[Test][Topi] Avoid depending on f32 rounding behavior for crop_and_divide tests #13773
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
|
Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.
Generated by tvm-bot |
Lunderberg
force-pushed
the
crop_and_resize_binary_fraction_box
branch
from
fa50f8d to
70aaff7
Compare
|
Failure wasn't reproducible locally with 1k repetitions, each with a different random input. Re-running CI with a fixed seed for random input data. |
|
The errors in the i386 test look like they are due to the same type of rounding issue that this PR was intended to avoid. That is, when an input pixel falls exactly on the edge of a bounding box, whether the |
The `crop_and_resize` operator uses floating-point arithmetic to determine whether an index is within a view-box. This can cause the use of `extrapolation_value` to depend on target-dependent rounding differences. For example, this issue was initially noticed on Vulkan during debugging of apache#13530, and was the result of computing `0.2*223.0 + 0.8*223.0 < 223.0`. If all intermediates are cast to float32, the left-hand side evaluates to `223.00002`. If intermediates are kept at a higher precision, the left-hand side evaluates to `223.0`. The floating-point indexing can't be removed, because the operator must match the API defined by TensorFlow's operator implementation. The TensorFlow documentation for [`CropAndResize`](https://www.tensorflow.org/api_docs/cc/class/tensorflow/ops/crop-and-resize) does not specify behavior in these cases, nor do the current TensorFlow unit tests check cases of rounding error. Since the TensorFlow unit tests only use binary fractions for the `boxes` argument, which largely avoids the rounding issue, this commit updates the TVM unit tests to avoid depending on floating-point precision.
This mimics the example usage of `tf.image.crop_and_resize`, whose API this operator is intended to follow. Using any boxes with edges representable as integer fractions has the potential for the in-bounds check to be impacted by rounding error (e.g. `0.2*x + 0.8*x < x`). Unfortunately, there's no way to remove this possibility altogether without changing the API, such as accepting the box location as an integer fraction, rather than a `float32`, but that would break compatibility. To avoid the risk of a flaky unit test based on the specific random boxes used, the PRNG is seeded prior to generation.
Lunderberg
force-pushed
the
crop_and_resize_binary_fraction_box
branch
from
0166f49 to
aad6f0e
Compare
Lunderberg
changed the title
masahi
approved these changes
Apr 5, 2023
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.
The
crop_and_resizeoperator uses floating-point arithmetic to determine whether an index is within a view-box. This can cause the use ofextrapolation_valueto depend on target-dependent rounding differences.For example, this issue was initially noticed on Vulkan during debugging of #13530, and was the result of computing
0.2*223.0 + 0.8*223.0 < 223.0. If all intermediates are cast to float32, the left-hand side evaluates to223.00002. If intermediates are kept at a higher precision, the left-hand side evaluates to223.0.The floating-point indexing can't be removed, because the operator must match the API defined by TensorFlow's operator implementation. The TensorFlow documentation for
CropAndResizedoes not specify behavior in these cases, nor do the current TensorFlow unit tests check cases of rounding error. Since the TensorFlow unit tests only use binary fractions for theboxesargument, which largely avoids the rounding issue, this commit updates the TVM unit tests to avoid depending on floating-point precision.