diff options
| author | Lev Proleev <levp@google.com> | 2019-12-17 13:19:18 +0000 |
|---|---|---|
| committer | Lev Proleev <levp@google.com> | 2020-01-02 18:35:20 +0000 |
| commit | 21d5907e55f6ed8a3e08c240efb9cf5d4a644fd8 (patch) | |
| tree | 6dc92827ee6ea2f9a1ab4b6b251168f9b26257cc /nn/runtime/test/specs/V1_3 | |
| parent | 63e4f723ead3b11de38b38896a0a4b9bdb9cec45 (diff) | |
| download | ml-21d5907e55f6ed8a3e08c240efb9cf5d4a644fd8.tar.gz | |
Add exceptions to the signed quantization coupling tests
There are several ops that are exceptions to the general quant8 types
coupling:
HASHTABLE_LOOKUP -- due to legacy reasons uses TENSOR_QUANT8_ASYMM
tensor as if it was TENSOR_BOOL. It doesn't make sense to have
coupling in this case.
LSH_PROJECTION -- hashes an input tensor treating it as raw bytes. We
can't expect same results for coupled inputs.
PAD_V2 -- pad_value is set using int32 scalar, so coupling produces a
wrong result.
CAST -- converts tensors without taking into account input's scale and
zero point. Coupled models shouldn't produce same results.
QUANTIZED_16BIT_LSTM -- the op is made for a specific use case,
supporting signed quantization is not worth the compications.
Bug: 143935412
Bug: 143934188
Bug: 143935353
Bug: 143934467
Bug: 143934184
Test: quantization coupling tests in CTS and VTS
Change-Id: Ic2b2336cbc12c8e75579d088273c13b7d01a59d3
Diffstat (limited to 'nn/runtime/test/specs/V1_3')
0 files changed, 0 insertions, 0 deletions