Digit recognition is a simple task of recognising digits from 0 to 9. A CNN is trained on the MNIST dataset which gave a good accuracy of around 98% on the test set. Choosing the proper model and employing suitable regularisation technique, the model also performed well during inference with an accuracy of close to 98%. The whole model is then converted to ONNX format and run on CPU with ONNX Runtime.
- Train the model using the 'MNIST classification using ConvNets' notebook. Save your models by specifying the path.
- Select the model that you want to convert to ONNX format. Specify the path to your saved model in the MNIST_ONNX notebook. Save your model in the onnx format (say, model.onnx).
- Load the model.onnx file and run an inference session on your CPU using ONNX Runtime. Make sure to specify the path in mnist_onnx.py
You can directly download the model in onnx format here https://drive.google.com/file/d/16y9_jh7H7La0M1vuQ3XcpdpYu-ibJbJB/view?usp=sharing.
