Hacker News Reader

Comment by abidlabs | Hacker News Reader

Asking for clarification so that we can improve it -- is it that the UI is not usable, or is that the model predicts things inaccurately? If it's the latter, it might say more about the model (a convolutional network with high accuracy on the MNIST dataset) than Gradio

View on HN

Probably not even the model, but the preprocessing - MNIST-trained models expect all the digits to have the same preprocessing that MNIST had (because that's all they saw), so if you apply a MNIST-optimized model directly to user-generated input without e.g. resizing and re-centering the images (or, alternatively, a "robustized" model trained with various data augmentations), then you're going to have horrible results.

The latter - the results are astoundingly bad. You should remove that demo ASAP.

I just drew a vertical line, slightly off center to the left, almost perfectly straight, starting about x=0.1 and proceeding to x=0.95. I expected it would show as a digit 1. Page said it was a 6. Tried a 1 with a serif on top. Page said it was an 8. Added a serif on the bottom. Still an 8.

Heck, if you mouse over the area without drawing, the model is torn between 1, 9 and 3.

Then I went to the next demo and uploaded a picture of a Christmas choir in a church. Model said "altar" and "vestments". Highlight that one.

The UI is fine. The model seems to struggle with me drawing 6 on desktop.

Reading "Sketch Recognition" I was expecting something akin to Google draw which recognises objects like cars and trees. The digit classifier is of course fine, but I felt a bit disappointed when I realised it "only" does numbers. I also think it's the least impressive of the demos you have.

Reply on news.ycombinator.com