NormalizeL2 transformation #1892
Conversation
mbencer
added
the
category: transformations
openvino-pushbot
added
category: IE Tests
inference-engine/src/transformations/src/transformations/normalize_l2_fusion.cpp
Outdated
Show resolved
Hide resolved
inference-engine/src/transformations/src/transformations/normalize_l2_fusion.cpp
Outdated
Show resolved
Hide resolved
inference-engine/src/transformations/src/transformations/normalize_l2_fusion.cpp
Show resolved
Hide resolved
inference-engine/src/transformations/src/transformations/normalize_l2_fusion.cpp
Outdated
Show resolved
Hide resolved
inference-engine/src/transformations/include/transformations/utils/utils.hpp
Outdated
Show resolved
Hide resolved
inference-engine/src/transformations/include/transformations/utils/utils.hpp
Outdated
Show resolved
Hide resolved
| return false; | ||
| } | ||
|
|
||
| if (shape_size(constant->get_shape()) > 1) { |
There was a problem hiding this comment.
What if the constant is empty? I mean if the size is equal to 0.
There was a problem hiding this comment.
@lazarevevgeny As I understand shape_size(constant->get_shape()) == 0 means that it is scalar
There was a problem hiding this comment.
Scalar is only if shapes are empty.
But if one dimension is equal to 0, shape_size also returns 0, if I remember right.
But in any way it is a really strange case.
There was a problem hiding this comment.
I think this check is not only for a scalar but constant with any shape like {1, 1, 1} that we can also cast to a single value.
So shape_size returns "Number of elements in spanned by a shape".
There was a problem hiding this comment.
added self-explanatory variable (is_scalar_or_single_elem) + used is_scalar helper
| return false; | ||
| } | ||
|
|
||
| if (shape_size(constant->get_shape()) > 1) { |
There was a problem hiding this comment.
Scalar is only if shapes are empty.
But if one dimension is equal to 0, shape_size also returns 0, if I remember right.
But in any way it is a really strange case.
inference-engine/src/transformations/src/transformations/normalize_l2_fusion.cpp
Show resolved
Hide resolved
| auto reduce_sum = std::make_shared<ngraph::opset4::ReduceSum>(pow, axes); | ||
| auto sqrt = std::make_shared<ngraph::opset4::Sqrt>(reduce_sum); | ||
| auto eps_const = ngraph::pattern::wrap_type<ngraph::opset4::Constant>(); | ||
| auto sqrt_add_eps = std::make_shared<ngraph::opset4::Add>(sqrt, eps_const); |
There was a problem hiding this comment.
Maybe we can join these two passes if check sqrt_max_eps or sqrt_add_eps types in the callback.
@GlebKazantaev what do you think about it?
|
@ilyachur , @GlebKazantaev , please, take a look. |
I asked @GlebKazantaev to look at this PR. I think we can join two passes from this PR. |
|
LGTM |
* first version of implementation * added unit tests * changed multiply to pow * doc + unit tests * more unit tests * code review remarks * missing new line * remarks * review remarks * Build fix - update constant check function in HSwishFusionWithClamp Co-authored-by: mitruska <[email protected]>
* first version of implementation * added unit tests * changed multiply to pow * doc + unit tests * more unit tests * code review remarks * missing new line * remarks * review remarks * Build fix - update constant check function in HSwishFusionWithClamp Co-authored-by: mitruska <[email protected]>
No description provided.