Add docs for custom loss layer #107
Conversation
Codecov ReportBase: 73.55% // Head: 73.28% // Decreases project coverage by
Additional details and impacted files@@ Coverage Diff @@
## main #107 +/- ##
==========================================
- Coverage 73.55% 73.28% -0.28%
==========================================
Files 14 14
Lines 2590 2590
==========================================
- Hits 1905 1898 -7
- Misses 685 692 +7
Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here. ☔ View full report at Codecov. |
There was a problem hiding this comment.
This looks great, thanks!
Of course, I'd note that adding @turbo to the loss and gradient should improve performance.
The pointer pu is in case the layer wants to allocate temporaries. As you noted here
As the other loss functions do this, we should define some functions to say that we don't want any preallocated temporary arrays
If you did want temporaries, you'd need to specify their size in layer_output_size and forward_layer_output_size.
Then, the memory the pointer pu points to would have that much extra space.
The recommended approach from there is to use a PtrArray for an array-like API, and then bump the pointer if the PtrArray is supposed to live beyond this function.
SimpleChains probably needs a better concept of lifetimes, to avoid overallocating memory.
Links to #106. This is a start at creating some documentation around implementing more loss functions for external users of the package, and also other open-source devs that want to contribute a varied set of loss functions.
This might still need some work, but it is enough for solving my initial issue.