박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 18 Jun 2019 03:00:28 +0000 (12:00 +0900)]
[loco] Enable Conv2D pad setter (#3833)
This commit enable (temporarily disabled) Conv2D pad setter.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 18 Jun 2019 02:25:35 +0000 (11:25 +0900)]
[moco/tf] add log for fix padding avgpool2d (#3829)
* [moco/tf] add log for fix padding avgpool2d
This will add result pad values of fix padding of AvgPool2D node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* move LOGGER to top of function
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 18 Jun 2019 02:23:31 +0000 (11:23 +0900)]
[moco/tf] Adding BiasAdd (#3815)
[moco/tf] Adding BiasAdd
This comits adds BiasAdd for moco/tf.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
Signed-off-by: Pavel Iliutchenko p.iliutchenk@samsung.com
Ivan Vagin/AI Tools Lab /SRR/Engineer/삼성전자 [Tue, 18 Jun 2019 01:23:09 +0000 (04:23 +0300)]
[moco/tf] provided helper for validate methods (#3804)
* [moco/tf] provided helper for frontend validate method
provided helper for frontend validate method
Signed-off-by: Ivan Vagin <ivan.vagin@samsung.com>
* Fixed formatting
Signed-off-by: Ivan Vagin <ivan.vagin@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 18 Jun 2019 01:17:49 +0000 (10:17 +0900)]
[moco/tf] Fix padding for MaxPool2D (#3805)
* [moco/tf] Fix padding for MaxPool2D
This will fix padding for MaxPool2D: same as AvgPool2D
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* fix formula
* fix pad formula
* fix typo
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 18 Jun 2019 01:09:01 +0000 (10:09 +0900)]
[moco/tf] test of shape inference for avgpool2d (#3825)
This wll add test of TestShapeTransform AvgPool2D for SAME and VALID padding type
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 17 Jun 2019 23:23:18 +0000 (08:23 +0900)]
[loco exporter] Support AvgPool2D (#3820)
This commit supports AvgPool2D loco node.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 17 Jun 2019 22:32:53 +0000 (07:32 +0900)]
[angkor] Introduce README.md (#3331)
* [angkor] Introduce README.md
This commit will add README file for `angkor`
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
* remove line
* Add words
* Fix headlines
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 17 Jun 2019 11:20:01 +0000 (20:20 +0900)]
[tf2tflite] Ignore name mismatch on testing (#3826)
tfttflite now does not count tensor name mismatch as a test failure.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 17 Jun 2019 11:16:35 +0000 (20:16 +0900)]
[moco/tf] Test pbtxt extracted from inception_v3 (#3824)
These test pbtxt files are extracted from inception_v3, containing 4, 7, 10 nodes respectively.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 17 Jun 2019 10:44:28 +0000 (19:44 +0900)]
[moco/tf] remove unnecessary comment (#3822)
This will remove unnecessary comment about padding
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 17 Jun 2019 10:42:30 +0000 (19:42 +0900)]
[nnkit-HDF5] Use index instead of name (#3821)
Previously, nnkit HDF5 export uses a tensor name when it create a
dataset. With this commit, nnkit HDF5 export uses a tensor index
instead.
This commit also updates nnkit HDF5 import action accordingly.
To maintain backward compatibility, HDF5 export stores the name of
each tensor as an independent attribute. This additional attribute
allows users to detect name mismatch same as before.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 17 Jun 2019 10:26:10 +0000 (19:26 +0900)]
[moco/tf] store padding case insensitive (#3819)
This will change to store padding string as case insensitive as it will be stored once and used many
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 17 Jun 2019 10:23:23 +0000 (19:23 +0900)]
[moco/tf] fix padding for avgpool2d (#3816)
* [moco/tf] fix padding for avgpool2d
This will fox padding calculation formula for AvgPool2D
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* with dilation and log
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 17 Jun 2019 10:02:41 +0000 (19:02 +0900)]
[loco exporter] Refactor Feature Codec export (#3818)
* [loco exporter] Refactor Feature Codec export
This commit amends detail structure to export FeatureEncode and
FeatureDecode nodes. Mainly, it redefine 'isIdentity()' utility function
and rename 'exportPermutation' into 'exportFeatureTranspose'
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
* Review fix: rename and notice
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 17 Jun 2019 09:26:16 +0000 (18:26 +0900)]
[nnkit-intf] Update README.md (#3817)
This commit updates nnkit-intf's README.md to be compatible with
Doxygen.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 17 Jun 2019 09:13:06 +0000 (18:13 +0900)]
[loco exporter] General use of checking padding (#3811)
* [loco exporter] General use of checking padding
This commit fixes getOpPadding function. It now supports SAME padding.
Also function signature is generalized to deal with any padding, not just
for MaxPool2d node.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
* Review fix: careful about unsigned variable
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 17 Jun 2019 07:23:50 +0000 (16:23 +0900)]
[moco/tf] Fix shape transformation for AvgPool2D (#3812)
This will fix shape transformation formula for AvgPool2D with padding type
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 17 Jun 2019 07:04:53 +0000 (16:04 +0900)]
[moco/tf] fix maxpool2d shape calculation (#3801)
* [moco/tf] fix maxpool2d shape calculation
This will fix maxpool2d shape calculation as like avgpool2d
- fix also op test model
- add PaddingData annotation
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* fix formula
* fix typo
* rename
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 17 Jun 2019 06:59:07 +0000 (15:59 +0900)]
[tf2tflite] Introduce test infrastructure (#3809)
* [tf2tflite] Introduce test infrastructure
This commit implements very basic test infrastructure for tf2tflite.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Fix typos
* Update testall
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 17 Jun 2019 05:10:20 +0000 (14:10 +0900)]
[moco/tf] Add comma to logging (#3813)
This will add comma in log string of ShapeTransform of AvgPool2D
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 17 Jun 2019 03:14:05 +0000 (12:14 +0900)]
[tftestinfo] isFloatTensor as a const method (#3798)
isFloatTensor does not update the content of ParsedTensor.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 17 Jun 2019 02:02:30 +0000 (11:02 +0900)]
[loco exporter] Parse and export graph I/O name (#3803)
* [loco exporter] Parse and export graph I/O name
Graph input and output names are parsed from loco graph's information
and then exported as name of input and output tensors.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
* Fix typo
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 17 Jun 2019 00:08:47 +0000 (09:08 +0900)]
[locop] Introduce SymbolTable alias (#3794)
This commit introduces SymbolTable alias. This alias make it easy to
refactor code layout.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 14 Jun 2019 09:53:19 +0000 (18:53 +0900)]
[loco] File-level comments for tensorflow.test.cpp (#3799)
This commit adds file-level comments at the top of tensorflow.test.cpp.
These file-level comments are valid for all the tests in this file.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 14 Jun 2019 09:14:05 +0000 (18:14 +0900)]
[loco] Introduce TensorBiasAdd node (#3783)
This commit introduces TensorBiasAdd node.
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 14 Jun 2019 09:09:05 +0000 (18:09 +0900)]
[loco exporter] FeatureCodec identity export (#3795)
Previously FeatureCodec nodes tried to be deleted when they are
identity, but that was not successful. This commit export those as
identity, i.e. CONCATENTATION with one input.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 14 Jun 2019 08:26:33 +0000 (17:26 +0900)]
[moco/tf] simplify adding annotation cleaning (#3796)
This will rewrite ClearAnnotTransform so that adding cleaning will be bit simple
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 14 Jun 2019 08:18:21 +0000 (17:18 +0900)]
[moco/tf] use auto for types in FixShapeTransform (#3793)
This will fix to use auto for types in FixShapeTransform methods
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 14 Jun 2019 08:09:28 +0000 (17:09 +0900)]
[moco/tf] Implement pad fixing for AvgPool2D (#3790)
This will implement pad fixing for AvgPool2D
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 14 Jun 2019 07:38:02 +0000 (16:38 +0900)]
[moco/tf] Conv2D test for the 1st Conv2D in inception v3 (#3787)
This adds Conv2D test for the 1st Conv2D in inception v3.
This test failed because the input name is a form of "input:0", which was not correctly processed by moco before.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 14 Jun 2019 06:41:19 +0000 (15:41 +0900)]
[tf2tflite] Fix build break (#3786)
Recent commit on moco changes the frontend interface, but this change is
not reflected into tf2tflite.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 14 Jun 2019 06:29:21 +0000 (15:29 +0900)]
[moco/tf] implement fix shape for avgpool2d (#3777)
This will implement fix_shape for avgpool2d node in fixshape transformation
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 14 Jun 2019 06:12:39 +0000 (15:12 +0900)]
[moco/tf] Make moco to use TensorName class (#3781)
This commit makes 1) moco to use TensorName class 2) renames some naming (e.g., renaming `nodes` to `tensors` if `nodes` actually means `tensors`)
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 14 Jun 2019 06:00:31 +0000 (15:00 +0900)]
[tf2tflite] Initial commit (#3779)
This commit introduces tf2tflite with minimal implementation.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 14 Jun 2019 05:10:22 +0000 (14:10 +0900)]
[moco/tf] add AvgPool test model (#3774)
This will add basic test material for AvgPool import
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 14 Jun 2019 05:08:23 +0000 (14:08 +0900)]
[loco] Introduce BiasEncode (#3766)
This commit extends loco with BiasEncode node which converts a tensor as
a bias.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 14 Jun 2019 05:06:00 +0000 (14:06 +0900)]
[moco/tf] TensorName class to represent the name of tensor (#3776)
* [moco/tf] TensorName class to represent the name of tensor
This class represents tensor name, which is a form of _nodeName_ ":" _number
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* adds a empty line
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 14 Jun 2019 04:44:31 +0000 (13:44 +0900)]
[moco/tf] Add shapeinfdata feature get,set (#3769)
* [moco/tf] Add shapeinfdata feature get,set
This will add getter and setter with FeatureShape for ShapeInferenceData
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* remove unused, fix comment
* fix typo
Ivan Vagin/AI Tools Lab /SRR/Engineer/삼성전자 [Fri, 14 Jun 2019 04:43:16 +0000 (07:43 +0300)]
[loco exporter] Implement loco exporter v0.1 (#3737)
This commit adds minimal implementation that is able to export
a single node network that includes one MaxPool2D op.
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
Signed-off-by: Efimov Alexander <a.efimov@samsung.com>
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 14 Jun 2019 04:43:05 +0000 (13:43 +0900)]
[moco/tf] Introduce AvgPool2D node builder (#3772)
This will introduce AvgPool2D node builder and related empty methods
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 14 Jun 2019 04:40:06 +0000 (13:40 +0900)]
[loco] Canonical Node Visitor with default implementation (#3771)
* [loco] Canonical Node Visitor with default implementation
With this commit, users are able to implement a node visitor with their
own default implementation.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Fix a typo
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 14 Jun 2019 01:59:33 +0000 (10:59 +0900)]
[locop] Show correct opname for canonical dialect (#3768)
With this commit, LinearV1 formatter shows correct opname for
canonical dialect.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 14 Jun 2019 01:43:23 +0000 (10:43 +0900)]
[moco/tf] Introduce str_toupper (#3770)
This will introduce str_toupper() method to convert string to upper case
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 13 Jun 2019 10:19:08 +0000 (19:19 +0900)]
[moco/tf] add missing fix_padding (#3764)
This will add missing fix_padding for Conv2D and FilterEncode
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 13 Jun 2019 09:55:39 +0000 (18:55 +0900)]
[moco/tf] clean padding annotation (#3760)
This will add padding data annotation free in transformation
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 13 Jun 2019 09:26:32 +0000 (18:26 +0900)]
[moco/tf] basic Conv2D (padding not yet implemented) (#3731)
* [moco/tf] basic Conv2D (padding not yet implemented)
This adds basic implementation of Conv2D.
Padding and `FixShapeTransform` is not yet implemented.
Note: To run the test, `FixShape()` for Conv2D should be implemented. So I added a minimum implementation in this commit.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* renaming vars
* split one GraphUpdate into two
* assert for padding
* supporting NCHW in FeatureDecode
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 13 Jun 2019 07:05:53 +0000 (16:05 +0900)]
[locop] Initial commit (#3756)
This commit creates "locop" which provides various pretty-printing
routines for loco IR.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 13 Jun 2019 07:00:59 +0000 (16:00 +0900)]
[loco] Canonical Nodes (#3759)
This commit updates the existing nodes to inherit canonical node
interface.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 13 Jun 2019 05:21:17 +0000 (14:21 +0900)]
[moco/tf] Introduce PaddingData (#3755)
This will introduce PaddingData as annotation for converting TF padding to loco padding
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 13 Jun 2019 03:56:38 +0000 (12:56 +0900)]
[loco] Introduce CanonicalNodeVisitor (#3748)
This commit introduces CanonicalNodeVisitor interface, and adds a related
test.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 13 Jun 2019 03:50:02 +0000 (12:50 +0900)]
[moco/tf] Introduce empty FixPaddingTransform (#3754)
This will introduce empty FixPaddingTransform to convert TF padding to loco padding
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 13 Jun 2019 03:49:47 +0000 (12:49 +0900)]
[loco_exporter] Initial commit (#3757)
This commit creates "loco_exporter" project with minimal implementation.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 12 Jun 2019 10:52:19 +0000 (19:52 +0900)]
[loco] Helper function to check equality of two dims (#3727)
* [loco] Helper function to check equality of two dims
This adds a helper operator `==` to check equality of two Dimensions.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* modified per comments
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 12 Jun 2019 10:43:53 +0000 (19:43 +0900)]
[moco/tf] make copy_shape_values as template (#3745)
This will change copy_shape_values() to support loco::Node type to reduce duplicate codes
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 12 Jun 2019 09:30:08 +0000 (18:30 +0900)]
[loco] Remove outdated comments in Nodes.h (#3746)
This commit removes several outdated comments in Nodes.h.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 12 Jun 2019 09:27:19 +0000 (18:27 +0900)]
[loco] Expose verifier and codec through loco.h (#3749)
This commit updates loco.h to include verifier and permuting codec.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 12 Jun 2019 08:51:58 +0000 (17:51 +0900)]
[moco/tf] class storing user-provided compiler arguments (#3734)
* [moco/tf] class storing user-provided compiler arguments
When a user run the compiler with arguments, these arguments are parsed and stored into this class.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* Using default constructor
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 12 Jun 2019 05:58:56 +0000 (14:58 +0900)]
[locomotiv] Conv2D (padding is not yet added) (#3725)
* [locomotiv] Conv2D (padding is not yet added)
This adds Conv2D implementation of locomotiv. Padding is not implemented yet.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* revised per comments
* Fixed type. Added TODO
* int to uint32_t
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 12 Jun 2019 05:33:40 +0000 (14:33 +0900)]
[moco/tf] a test case for Conv2D (#3728)
This adds Conv2D_000 test case for moco/tf.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 12 Jun 2019 04:48:13 +0000 (13:48 +0900)]
[loco] Introduce CanonicalNode interface (#3743)
This commit introduces CanonicalNode interface, and rewrites Forward
node as an example.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 12 Jun 2019 03:45:55 +0000 (12:45 +0900)]
[locomotiv] Add MaxPool2D test case for padding (#3742)
This commit test MaxPool2D execution with padding support.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 12 Jun 2019 03:45:33 +0000 (12:45 +0900)]
[moco/tf] Clean frontend annotations (#3740)
This will clean frontend temporary annotations
- remove test that checks shape information
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 12 Jun 2019 01:54:34 +0000 (10:54 +0900)]
[loco] Introduce CanonicalOpcode enum class (#3741)
This commit introduces CanonicalOpcode enum class which serves as a
unique tag for each CanonicalNode.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 11 Jun 2019 10:38:18 +0000 (19:38 +0900)]
[moco/tf] Test case for Concat axis with -1 (#3721)
* [moco/tf] Test case for Concat axis with -1
This will fill the test code for Cocat axis with -1 value
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* check values
* update comment and formula
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 11 Jun 2019 10:04:06 +0000 (19:04 +0900)]
[loco] Extend Node with dialect method (#3730)
This commit extends Node with dialect method as a step toward
multi-dialect support.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 11 Jun 2019 10:03:51 +0000 (19:03 +0900)]
[moco/tf] Enable concat inference test (#3724)
This will enable full Concat inference test
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 11 Jun 2019 10:02:09 +0000 (19:02 +0900)]
[loco] Introduce AvgPool2D (#3720)
* [loco] Introduce AvgPool2D
This will introduce AvgPool2D for average pool operator.
Attributes are same as that of exising MaxPool2D.
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* add divisor
* rebase
* rebase and add todo
* use full, valid
* uncomment pad
* update note
* use convention
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 11 Jun 2019 07:19:20 +0000 (16:19 +0900)]
[loco] Introduce CanonicalDialect singleton (#3729)
This commit introduces CanonicalDialect singleton which serves as an
in-memory unique identifier.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 11 Jun 2019 04:18:16 +0000 (13:18 +0900)]
[loco] Enable MaxPool2D Pad setter (#3723)
This commit enables (temporarily disabled) MaxPool2D Pad setter, and add
a related test.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 11 Jun 2019 03:06:45 +0000 (12:06 +0900)]
[moco/tf] add include for stringstream (#3722)
This will add include of sstream that was missed in previous commit
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 11 Jun 2019 03:05:06 +0000 (12:05 +0900)]
[locomotiv] MaxPool2D support padding (#3719)
* [locomotiv] MaxPool2D support padding
By this commit, MaxPool2D node execution is able to deal with padding.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
* Review feedback: local function name
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 11 Jun 2019 02:20:27 +0000 (11:20 +0900)]
[locomotiv] Support FilterEncode execution (#3716)
This commit supports FilterEncode node execution.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 11 Jun 2019 01:00:14 +0000 (10:00 +0900)]
[locomotiv] Combine Session test in one file (#3700)
Previously Session tests was divided into two files, public and internal
test. This commit combine them into one.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 11 Jun 2019 00:56:50 +0000 (09:56 +0900)]
[loco] Introduce replace helper (#3697)
This commit introduce "replace" helper which facilitate graph rewriting.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 11 Jun 2019 00:55:35 +0000 (09:55 +0900)]
[loco] Rewrite FixedArityNode as a mix-in (#3699)
* [loco] Rewrite FixedArityNode as a mix-in
FixedArityNode now serves as a mix-in rather than a concrete
implementation.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Inherit Base
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 10 Jun 2019 23:43:47 +0000 (08:43 +0900)]
[moco/tf] Implement fix_shape for TensorConcat (#3694)
* [moco/tf] Implement fix_shape for TensorConcat
This will implement fix_shape in shape inference for TensorConcat node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* add assert and comment
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 10 Jun 2019 10:01:33 +0000 (19:01 +0900)]
[moco/tf] const test with shape inference (#3709)
* [moco/tf] const test with shape inference
This will add test of ShapeInferenceData for ConstGen node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* use name shape
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 10 Jun 2019 07:41:01 +0000 (16:41 +0900)]
[locomotiv] Amend runtime error message for executions (#3712)
This commit edits runtime error messages for executions. It unifies
message format and style and correct some wrong infos.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 10 Jun 2019 07:29:11 +0000 (16:29 +0900)]
[moco/tf] Dump ShapeInferenceData for outputs (#3695)
This will dump ShapeInferenceData for output nodes with moco log library
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 10 Jun 2019 01:58:04 +0000 (10:58 +0900)]
[locomotiv] Execution verify and annotate domain info (#3701)
This commit makes use of domain annotation tools in node executions.
Now node executions check appropriate domain info preparation, and
annotate appropriate domain info on execution run.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 5 Jun 2019 06:39:10 +0000 (15:39 +0900)]
[locomotiv] Node domain annotation for Session (#3682)
Session now annotate node domain on input setting, and delete annotated
domain information on destruction.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 5 Jun 2019 04:32:56 +0000 (13:32 +0900)]
[moco/tf] Implement fix_shape for ConstGen (#3692)
This will implement fix_shape in shape inference for ConstGen node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 5 Jun 2019 03:54:47 +0000 (12:54 +0900)]
[moco/tf] Implement fix_shape for Pull (#3693)
This will implement fix_shape in shape inference for Pull node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 5 Jun 2019 03:37:48 +0000 (12:37 +0900)]
[loco] Add Conv2D_Float_000 testcase (#3691)
* [loco] Add Conv2D_Float_000 testcase
This commit adds TensorFlowTest.Conv2D_Float_000 testcase which
checks whether loco is able to encode a simple TensorFlow conv2d
network.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Update per feedback
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 5 Jun 2019 01:54:49 +0000 (10:54 +0900)]
[tfkit] Fix for already packed or unpacked (#3688)
This will fix tfkit to run pack or unpack for data are already packed or unpakced
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 5 Jun 2019 01:53:38 +0000 (10:53 +0900)]
[tfkit] support pack,unpack for int32 type (#3690)
This will make pack, unpack to support int32 data type
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 5 Jun 2019 01:17:36 +0000 (10:17 +0900)]
[moco/tf] Implement fix_shape for copy (#3685)
This will implement fix_shape() methods for nodes that just will do with copy input shape to output
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 5 Jun 2019 00:29:17 +0000 (09:29 +0900)]
[loco] Support Padding over Conv2D (#3686)
This commit extends Conv2D nodes with pad field and its getter.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 5 Jun 2019 00:28:50 +0000 (09:28 +0900)]
[loco] Introduce Filter Permuting Encoder (#3684)
This commit introduces a permuting encoder for filter values.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 5 Jun 2019 00:28:41 +0000 (09:28 +0900)]
[moco/tf] Change calling transform run (#3687)
This will change calling transform run by always call and update changed bool value
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 5 Jun 2019 00:27:54 +0000 (09:27 +0900)]
[moco/tf] Register two transforms (#3683)
This will register and run (currently dummy) two transforms after loading the graph
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 4 Jun 2019 23:40:45 +0000 (08:40 +0900)]
[loco] Introduce Pad field and its getter on MaxPool2D (#3679)
This commit extends MaxPool2D node with a padding field and its getter.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 4 Jun 2019 09:30:51 +0000 (18:30 +0900)]
[loco] Implement FilterAxis Permutation (#3675)
* [loco] Implement FilterAxis Permutation
This commit implements Permutation over FilterAxis similarly as that
over FeatureAxis.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Return const reference
* Fix a typo
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 4 Jun 2019 08:22:00 +0000 (17:22 +0900)]
[moco/tf] introduce FixShapeTransform (#3673)
* [moco/tf] introduce FixShapeTransform
This will introduce FixShapeTransform class to perform shape inference from known nodes and place holders for each node types.
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* reduce line
* sort
* use ||
* fix call call condition
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 4 Jun 2019 08:16:44 +0000 (17:16 +0900)]
[loco] Introduce Dialect interface (#3676)
This commit introduces Dialect interface class.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 4 Jun 2019 08:11:21 +0000 (17:11 +0900)]
[moco/tf] Add Concat node converter (#3671)
* [moco/tf] Add Concat node converter
This will add Concat node converter for TF loading
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* rename
* update comments
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 4 Jun 2019 04:53:07 +0000 (13:53 +0900)]
[locomotiv] Fix ASSERT_EQ for integer test of ConstGen (#3678)
It seems good to use ASSERT_EQ for integer test, ASSERT_FLOAT_EQ for
float test.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 4 Jun 2019 02:28:30 +0000 (11:28 +0900)]
[moco/tf] Introduce ClearAnnotTransform (#3672)
This will introduce ClearAnnotTransform class as a place holder to remove temporary annotation while loading TF model
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 3 Jun 2019 23:53:33 +0000 (08:53 +0900)]
[loco] Introduce FilterAxis enum class (#3669)
This commit introduces FilterAxis enum class which denotes the name of
each axis of any convolutional filter data.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>