rishabh55122/aicoach

Aicoach is a Python library that watches the numbers coming out of a machine learning training run and gives plain-English advice about what might be going wrong. Instead of training any models itself or needing a GPU, it simply looks at metrics you feed it, like training loss, validation loss, learning rate, and accuracy, and points out patterns that usually signal trouble. You create a Coach object and call observe once per epoch with your current loss and other numbers. After enough epochs, calling get advice returns a list of plain-English tips describing anything it noticed. There is also a separate function for checking class imbalance in your dataset labels before training even starts. The library checks for five common training problems. It flags overfitting when validation loss keeps rising while training loss keeps falling. It flags a plateau when a metric has stopped moving in any meaningful way. It checks whether the learning rate looks too high, causing the loss to bounce around, or too low, causing painfully slow progress. It flags class imbalance when one label in your dataset has far more examples than another. Finally, it flags divergence, a more serious case where the loss becomes NaN, infinite, or explodes rapidly, which suppresses the other warnings since divergence matters most. Each check returns either nothing, meaning no issue found, or an advice object containing a plain-English message, a severity level of info, warning, or critical, a short code identifying the issue, and some diagnostic details. All of the detection thresholds, like how many epochs count as a plateau or how much loss growth counts as divergence, are documented and adjustable. Aicoach has no machine learning framework dependencies itself and works alongside whatever training setup you already use. It installs with pip, includes over 280 tests, and is released under the MIT license.