-
Notifications
You must be signed in to change notification settings - Fork 333
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
(OD input pipeline Step 3/4) Add RandomAspectRatio layer #1054
Conversation
25dc0b0
to
df3782b
Compare
ok @ianstenbit @tanzhenyu @qlzh727 I believe this is prepared! |
/gcbrun |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for the PR!
Implementation wise looks good to me.
Can you add a description on why we need this? (AFAIK, the OD models are expected to train well on JitteredResize layer?)
Sure thing, I updated a bit but JitteredResize doesn't touch aspect ratio whereas this ONLY touches aspect ratio. |
…1054) * Add ragged image augmentations * Add ragged image augmentation support * Add ragged image test * Reformat and fix * Batchwise support * Add resize with pad only support * Remove pad only * Remove pad only * Add JitteredResize layer * Add jittered resize to ragged image test * Add jittered resize to serialization test * Add ragged augmentation layers * Add ragged image test * checkout origin/master -- files * Begin adding unit tests * Update converter for bounding boxes * Add mixed precision test * Add serialization test * Update ragged crop * Fix lint * Fix * Seed fix * Add serialization test
This PR adds a RandomAspectRatio layer that takes a Dense image tensor and always outputs a ragged tensor of aspect-ratio distorted images.
JitteredResize does not tweak aspect ratios, so we can use this layer to fill all aspect ratio distortion based needs across tasks.