Lev-etd / intel_course_project

🎓Multimodal Emotion Recognition with OpenVINO™

Demo

Model convertation

⌚ Convert and optimize

To convert PyTorch model to ONNX format, use the code below:

PATH = r'models\enet_b0_8\enet_b0_8.pt' # path to your model

feature_extractor_model = torch.load(PATH) # load model
feature_extractor_model.eval() # set the model in evaluation mode

dummy_input = torch.randn(1, 3, 224, 224).cuda() # Create dummy input for the model. It will be used to run the model inside export function.

torch.onnx.export(feature_extractor_model, (dummy_input, ), 'enet_b0_8.onnx') # call the export function

To convert ONNX model to IR format in cmd use a command like:

mo --input_model your_ONNX_model -o output_directory_for_IR_model --data_type FP32_or_FP16_or_FP8

In our case we used the following commands:

FP32:

mo --input_model D:\Users\amira\openvino_env\Lib\site-packages\openvino\model_zoo\models\group_project\enet_b0_8\enet_b0_8.onnx -o D:\Users\amira\openvino_env\Lib\site-packages\openvino\model_zoo\models\group_project\enet_b0_8 --data_type FP32

FP16:

mo --input_model D:\Users\amira\openvino_env\Lib\site-packages\openvino\model_zoo\models\group_project\enet_b0_8\enet_b0_8.onnx -o D:\Users\amira\openvino_env\Lib\site-packages\openvino\model_zoo\models\group_project\enet_b0_8 --data_type FP16

Notes

Models to evaluate and the code to run them were borrowed from this repository.

Please be sure that EfficientNet models for PyTorch are based on old timm 0.4.5 package, so that exactly this version should be installed by the following command:

pip install timm==0.4.5