Error loading MixNet model

[RicardBorras]

Hello,

when loading a MixNet model converted from ONNX using model converter, I get an error saying "Convolution initialization. Group splitted mode are used together with direct group specification" when model is loaded. Error is reproduced in 2020.3 and 2020.4 releases in Ubuntu Linux 18.04

ONNX model is converted from a pretrained Pytorch model from https://github.com/rwightman/gen-efficientnet-pytorch.

and converted using:

python mo_onnx.py --input_model=<ONNX_FILE> --output_dir=<OUTPUT_FOLDER>

Find attached ONNX file and converted models

Any help?

Thanks,

Ricard

[Munesh_Intel]

Hi Ricard,

EfficientNet B0, EfficientNet B5 and EfficientNet B7 are available as public models in Open Model Zoo. For your case, I’ve validated EfficientNet B0 using OpenVINO 2020.4 in Ubuntu 18.04 LTS, and the model is working fine with Image Classification Python Sample.

I suggest you try the following:

I converted the PyTorch model to ONNX using the following command available from https://github.com/rwightman/gen-efficientnet-pytorch

python3 onnx_export.py --model efficientnet_b0 ./efficientnet-b0.onnx

The Model Optimizer command that I used is as follows:

python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo_onnx.py --input_model /home/username/Desktop/efficientnet-b0.onnx --reverse_input_channels --input_shape [1,3,224,224]

The Model Optimizer conversion parameters for efficientnet-b0-pytorch is available at

https://github.com/openvinotoolkit/open_model_zoo/blob/master/models/public/efficientnet-b0-pytorch/model.yml#L64

Similarly, Model Optimizer conversion parameters for efficientnet-b5-pytorch and efficientnet-b7-pytorch are available at

https://github.com/openvinotoolkit/open_model_zoo/blob/master/models/public/efficientnet-b5-pytorch/model.yml#L70

https://github.com/openvinotoolkit/open_model_zoo/blob/master/models/public/efficientnet-b7-pytorch/model.yml#L69

Regards,

Munesh

[RicardBorras]

Hi,

I haven't tryied EfficientNET models from this repo, I am trying to use MIXNET models (following same steps that you supplied for EfficientNET). Can you verify if your approach works for Mixnet models?

thanks,

Ricard 

[Munesh_Intel · Accepted solution]

Hi Ricard,

I’ve validated your MixNet model, and am seeing the same errors as well. This is due to MixNet not being supported officially by OpenVINO.

Details about supported PyTorch models are available at the following page:

https://docs.openvinotoolkit.org/2020.4/openvino_docs_MO_DG_prepare_model_convert_model_Convert_Model_From_ONNX.html#supported_pytorch_models_via_onnx_conversion

Regards,

Munesh

[Munesh_Intel]

Hi Ricard,

This thread will no longer be monitored since we have provided a solution. If you need any additional information from Intel, please submit a new question.

Regards,

Munesh