose=False, opset_version=args.opset, File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/onnx/init.py", line 275, in export return utils.export(model, args, f, export_params, verbose, training, File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/onnx/utils.py", line 88, in export _export(model, args, f, export_params, verbose, training, input_names, output_names, File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/onnx/utils.py", line 689, in _export_model_to_graph(model, args, verbose, input_names, File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/onnx/utils.py", line 458, in _model_to_graph graph, params, torch_out, module = _create_jit_graph(model, args, File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/onnx/utils.py", line 422, in _create_jit_graph graph, torch_out = _trace_and_get_graph_from_model(model, args) File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/onnx/utils.py", line 373, in _trace_and_get_graph_from_model torch.jit._get_trace_graph(model, args, strict=False, _force_outplace=False, _return_inputs_states=True) File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/jit/_trace.py", line 1160, in _get_trace_graph outs = ONNXTracedModule(f, strict, _force_outplace, return_inputs, _return_inputs_states)(*args, kwargs) File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl return forward_call(*input, kwargs) File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/jit/_trace.py", line 127, in forward graph, out = torch._C._create_graph_by_tracing( File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/jit/_trace.py", line 118, in wrapper outs.append(self.inner(trace_inputs)) File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl return forward_call(input, kwargs) File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1039, in _slow_forward result = self.forward(*input, kwargs) File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/nn/modules/container.py", line 139, in forward input = module(input) File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl return forward_call(*input, kwargs) File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1039, in _slow_forward result = self.forward(*input, *kwargs) File "/mnt/sda1/Deep_learning/code/yolov8/ultralytics/ultralytics/nn/tasks.py", line 45, in forward return self.predict(x, args, kwargs) File "/mnt/sda1/Deep_learning/code/yolov8/ultralytics/ultralytics/nn/tasks.py", line 62, in predict return self._predict_once(x, profile, visualize) File "/mnt/sda1/Deep_learning/code/yolov8/ultralytics/ultralytics/nn/tasks.py", line 82, in _predict_once x = m(x) # run File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl return forward_call(*input, *kwargs) File "/home/hx/anaconda3/envs/yolov8/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1039, in _slow_forward result = self.forward(input, **kwargs) File "/mnt/sda1/Deep_learning/code/yolov8/ultralytics/ultralytics/nn/modules/transformer.py", line 81, in forward pos_embed = self.build_2d_sincos_position_embedding(w, h, c) File "/mnt/sda1/Deep_learning/code/yolov8/ultralytics/ultralytics/nn/modules/transformer.py", line 90, in build_2d_sincos_position_embedding grid_w, grid_h = torch.meshgrid(grid_w, grid_h, indexing='ij') TypeError: meshgrid() got an unexpected keyword argument 'indexing' `

2. I am going to try the Paddle version. Have you tried this? no ，did it works？
3. when i use ultralytics yoloV8，i set onnx opset version as 12 it works on jetson

[IronmanVsThanos]

the author of RT-DETR told me that we need to upgrade our TensorRT version >= 8.5.1 to support some operator in rt-detr

[pullmyleg]

Thanks for sharing @IronmanVsThanos!

I think we will need to use a later version of Deepstream. x86 uses a later version of cuda (11.4) which is why it is working and Jetson with Cuda 10.2 won't support this.

[IronmanVsThanos]

yes，I agree with you. 6.0.1 is too late，by the way，the author of RT-DETR told me that they will release RT-DETRV2 and this version does not rely much on the Tensor RT version

[pullmyleg]

Thanks exciting. @IronmanVsThanos did he mention an estimated release date?

[IronmanVsThanos]

He did not mention a specific time, but said that it would be in the near future.