[Error] Action Recognition : Convert ONNX to IR file

[Nhoodangggggg]

I tried to to convert the ONNX models to OpenVINO format. My model is resnet34_vtn_rgb_decoder.onnx and converted by the following commands

mo.py \
--data_type FP16 \
--input_model ./pretrained_models/resnet34_vtn_rgb_decoder.onnx \
--input_shape [1,16,512]

But it failed and I got this message :

[ ERROR ] Cannot infer shapes or values for node "258/Broadcast".
[ ERROR ] 'numpy.float64' object cannot be interpreted as an integer
[ ERROR ]
[ ERROR ] It can happen due to bug in custom shape infer function <function Broadcast.infer at 0x10cbe24d0>.
[ 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.
[ ERROR ] 'numpy.float64' object cannot be interpreted as an integer
Stopped shape/value propagation at "258/Broadcast" node.
For more information please refer to Model Optimizer FAQ (https://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_Model_Optimizer_FAQ.html), question #38.

I downloaded the pretrain model from https://github.com/opencv/openvino_training_extensions/tree/develop/pytorch_toolkit/action_recognition

and then I covert to ONNX on google colab by the following command

!python "/content/openvino_training_extensions/pytorch_toolkit/action_recognition/main.py"
--clip-size 16 --st 2 --model resnet34_vtn_decoder \
--pretrain-path ./resnet_34_vtn_rgb_kinetics.pth
--onnx ./resnet_34_vtn_rgb_kinetics_onnx.onnx
--no-layer-norm --no-cuda

Could you tell me how to solve this problem ?

[minhhoangbui]

I have a same issue. Do you use which version of OpenVino to convert onnx file?

[Nhoodangggggg]

I used Openvino 2019_R3.1 and also 2020.1

[minhhoangbui]

I use 2019R2 and 2020R1, I wonder they use which version :(