Do I need a GPU to build an image classifier?

No. The digits dataset holds 1,797 tiny images, and the whole training run finishes in seconds on a normal laptop CPU. A GPU only starts to matter with large photo datasets and deep neural networks. For this tutorial, and for most first classifiers, a CPU is more than enough, and that keeps the build free and simple.

Why does this tutorial use 8 by 8 images instead of real photos?

Small images keep the tutorial fast, free, and honest. Each digit is 64 numbers, so training takes seconds and the results are easy to inspect. Real photos carry thousands of pixels and need convolutional networks, which are a bigger build. The workflow is the same either way: turn pictures into numbers, split the data, train, and measure.

Can I use my own handwritten digits with this code?

Yes, with one condition. Draw each digit dark on a white background, save it as a PNG, and pass the path to predict.py. The script resizes an image to 8 by 8 and rescales the pixels, so size does not matter. Digits that look like the training set, centred and clean, score best, while tilted or faint writing confuses it.

What is the difference between accuracy and a classification report?

Accuracy is one number for the whole test set, and ours is 98.9 percent. A classification report breaks that number down per digit, showing precision and recall for each one. Our report shows digit 8 is weakest, with recall of 0.953, a fact accuracy alone hides. Use accuracy for a quick check, and the report to find what to fix.

How can I push accuracy past 98.9 percent?

Start with the data, not the model. More training images, especially more 8s, would fix the weakest class directly. Tuning the SVM gamma and C values with a grid search is next, and the digits set is small enough that a search runs in seconds. Beyond that, gains shrink fast, and a bigger model would mostly memorise the same images.

Build an Image Classifier with scikit-learn