Add tile and repeat_values procedures
#648
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
These procedures let you repeat the values of a Tensor multiple times. This functionality exists both in numpy (`repeat`) and in Matlab (`repelem`). A different name was chosen here to avoid confusion with nim's `repeat` function, which behaves differently (it repeats the whole input sequence, like numpy's `tile` or Matlab's `repmat` functions), and to make the name more self explanatory. There are two versions of this procedure (with multiple overloads): - One that repeats all values the same amount of times over a given axis. - One that repeats each value a different amount of times, but returns a rank-1 tensor. Note that the second one is implemented as 2 procs with different argument types (openArray[int] and Tensor[int]). I measured the performance using the timeit library. The results show that the performance is comparable to numpy's `repeat` function. In particular, a small example which takes numpy's `repeat` ~2-3 usec per iteration, takes ~4 usec in --d:release mode, and ~1-2 usec in --d:danger mode.
This procedure lets you construct a new tensor by repeating the input tensor a number of times on one or more axes. This is similar to numpy's `tile` and Matlab's `repmat` functions. I measured the performance using the `timeit` library. The results show that the performance is comparable to (but not as good as) numpy's `tile`. In particular, a small example which takes numpy's `tile` ~3-4 usec per iteration, takes ~8-9 usec in --d:release mode, and ~5-6 usec in --d:danger mode. I believe that the performance could be improved further by preallocating the result Tensor before the tiling operation. The current implementation is not as efficient as it could be because it is based on calling `concat` multiple times, which requires at least as many tensor allocations (of increasing size).
Vindaar
reviewed
May 7, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Add the following new procedures:
repeat_values: Procedures that let you repeat the values of a Tensor multiple times. This functionality exists both in numpy (repeat) and in Matlab (repelem).tile: Procedure that lets you construct a new tensor by repeating the input tensor a number of times on one or more axes. This functionality exists both in numpy (tile) and in Matlab (repmat).I didn't follow numpy's naming convention for
repeat_valuesto avoid a confusing with nim'ssequtils.repeat, which does not repeat individual values but the whole sequence (liketiledoes).I measured the performance of these functions and it is comparable to numpy's.
repeat_valuesis faster than numpy in --d:danger mode and a bit slower in --d:release mode.tileis slower in both cases (but the difference is not too large).tile's implementation could be improved (to avoid intermediate tensor allocations) in the future.