Runtime Error and Segmentation Fault during Video Inference with SAM2 Model

[georgebaknis]

1. Export ONNX Models:

   • Cloned [SAM2Export](https://github.com/Aimol-l/SAM2Export/tree/main) repository.
   • Ran the export_onnx.py script to generate the four ONNX files.

2. Setup OrtInference:

   • Cloned OrtInference repository.
   • Installed all required dependencies according to the provided instructions.

3. Encountered Issue when running ./main:
   Initially faced the following error:

   terminate called after throwing an instance of 'Ort::Exception'
   what(): not enough space: expected 8388608, got 16
   

Temporary Fix (Uncertain correctness):

• Commented out the following lines in SAM2.cpp within the function img_decoder_infer:

input_tensor.push_back(Ort::Value::CreateTensor<int64>(memory_info, frame_size.data(), frame_size.size(),
                        this->img_decoder_input_nodes[2].dim.data(),
                        this->img_decoder_input_nodes[2].dim.size()));

• After commenting these lines, the program runs, but then another issue occurs.

Persistent Issue (Segmentation Fault):

• Running inference on video leads to a segmentation fault.
• The frame number at which the segmentation fault occurs varies each run, indicating potential memory handling or concurrency issues.

Environment:

• CUDA: 12.9
• cuDNN: 9.10.2
• GPU: NVIDIA RTX 5090

Questions:

• Is commenting out the mentioned lines the correct approach?
• Could the segmentation fault be related to incorrect tensor shapes or memory management?
• How can I ensure stable inference over video without encountering segmentation faults?

[ChaceYoung]

1. Export ONNX Models:

   • Cloned [SAM2Export](https://github.com/Aimol-l/SAM2Export/tree/main) repository.
   • Ran the export_onnx.py script to generate the four ONNX files.

2. Setup OrtInference:

   • Cloned OrtInference repository.
   • Installed all required dependencies according to the provided instructions.

3. Encountered Issue when running ./main:
   Initially faced the following error:

   terminate called after throwing an instance of 'Ort::Exception'
   what(): not enough space: expected 8388608, got 16
   

Temporary Fix (Uncertain correctness):

• Commented out the following lines in SAM2.cpp within the function img_decoder_infer:

input_tensor.push_back(Ort::Value::CreateTensor(memory_info, frame_size.data(), frame_size.size(),
this->img_decoder_input_nodes[2].dim.data(),
this->img_decoder_input_nodes[2].dim.size()));

• After commenting these lines, the program runs, but then another issue occurs.

Persistent Issue (Segmentation Fault):

• Running inference on video leads to a segmentation fault.
• The frame number at which the segmentation fault occurs varies each run, indicating potential memory handling or concurrency issues.

Environment:

• CUDA: 12.9
• cuDNN: 9.10.2
• GPU: NVIDIA RTX 5090

Questions:

• Is commenting out the mentioned lines the correct approach?
• Could the segmentation fault be related to incorrect tensor shapes or memory management?
• How can I ensure stable inference over video without encountering segmentation faults?

same issue here, any progress?

[ycxia]

same issue