Diagnosing from the record alone
Why it matters Often you cannot rerun anything: the run was someone else's, or it cost a week. All you have is what was logged. Two recorded experiments from this course hold real failures.
In gpt-width-lr, a setting is broken by its learning rate if its final held-out loss, averaged over the three recorded seeds, is more than 0.5 worse than the best setting of the same width. In cnn-depth-shortcuts, a single run is stuck if its test accuracy is exactly the same at every epoch: it is answering one class for everything (there are four shape classes, so about 0.25).
Predict before you look: which GPT runs are broken, and which CNN is stuck?
Fill broken with (width, lr) pairs and stuck with (depth, shortcuts, seed) triples, computed from the records. exp.seeds(d=..., lr=...) gives the three runs of one setting.
Spotted a problem in this lesson? Report it · Ask about this lesson
This exercise opens soon
Finish the lesson before it and this one unlocks. The whole course is free, no account needed, and you can read this lesson right now.
This browser is blocking cookies for learn.floati.life, so lessons you finished here cannot open in order. Allow cookies for this site and reload.