[nnc] Fix tensorflow lite import of depthwise conv (#985)
authorDmitry Mozolev/AI Tools Lab /SRR/Engineer/삼성전자 <d.mozolev@samsung.com>
Tue, 14 Aug 2018 16:45:28 +0000 (19:45 +0300)
committerSergey Vostokov/AI Tools Lab /SRR/Staff Engineer/삼성전자 <s.vostokov@samsung.com>
Tue, 14 Aug 2018 16:45:28 +0000 (19:45 +0300)
commit1e2308c2b60ff8504a0d1028d8dbf942b8709ea4
tree8996cfc452ac8a731d78695117061622fec0f3f7
parent674a80d7f28fc7ab259ce4fb7b7df047818a8727
[nnc] Fix tensorflow lite import of depthwise conv (#985)

Depthwise convolution in Tensorflow Lite has channel multiplier
as its kernel's first dimension, while current Model IR
assumes it is the last dimension. This commit fixes this issue
by appropriately transposing the kernel during import.

Signed-off-by: Dmitry Mozolev <d.mozolev@samsung.com>
contrib/nnc/libs/frontend/tflite/src/tflite_ir_visitor.cpp