Experiment tracking logs every training run and its results.
You record hyperparameters, metrics, artefacts and the environment.
This lets you compare dozens of runs without a spreadsheet.
Weights and Biases and MLflow are the common choices.
Untracked experiments are effectively lost work.
Experiment tracking logs every training run and everything about it.
Hyperparameters, metrics, artefacts, code version and environment.
Without it you are relying on memory and a spreadsheet, and both fail by the twentieth run.
The question it answers is which run produced this result and what was different.
That is unanswerable a week later if nothing was recorded.
Log at the point of running rather than afterwards, since afterwards never happens.
One line at the start and one at the end of a run is enough to begin.
Weights and Biases and MLflow are the common choices.
Both give a dashboard comparing runs and plotting metrics against each other.
Log the losing runs too, because knowing what failed prevents repeating it.
Record the random seed, or a good run cannot be recreated.
Record the data version alongside the code version.
Tag runs with the question they were meant to answer.
Untracked experiments are effectively lost work, however good the result was.
The habit costs a few minutes and repays it the first time someone asks how you got a number.
One question on this card. Every option explains itself once you pick, right or wrong.
22 concepts in this module, then a graded test with every answer explained.