[GNA] Depth-wise separable convolution support #7281
Conversation
sirzabek
force-pushed
the
gna_dwsc
branch
2 times, most recently
from
05e48e2 to
2d6cdaa
Compare
sirzabek
requested review from
a team,
kbruniec,
nkogteva,
evkotov and
elilobanova
inference-engine/src/gna_plugin/transformations/convert_dwsc_to_scaleshifts.cpp
Outdated
Show resolved
Hide resolved
inference-engine/src/gna_plugin/transformations/convert_dwsc_to_scaleshifts.cpp
Outdated
Show resolved
Hide resolved
inference-engine/src/gna_plugin/transformations/convert_dwsc_to_scaleshifts.cpp
Show resolved
Hide resolved
| #include "transformations/convert_dwsc_to_scaleshifts.hpp" | ||
| #include "common_test_utils/ngraph_test_utils.hpp" | ||
| #include <ngraph/function.hpp> | ||
| #include <ngraph/opsets/opset7.hpp> |
There was a problem hiding this comment.
is it ok to use opset7 directelly?
isn't there sth like "opset_newest" ?
There was a problem hiding this comment.
Suggestion by ngraph developers was to use opset7, so I believe this is the way to go for now.
| auto bias_size = shape_size(add_const->get_shape()); | ||
| auto conv_filter_count = conv->get_output_shape(0)[1]; | ||
| if (bias_size == conv_filter_count) | ||
| return add_const; |
There was a problem hiding this comment.
Consider to make more strict const shape comparison,
i.e.: not only compare total shape size but check that each dimension matches expected one by one.
There was a problem hiding this comment.
It's assumed that bias const is flat and should be broadcasted, but I'll take notice of this suggestion in next PR.
| @@ -72,4 +73,29 @@ std::shared_ptr<ngraph::opset7::StridedSlice> FlatCrop(ngraph::Output<ngraph::No | |||
| std::vector<int64_t>{1, 0}); // end mask | |||
| } | |||
|
|
|||
| std::shared_ptr<ngraph::Node> VerifyBiasGetConst(std::shared_ptr<ngraph::Node> conv, std::shared_ptr<ngraph::Node> bias) { | |||
There was a problem hiding this comment.
Consider not to pass conv.
probaby conv filter count could be derived from bias->input(1)->shape[1].
consider to rename VerifyBiasGetConst() into GetVerifiedConstBiasPerFilter()
consider to rename bias into add (or addToVerify)
There was a problem hiding this comment.
If bias->input(1)->shape[1] is compared instead of conv_filter_count then it will be compared with itself.
* [GNA] Add support for DWSC, other fixes and code refactoring. * [GNA] Change supported layout to NHWC * [GNA] Detect bias const only on second position, move verification of dwsc to matcher
* [GNA] Add support for DWSC, other fixes and code refactoring. * [GNA] Change supported layout to NHWC * [GNA] Detect bias const only on second position, move verification of dwsc to matcher
Details:
Tickets: