Project idea · ungraded

Training Diagnostics Notebook

Build a reusable notebook (or script) of experiments that explain why a model trains well or badly: a learning-rate finder, loss curves for good and bad settings, an overfitting demonstration and gradient-norm plots. Use it on your digit recogniser or any other model.

You will use

Milestones

Build the smallest working version first. Ticks are saved in this browser only.

01Write a function that trains a model for a few epochs and returns train and validation losses.
02Learning-rate finder: increase the learning rate exponentially over one epoch and plot loss against learning rate.
03Show three curves side by side: learning rate too low, about right and too high.
04Demonstrate overfitting with a large model on 200 examples, then fix it with dropout or weight decay.
05Log the gradient norm of every layer per step and show the effect of gradient clipping.

Stretch goals

Add a learning-rate schedule (cosine or one-cycle) and compare final accuracy.
Track experiments in a CSV file and compare runs with different seeds.
Use TensorBoard to view the curves live.

Data

Any small classification dataset works: sklearn.datasets.load_digits() (bundled) or synthetic data from sklearn.datasets.make_moons(n_samples=1000, noise=0.25, random_state=0).

Run it yourself

pip install torch scikit-learn matplotlib jupyter
jupyter notebook