[Data] Modify ImageDatasource
to use Image.BILINEAR
as the default image resampling filter.
#43484
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Outdated suggestions cannot be applied.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.
Why are these changes needed?
Ray Data's current
ImageDatasource
uses PIL'sresize
function as the image processing backend, which by default uses theImage.BICUBIC
resampling filter. In practice, we found this is around 20% slower than using theImage.BILINEAR
filter, which is the default option in torch vision.Here are some benchmark results:
This is a unit test where we load one single image and resize it, repeated 10000 times. Note the time difference between the current (Image.BICUBIC) filter vs. proposed (Image.BILINEAR).
This is an end-to-end benchmark modified from @stephanie-wang 's image loader microbenchmark. Here, we demonstrate the actual effect on the ray data image loading pipeline. Using the
BILINEAR
filter as default leads to a ~22% increase in throughput.If a user still wishes to use the
BICUBIC
filter, this is still easily achievable by applying a UDFresize_fn
after the images have been read. For instance, we can choosecv2
's INTER_CUBIC or PIL's original resize (this requires the use of PIL'sfromarray
though, as the output ofread_image
are numpy arrays; this conversion lowers throughput).Related issue number
Checks
git commit -s
) in this PR.scripts/format.sh
to lint the changes in this PR.method in Tune, I've added it in
doc/source/tune/api/
under thecorresponding
.rst
file.