the repository is about the conversion of CRNN model, which is widely used for text recognition. the CRNN model is converted from PyTorch to TensorRT via ONNX
To run this project, some packages with special version will be installed:
| Package Name | Version | Description |
|---|---|---|
| PyTorch | 1.3.1 | Deep learning tool |
| onnx | 1.6.0 | Conversion medium for different frames |
| onnxruntime | 1.3.0 | Do the inference of an onnx file |
| TensorRT | 7.0.0.11 | Plan to Optimize the inference efficiency on Nvidia GPUs |
| pillow | 7.1.2 | Image processing |
| opencv-python | 4.2.0.34 | Image Processing |
| pycuda | 2019.1.2 | Help to allocate memory on cpu and gpu when using TensorRT for inference |
| If other packages are to installed, please follow the information in CMD |
The project is greatly helped by the project of CRNN-Pytorch Thanks for meijieru's contributions!
git clone https://github.com/YIYANGCAI/CRNN-Pytorch2TensorRT-via-ONNX
cd CRNN-Pytorch2TensorRT-via-ONNX
Find the pretrained model from meijieru's project mentioned above Copy the pth model into ./data in the project
python pytorch2onnx.py
Then you can find onnx model ./new_crnn.onnx is created. You can test the input and output of pth and onnx model by doing their inferences.
python onnx2tensorrt.py
I have test the inference time on TITAN-RTX and the inference time can be fast as 3 ms, however, the inference by INT8 is not applied in this project, I will do this later.