Torch.manual_seed(3407) is all you need: On the influence of random seeds in deep learning architectures for computer vision
 is all you need.png)
Context
- It is a common practice in the machine learning community to report results from a single run (a single random seed). This many compromise the statistical validity of the presented results.
Takeaways
-
The random seed has an important influence on the final accuracy of computer vision models.
-
The accuracy difference between a “bad seed” and a “good seed” was around 1% on the studied benchmarks. But the gap may be even higher if more seeds were evaluated.
-
Dependending on the benchmark, improvements greater than 0.5% on the current state-of-art method are generally considered relevant by the computer vision community.
-
It is recommended that studies evaluate many runs with different random seeds. The accuracy should be published in terms of average, standard deviation, and minimum/maximum values.