博客
Reproducing a paper without pretending it was smooth
Notes on the difference between understanding a result and building a reliable reproduction.
A lot of reproduction notes become retroactive victory laps. The code runs once, the plot looks roughly correct, and the writeup quietly compresses a week of confusion into three neat bullet points.
This site is meant to keep more of the real texture.
What usually breaks first
The first problem is almost never the core model idea. It is usually one of the surrounding assumptions.
Data conventions
A dataset might claim to be standardized while still hiding an important camera convention or normalization choice.
Evaluation details
Metrics often depend on small filtering rules that are not highlighted in the abstract or headline table.
Training defaults
Even learning-rate schedules, augmentation order, or batch construction can move results enough to waste days.
A more useful reproduction note
A useful note should record:
- what matched the paper quickly,
- what stayed ambiguous,
- what was inferred from code,
- and what still feels uncertain after the run succeeds.
That kind of honesty is more reusable than a perfect-looking summary.