diff options
author | Gunes Bayir <gunes.bayir@arm.com> | 2023-02-10 11:07:37 +0000 |
---|---|---|
committer | Gunes Bayir <gunes.bayir@arm.com> | 2023-02-10 14:29:57 +0000 |
commit | f4230aafcc106305d4da9798bd5211fe7a7809ea (patch) | |
tree | fa1b7f1cd98f730206106c4020f1fe62a62224f6 /src/graph | |
parent | 06db85ed4ca0b3f4ee525787cb17a814a89006ce (diff) | |
download | ComputeLibrary-f4230aafcc106305d4da9798bd5211fe7a7809ea.tar.gz |
Fix DeconvolutionLayer tolerance issues in FP16 tests
This patch increases the tolerance value used for FP16 tests in Neon(TM) backend. The tolerance number means 0.01f means it is ok to have 1% mismatch in the resulting tensor between the reference and the target. The value adopts a slightly stricter threshold compared to ConvolutionLayer (which is currently at 7%). This increase makes sense because Deconvolution layer uses convolution under the hood.
Resolves: COMPMID-5841
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Change-Id: Ie0ebf5cce1e9753dc641a947d84128dd6da402d4
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9120
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-by: Sang Won Ha
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/graph')
0 files changed, 0 insertions, 0 deletions