[IE][nGraph] resolve division by zero for FP16 #8676
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
pavel-esir
force-pushed
the
resolve_division_to_zero
branch
from
0ee15e5
to
c4aa420
Compare
pavel-esir
requested review from
a team,
sadolini,
achetver,
yekruglov,
rkazants and
GlebKazantaev
and removed request for
a team
11 tasks
sadolini
approved these changes
Nov 30, 2021
rkazants
reviewed
Dec 1, 2021
...nsformations/include/transformations/common_optimizations/division_to_zero_fp16_resolver.hpp
Outdated
Show resolved
Hide resolved
rkazants
reviewed
Dec 1, 2021
.../transformations/src/transformations/common_optimizations/division_to_zero_fp16_resolver.cpp
Outdated
Show resolved
Hide resolved
rkazants
reviewed
Dec 1, 2021
.../transformations/src/transformations/common_optimizations/division_to_zero_fp16_resolver.cpp
Outdated
Show resolved
Hide resolved
rkazants
reviewed
Dec 1, 2021
...ne/tests/functional/inference_engine/transformations/division_to_zero_fp16_resolver_test.cpp
Outdated
Show resolved
Hide resolved
achetver
approved these changes
Dec 2, 2021
pavel-esir
changed the title
…e minor corrections
rkazants
approved these changes
Dec 2, 2021
GlebKazantaev
suggested changes
Dec 6, 2021
.../transformations/src/transformations/common_optimizations/division_by_zero_fp16_resolver.cpp
Outdated
Show resolved
Hide resolved
.../transformations/src/transformations/common_optimizations/division_by_zero_fp16_resolver.cpp
Outdated
Show resolved
Hide resolved
|
|
||
| /** | ||
| * @ingroup ie_transformation_common_api | ||
| * @brief: clamps eps into fp16 minimal normalized value in input_1/Maximum(input_2, eps); input_1/Add(input_2, eps); |
There was a problem hiding this comment.
nit: looks like this is not al patterns that are supported. Better to keep patterns description in cpp.
GlebKazantaev
approved these changes
Dec 7, 2021
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.
Root cause analysis: In some models patterns
x/max(y, eps)andx/(y + eps)are used to prevent division by zero. But usually in FP32 networks eps is such small (e.g. 1e-9, 1e-12, ...) that after casting to FP16 it's collapsed to zero. This can lead to division by zero ifxandyhappen to be also zero.Solution: To prevent that in such patterns we change eps to FP16 smallest normalized value (0.0000610.., details in Half_precision_examples) in transformation
DivisionByZeroFP16Resolver.Tickets:
Ticket: CVS-66045, CVS-65735
Code:
Validation: