Just adapted and integrated a numeric detector into my website. Unfortunately there were a few issues that I had to overcome so it took longer than desired.
Creating a separate test worked wonders, starting with an all-in-one HTML file. This went well with the rawhtml tag, provided that I cut some unnecesary things out.
Now I can start to improve the model although it’s already at 99% accuracy. 😅️
Looking at this post (Going beyond 99%) I learned that it’s an exercise from kaggle and that the script trains for 30 epochs- It’s also using Tensorflow. So much for my idea of using it directly but changing the number of epochs is easy enough.
And so the Model reached an accuracy of 99.18%: Test set: Average los1s: 0.0262, Accuracy: 9918/10000 (99%).
Got a errors on the page. To fix it, add opset_version=9 to your conversion function (torch.onnx.export).
From my tessts this is far worse than the example onnx model file. Detecting a 4 appears to be beyond it.