Add RunInference.with_errors() API #14
Draft
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
added 8 commits
July 29, 2020 19:19
tfx_bsl/beam/run_inference.py
Outdated
| return beam.pvalue.TaggedOutput(_get_operation_type(batch[0]), batch) | ||
| else: | ||
| try: | ||
| return beam.pvalue.TaggedOutput(_get_operation_type(batch[0]), batch) |
There was a problem hiding this comment.
Will this throw error? It seems _get_operation_type will return str or unicode
There was a problem hiding this comment.
_get_operation_type has return value Text (str/unicode) and beam.pvalue.TaggedOutput accepts either str or unicode for a tag. See: https://github.com/apache/beam/blob/master/sdks/python/apache_beam/pvalue.py#L328-L343
added 13 commits
August 6, 2020 01:01
Benchmarks showed that TagByOperation was a performance bottleneck* as it requires disc access per query batch. To mitigate this I implemented operation caching inside the DoFn. For readability, I also renamed this operation to "SplitByOperation" as that more accurately describes its purpose. On a dataset with 1m examples, TagByOperation took ~25% of the total wall time. After implementing caching, this was reduced to ~2%.
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Outdated suggestions cannot be applied.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.
with_errors()APIThis PR introduces the
RunInference(...).with_errors()API which allows users to catch runtime errors as a separate PCollection stream.By default, runtime errors (for example invalid model specs or invalid examples) are thrown which can crash a pipeline. If this is not desirable, users can use
.with_errors()to catch runtime errors:The error output stream has the type:
Tuple[Exception, Any]and contains both the original error and whatever object is relevant to the error.Note: when runtime errors are allowed to be raised, they are raised from their original location (e.g. inside a nested PTransform) which makes debugging easier.
Internal details
RunInferenceImplis now a class (rather than a function with @beam.ptransform_fn). This enables us to add awith_errorsmethod that can take effect in theexpandmethod.RunInferenceCorereturns a dict containing{'predictions': ..., 'errors': ...}and takes an additionalcatch_errors: bool = Falseparameter which indicates whether to catch or allow runtime errors.Added the
_ParDoExceptionWrapperutility which runsbeam.ParDoon a providedbeam.DoFnand optionally catches exceptions raised in theprocess()method.Operation wrapper transforms (e.g.
_Classify,_Regress, ...) accept an additionalcatch_errors: bool = Falseparameter and return a dict containing{'predictions': ..., 'errors': ...}Dependencies
This PR depends on #13