onnxruntime.capi.onnxruntime_pybind11_state.InvalidGraph: [ONNXRuntimeError] : 10 : INVALID_GRAPH : Load model from ./model1.onnx failed:This is an invalid model. Type Error: Type 'tensor(int64)' of input parameter (655) of operator (Clip) in node (Clip_354) is invalid.

[quanliu1991]

Describe the bug
Use torch.onnx.export() had coverted faster_rcnn_R_50_C4_1x.yaml file of Detectron2 FasterRCNN model to model2.onnx, but when
sess = onnxruntime.InferenceSession(model_path, sess_options=sess_opt, providers=providers) ,an InvalidGraph error occurred.

System information

• OS Platform :Linux ContOS 7.9

• ONNX Runtime installed from binary:

• ONNX Runtime version:gpu 1.10

• Python version:3.7.11

• CUDA/cuDNN version:11.4/8.2.4

• detectron2: 0.6

To Reproduce

• pytorch covert onnx
  use detectron2/tools/deploy/export_model.py --config-file faster_rcnn_R_50_C4_1x.yaml export onnx.
• onnxruntime run onnx

          providers = [
                ('CUDAExecutionProvider', {
                    'device_id': device_id,
                    'cudnn_conv_algo_search': 'HEURISTIC',
                })
            ]
        sess_opt = rt.SessionOptions()
        sess = rt.InferenceSession("./model2.onnx", sess_options=sess_opt, providers=providers)

error log:

onnxruntime.capi.onnxruntime_pybind11_state.InvalidGraph: [ONNXRuntimeError] : 10 : INVALID_GRAPH : Load model from ./model1.onnx failed:This is an invalid model. Type Error: Type 'tensor(int64)' of input parameter (655) of operator (Clip) in node (Clip_354) is invalid.

Expected behavior
model2.onnx run work.

Additional context
moeld2.onnx download link:
https://drive.google.com/file/d/18sRJ6GR2LkhycG3EdYswWKQfDQiuB7HV/view?usp=sharing

[garymm]

This is probably a bug in torch.onnx.export, so if you can repro this on the latest PyTorch nightly you can file a bug in github.com/pytorch/pytorch.

However I think you can probably work around this by exporting with op set version 12.

I guess edit this line:
https://github.com/facebookresearch/detectron2/blob/main/tools/deploy/export_model.py#L127

Why I think this could work:
int64 support for Clip was added in op set 12.
Compare Clip-12 with Clip-11.

[quanliu1991]

@garymm
First , thank you very much, after I re-exported onnx named model2with12.onnx with op set version 12, it works on CUDA.

However, I try to use the same onnx when ep is Tensorrt a new error occurs: TensorRT input: 717 has no shape specified.
code show as below:

providers = [
                ('TensorrtExecutionProvider', {
                    'device_id': 0,
                })]
sess_opt = onnxruntime.SessionOptions()
sess = onnxruntime.InferenceSession(model_path, sess_options=sess_opt, providers=providers)

image=np.random.randint(1, 255, size=(3, 800, 1202), dtype=np.nint8)
sess.run([sess.get_outputs()[0].name], {sess.get_inputs()[0].name: image})

The following error occurs in onnxruntime.InferenceSession:

2022-01-30 19:02:16.145065205 [E:onnxruntime:, inference_session.cc:1448 operator()] Exception during initialization: /onnxruntime_src/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc:925 SubGraphCollection_t onnxruntime::TensorrtExecutionProvider::GetSupportedList(SubGraphCollection_t, int, int, const onnxruntime::GraphViewer&, bool*) const [ONNXRuntimeError] : 1 : FAIL : TensorRT input: 717 has no shape specified. Please run shape inference on the onnx model first. Details can be found in https://www.onnxruntime.ai/docs/reference/execution-providers/TensorRT-ExecutionProvider.html#shape-inference-for-tensorrt-subgraphs

run symbolic_shape_infer have an error occurred.

python symbolic_shape_infer.py --input ./model2with12.onnx  --output ./out_model2with12.onnx --auto_merge --verbose 3

I don't know how to solve this kind of problem，I expected the onnx model to work on ep:Tensorrt.

model2with12.onnx download link:
https://drive.google.com/file/d/1_egymUZukkjzNfNDSVYIzLGpGRBfuIRQ/view?usp=sharing
input image ndarray info:
shape is [3,800,1202]
dtype is unit8

[garymm]

@quanliu1991 please open a new issue for that.
Closing this since you worked around the original issue.

[quanliu1991]

@garymm thank you, I had open a new issue #10443 #.