commit eeec8957ab4bde841731c807a11eec6a8fbb61f2
author Przemyslaw Szczepaniak <pszczepaniak@google.com> Fri Dec 28 11:47:19 2018 +0000
committer Miao Wang <miaowang@google.com> Thu Jan 24 18:08:03 2019 -0800
tree 99a5f116c7fe62a829abb893ae14d7e57cbe3a1b
parent 415ba4e66f6804cd75a071efdb9bb8917081ef6d

Per-channel quantization support in DataTypeConverter.

- Fixed bug from 425963e97d6ccd7c7fc0fea3ad502ab192a7a14c which was
caused by incorrect packing of negative int8_t values for vts tests.
- Added converters for per-channel quantization for models in
conv2d_v1.2.mod.py and depthwise_conv2d_v1.2.mod.py.
- Added 'hide' flag to SymmPerChannelQuantParams, as it's
needed for conversion, but not actually used in the bias TENSOR_INT32
type.

Bug: 119255406
Test: NeuralNetworksTest_static and VtsHalNeuralnetworksV1_2TargetTest
Change-Id: I9689240e06ba423d8bea6353fae3d331d7566279
Merged-In: I9689240e06ba423d8bea6353fae3d331d7566279
(cherry picked from commit da8b3508227209b00e6815e8b5874aa2db76f8b1)