5,337
edits
Debugging ML Models: Difference between revisions
→NaNs and Infs
| Line 68: | Line 68: | ||
* The derivatives of \(\arccsc(x)\) and \(\arcsec(x)\) are Inf near 0 and 1. | * The derivatives of \(\arccsc(x)\) and \(\arcsec(x)\) are Inf near 0 and 1. | ||
If you must use one of these functions near the unstable points, I suggest performing a linear or | If you must use one of these functions near the unstable points, I suggest performing a linear or taylor series approximation near 0 and 1. | ||
Other ways to mitigate NaNs and Infs are: | Other ways to mitigate NaNs and Infs are: | ||