I am trying to convert an onnx model generated from Paddle inference to IR.
Openvino version: 2021
Command: python mo.py --input_model "SVTR_Tiny.onnx"
[ ERROR ] The ExpandDims node p2o.Unsqueeze.0 has more than 1 input
[ ERROR ] The ExpandDims node p2o.Unsqueeze.1 has more than 1 input
[ ERROR ] The ExpandDims node p2o.Unsqueeze.2 has more than 1 input
[ ERROR ] The ExpandDims node p2o.Unsqueeze.3 has more than 1 input
[ ERROR ] The ExpandDims node p2o.Unsqueeze.4 has more than 1 input
[ ERROR ] Cannot infer shapes or values for node "p2o.Unsqueeze.4".
[ ERROR ] Wrong number of inputs to the layer p2o.Unsqueeze.4
[ ERROR ]
[ ERROR ] It can happen due to bug in custom shape infer function <function ExpandDims.infer at 0x0000016E198483A8>.
[ ERROR ] Or because the node inputs have incorrect values/shapes.
[ ERROR ] Or because input shapes are incorrect (embedded to the model or passed via --input_shape).
[ ERROR ] Run Model Optimizer with --log_level=DEBUG for more information.
The model is a transformer-based text recognition model, This model primarily uses Grid_Sampler, which I have replaced with a custom function.
Link for the original model: PaddleOCR/algorithm_rec_svtr_en.md at release/2.5 · PaddlePaddle/PaddleOCR (github.com)
Converted PaddleIR and ONNX(using Paddle2onnx): https://drive.google.com/drive/folders/1lnpgDhDp8FqmBfb9UoKssWZavkF0tfJ3?usp=sharing
Can you please provide a solution for this issue.
Thanks for reaching out to us.
I’m able to convert your onnx model into Intermediate Represenation using the latest version of OpenVINO™ toolkit. Any opportunity from your end to try the latest version of OpenVINO™ toolkit? The latest version of OpenVINO™ toolkit 2022.1 provides functional bug fixes, and capability changes for the previous 2021.4.2 LTS release.
Thanks a lot for your reply. Moving to v2022 is a great solution but I am working with RPI 4(64 bit) for which a stable release of v22 is not yet present. Hence, I am currently working with v21.
Is there any specific model structure change that could solve this issue?
We've got a reply from the relevant team. There are a few libraries file changes to resolve the issue and it is available in the OpenVINO™ toolkit 2022.1 release. For a stable release, please stay tuned! Else, you can compile the OpenVINO™ toolkit from GitHub.
Thanks for your question.
This thread will no longer be monitored since we have provided suggestions.
If you need any additional information from Intel, please submit a new question.