Failed to convert pre-trained action recognition model to ONNX format

[elainewangprc]

I'd like to convert the pre-trained action recognition model to OpenVINO IR files. So I followed the instructions at https://github.com/opencv/openvino_training_extensions/tree/develop/pytorch_toolkit/action_recognition. In the first step that convert the model to ONNX, I saw below error message:

Traceback (most recent call last): File "main.py", line 385, in main() File "main.py", line 330, in main export_onnx(args, model, args.onnx) File "main.py", line 36, in export_onnx model = model.module File "/usr/local/lib/python3.5/dist-packages/torch/nn/modules/module.py", line 539, in getattr type(self).name, name)) AttributeError: 'VideoTransformer' object has no attribute 'module'

Steps to Reproduce (for bugs)

1. git clone https://github.com/opencv/openvino_training_extensions.git
2. https://download.01.org/opencv/openvino_training_extensions/models/action_recognition/resnet_34_vtn_rgbd_kinetics.pth
3. Under directory openvino_training_extensions/pytorch_toolkit/action_recognition, run python3 main.py --model resnet34_vtn_rgbdiff --clip-size 16 --st 2 --pretrain-path ~/resnet_34_vtn_rgbd_kinetics.pth --onnx resnet34_vtn_rgbd.onnx

I want to convert the pre-trained decoder and encoder model to OpenVINO IR files. Like https://github.com/opencv/open_model_zoo/blob/master/intel_models/action-recognition-0001-encoder/description/action-recognition-0001-encoder.md and https://github.com/opencv/open_model_zoo/blob/master/intel_models/action-recognition-0001-decoder/

Your Environment

• Linux Ubuntu 16.04:
• TensorFlow installed from binary
• TensorFlow version (use command below): v1.13.1-0-g6612da8951

  python -c "import tensorflow as tf; print(tf.GIT_VERSION, tf.VERSION)"

• Python version: Python 3.5.2
• torch version: 1.1.0
• CUDA/cuDNN version: not available
• GPU model and memory: Intel Coffeelake GT3

[AlexanderDokuchaev]

Hi @elainewangprc Fixed some problem. Try use follow command:

python3 main.py --model resnet34_vtn_rgbdiff --clip-size 16 --st 2 --pretrain-path ~/resnet_34_vtn_rgbd_kinetics.pth --onnx resnet34_vtn_rgbd.onnx --no-cuda --no-layer-norm

--no-cuda - to run in environment without cuda --no-layer-norm - to ignore LayerNormalization that can not be converted to ONNX, but this may lead to decrease accuracy of final model.

[elainewangprc]

Hi Alex, Thank you for fixing the issue! Now I'm able to convert the pre-trained model into ONNX format.