채성우/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 16 Sep 2019 03:26:53 +0000 (12:26 +0900)]
[loco] Replace Pad.h with Padding2D.h (#7440)
This commit replaces Pad.h with Padding2D.h in loco.
Related: #7357
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 16 Sep 2019 02:27:19 +0000 (11:27 +0900)]
[moco-tf] Use generated opname (#7435)
This commit revises TFFormattedGraph to use generated opname instead of
hardcoded one.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 16 Sep 2019 01:44:48 +0000 (10:44 +0900)]
[moco-tf] Introduce TFDialect for Mean (#7320)
* [moco-tf] Introduce TFDialect for Mean
This commit will introduce TFDialect for `Mean`
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
* apply comments
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 16 Sep 2019 01:32:37 +0000 (10:32 +0900)]
Make to support PReLU op for acl neon (#7330)
This commit makes to support PReLU op for acl neon.
- Add PReLU kernel of acl neon
- Make neurun supports PReLU op for acl neon
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 16 Sep 2019 01:30:38 +0000 (10:30 +0900)]
[loco] Rename Pad class to Padding2D (#7395)
This commit is for renaming existing Pad class to Padding2D in loco.
Related: #7357
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 16 Sep 2019 01:24:31 +0000 (10:24 +0900)]
Enable QASYMM8 of PReLU op for acl cl (#7292)
This commit enables QASYMM8 of PReLU op for acl cl.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 16 Sep 2019 00:35:16 +0000 (09:35 +0900)]
[exo-tflite] Adding MaxPool2DConverter into convert.cpp (#7400)
Now MaxPool2DConverter will work with knob on.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
[nnc] Remove frontend passes (#7423)
Remove frontend passes. Call importers directly instead.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[nnc] Remove frontend passes in tests (#7424)
Remove frontend passes. Call importers directly instead.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[nnc] Remove frontend passes in utils (#7425)
Remove frontend passes. Call importers directly instead.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir] Remove DTYPE (#7422)
Remove no longer used type alias `DTYPE`.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir_onnx] Replace DTYPE with DataType. (#7421)
Replace deprecated `DTYPE` with `DataType`.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
Павел Ильютченко/AI Tools Lab /SRR/Engineer/삼성전자 [Thu, 12 Sep 2019 17:04:41 +0000 (20:04 +0300)]
[mir_onnx] Eliminate TransposeOp in Gemm when weights is ConstantOp (#7416)
* Fixed Gemm importer for elimination extra Transposes
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
[mir] Remove default value of `data_format` parameter (#7420)
* Remove default value of `data_format` parameter of conv-like op constructors.
* Rename `CommonProps.h` -> `PaddingType.h`.
* Remove deprecated enum members of `PaddingType`.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir_tflite] Set `data_format` parameter explicitly (#7415)
* Pass `data_format` to constructors of conv-like ops explicitly.
* Replace deprecated `PaddingType::Same` with equivalent `PaddingType::SameUpper`.
* Remove unused includes.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir_caffe] Remove unused include (#7417)
Remove one unused include directive
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir_caffe2] Remove unused include (#7418)
Remove one unused include directive
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir2loco] Set data format explicitly (#7419)
Pass `data_format` argument to conv-like constructors explicitly.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir] Fix padding computation (#7413)
Fix padding computation for `DeConv2DOp`.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[nnc] Pass data format to conv-like op constructors explicitly (#7409)
* Pass data format to constructors of `Conv2DOp`, `DeConv2DOp`, `DepthwiseConv2DOp` explicitly.
* Add missing parameter to call of `DeConv2DOp` constructor.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
Павел Ильютченко/AI Tools Lab /SRR/Engineer/삼성전자 [Wed, 11 Sep 2019 12:56:47 +0000 (15:56 +0300)]
[nnc] Unittests for Layout Switcher (#7321)
* Implemented tests for all layout related operations
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
[mir] Remove old `DeConv2DOp` constructors (#7405)
* Remove old `DeConv2DOp` constructors which accept only one padding vector instead of two.
* Add enum members to `PaddingType` to match ONNX specification.
* Adjust padding and output shape calculation for `DeConv2DOp`.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 08:58:00 +0000 (17:58 +0900)]
[exo-tflite] Make unit test use perm initialized with NHWC layout (#7396)
This make FeatureEncode and FeatureDecode use NHWC layout. Previous test code does not have this info and this test will fail when MaxPool2DConverter is enabled.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 08:20:08 +0000 (17:20 +0900)]
[exo-tflite] Introducing MaxPool2DConverter (#7387)
* [exo-tflite] Introducing MaxPool2DConverter
This adds MaxPool2DConverter, which converts loco::MaxPool2D to locoex::TFLMaxPool2D.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* adds template instantiation
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 11 Sep 2019 08:14:33 +0000 (17:14 +0900)]
[loco] Add TransposedConv2DLayer for test GraphBuilder (#7392)
This commit adds TransposedConv2DLayer for test GraphBuilder.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
Павел Ильютченко/AI Tools Lab /SRR/Engineer/삼성전자 [Wed, 11 Sep 2019 08:12:43 +0000 (11:12 +0300)]
[loco] Introduce Matrix Domain with its codecs (#7271)
* Added Domain::Matrix
* Added MatrixShape, MatrixAxis, MatrixIndex
* Added and implemented MatrixEncoder and MatrixDecoder
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
Павел Ильютченко/AI Tools Lab /SRR/Engineer/삼성전자 [Wed, 11 Sep 2019 08:08:55 +0000 (11:08 +0300)]
[mir2loco] Support DataFormat from mir in layout related ops (#7322)
* Created createHWIOFilterEncode
* Fixed createFeatureEncode, createFeatureDecode using DataFormat from
mir operations
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 11 Sep 2019 08:06:03 +0000 (17:06 +0900)]
[ann-ref] Resolve SOLVe issues in ann-ref (#7380)
This commit will add missing Tensorflow author and source statements
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 11 Sep 2019 08:05:36 +0000 (17:05 +0900)]
[exo-tflite] Rename existing Pad class (#7388)
This commit is for renaming existing Pad class in exo-tflite.
Related: #7357
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 11 Sep 2019 08:04:55 +0000 (17:04 +0900)]
[loco] Rename existing Pad class (#7393)
This commit is for renaming existing Pad class in loco.
Related: #7357
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 08:01:39 +0000 (17:01 +0900)]
[exo-tflite] Fix wrong ifdef guard (#7391)
This file was previously in `conversion` directory and ifdef guard contains CONVERSION. Now the file is in `src` so `CONVERSION` was removed.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 11 Sep 2019 07:58:05 +0000 (16:58 +0900)]
[tfkit] Resolve SOLVe issues in tfkit (#7381)
This commit will add Tensorflow author and remove redundant license statements
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 07:56:52 +0000 (16:56 +0900)]
[logo] Preparing SimplifyDomainConversionPass to add more visit() (#7157)
* [logo] Preparing SimplifyDomainConversionPass to add more visit()
This is a preparation to introduce other visit(..).
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* nullifying input, etc
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 11 Sep 2019 07:56:07 +0000 (16:56 +0900)]
[nnc] Resolve SOLVe issues in nnc (#7383)
This commit will modify Apache 2.0 License to BSD 2-Clause License
since the code itself is 100% matched with original code and
caffe is under BSD 2-Clause License
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 11 Sep 2019 07:51:16 +0000 (16:51 +0900)]
[mir2loco] Rename existing Pad class (#7386)
This commit is for renaming existing Pad class in mir2loco.
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 11 Sep 2019 07:40:55 +0000 (16:40 +0900)]
[mocotest-tf] Enable Some Broadcasting Tests (#7385)
This commit enables supported broadcasting tests.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 07:36:33 +0000 (16:36 +0900)]
[nnpkg] Update specification for operators (#7355)
It specifies the operators will use the same input, output and
attributes from tensorflow lite.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 11 Sep 2019 07:17:20 +0000 (16:17 +0900)]
[moco-tf] Rename existing Pad class (#7384)
This commit is for renaming existing Pad class in moco-tf.
Related: #7357
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 11 Sep 2019 07:09:36 +0000 (16:09 +0900)]
[moco-tf] Support Broadcasting (#7378)
This commit revises canonicalize_eltwise_binary_node to insert proper
boradcasting nodes if necessary.
This change will affect the canonicalization process of TFAdd, TFSub, TFMul, TFDiv.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 07:09:23 +0000 (16:09 +0900)]
[exo-tflite] exporting TFLMaxPool2D (#7382)
Code to export TFLMaxPool2D was added. Since much of this code is duplicated with TFLAveragePool2D, common code was extracted into a separate method.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 11 Sep 2019 07:08:57 +0000 (16:08 +0900)]
Add UNIT_Add_005 TensorFlow testcase (#7379)
This commit adds UNIT_Add_005 TensorFlow testcase. This test is almost
same as UNIT_Add_001 except that it requires broadcasting over LHS (not RHS).
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
[neurun] Rename Scheduler to HEScheduler (#7340)
Since this is a scheduler for heterogeneous execution scheduling,
it must be called appropriately
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 07:02:06 +0000 (16:02 +0900)]
[exo-tflite] Introduce TensorShapeExpander (#7373)
* [exo-tflite] Introduce TensorShapeExpander
This will introduce TensorShapeExpander in TFLShapeInferenceRule for shape expansion for operations that has broadcasting
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* remove incs
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 11 Sep 2019 06:31:17 +0000 (15:31 +0900)]
[loco] add alias before renaming pad class (#7377)
This commit is getting ready for renaming pad class.
Related: #7357
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
이춘석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 11 Sep 2019 05:44:58 +0000 (14:44 +0900)]
Add BSD-2-Clause license (#7370)
- Add license in nnfw.spec
- Add license text
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 11 Sep 2019 05:34:54 +0000 (14:34 +0900)]
[locomotiv] Simplify Element-wise Unary Op Kernels (#7358)
* [locomotiv] Simplify Element-wise Unary Op Kernels
This commit rewrites element-wise unary op kernels on top of shared
infrastructure.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Fix copy-and-paste error
* Fix error message
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 05:33:50 +0000 (14:33 +0900)]
[exo-tflite] Adding ReluConverter test (#7287)
* [exo-tflite] Adding ReluConverter test
Two test cases were added: loco::relu with tensor input/putput, loco::relu with feature input/output.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* remove code that checks knob
* using plan.at(n) instead of bunch of statements
* using find_first_node_bytype
* no loop, used iterator
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 05:32:47 +0000 (14:32 +0900)]
[exo-tflite] introducing knob for TFLMaxPool2D (#7376)
This adds knob for TFLMaxPool2D, for backward compatibility reason.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 11 Sep 2019 05:31:25 +0000 (14:31 +0900)]
[moco-tf] Introduce CanonicalEltwiseInputConnector (#7368)
* [moco-tf] Introduce CanonicalEltwiseInputConnector
This commit introduces CanonicalEltwiseInputConnector which facilitates
pipeline extension.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Use NodeTy
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 05:29:59 +0000 (14:29 +0900)]
[exo-tflite] Adding AvgPool2DConverter (#7317)
* [exo-tflite] Adding AvgPool2DConverter
AvgPool2DConverter is added, which converts loco::AvgPool2D to locoex::TFLAveragePool2D.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* fix build error
* fix comments
* fix typo
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 05:29:09 +0000 (14:29 +0900)]
[exo-tflite] TypeInference for TFLMaxPool2D (#7350)
This adds type inference for `TFLMaxPool2D`.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 05:28:25 +0000 (14:28 +0900)]
[exo-tflite] shape inference for TFLMaxPool2D (#7374)
Adding shape inference for TFLMaxPool2D
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 05:23:33 +0000 (14:23 +0900)]
[logo] Revise test to use make_graph (#7371)
This will revise ConstantFoldingPass test to use make_graph() method instead of using stack value
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 11 Sep 2019 05:04:50 +0000 (14:04 +0900)]
[locop] Do NOT show Shape attribute of Pull node (#7338)
"shape" attribute in Pull node is deprecated.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 05:04:37 +0000 (14:04 +0900)]
[exo-tflite] shape inference for TFLAveragePool2D (#7345)
Adding shape inference for TFLAveragePool2D and two test cases.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 11 Sep 2019 05:02:47 +0000 (14:02 +0900)]
[moco-tf] Introduce Broadcast Helper (#7334)
* [moco-tf] Introduce Broadcast Helper
This commit introduces Broadcast Helper module which facilitates
numpy-style broadcasting support.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Fix a typo
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 03:52:11 +0000 (12:52 +0900)]
[moco-tf] Remove unused CanonicalNodes.lst (#7366)
This will remove not used anymore CanonicalNodes.lst file
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 03:51:46 +0000 (12:51 +0900)]
[moco-tf] Remove unused NodeShape (#7367)
This will remove not used anymore moco::tf::NodeShape
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 03:34:32 +0000 (12:34 +0900)]
[exo-tflite] Convert for EltwiseAdd (#7353)
This will add converter for canonical EltwiseAdd to TFLAdd
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 11 Sep 2019 03:34:05 +0000 (12:34 +0900)]
[exo-tflite] Type inference for TFLAdd (#7354)
This will add type inference for TFLAdd node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 11 Sep 2019 03:33:12 +0000 (12:33 +0900)]
[tfinfo] Rename targets (#7369)
* [tfinfo] Rename targets
This commit renames targets in tfinfo project to have tfinfo as prefix.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Fix a typo
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 11 Sep 2019 00:44:03 +0000 (09:44 +0900)]
[moco-tf] Introduce canonicalize_eltwise_binary_node (#7327)
* [moco-tf] Introduce canonicalize_eltwise_binary_node
A similar pattern is repeated across several element-wise binary
operation converters.
This commit extracts this pattern as canonicalize_eltwise_binary_node
to reduce code duplication.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Put implementations in anonymous namespace
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 11 Sep 2019 00:22:35 +0000 (09:22 +0900)]
[loco] Remove unused member of FeatureEncodeLayer (#7344)
This commit removes unused member '_perm' of FeatureEncodeLayer.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 11 Sep 2019 00:22:13 +0000 (09:22 +0900)]
[loco] Add FilterEncodeLayer for test GraphBuilder (#7343)
This commit adds FilterEncodeLayer for test GraphBuilder.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
[mir] Switch to new `DeConv2DOp` constructor (#7361)
Construct `DeConv2DOp` with the new constructor.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir_caffe] Switch to new `DeConv2DOp` constructor (#7362)
Construct `DeConv2DOp` with the new constructor.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir] Add a new constructor to `DeConv2DOp` (#7359)
All existing constructors miss the "padding after" parameter. They will be replaced with the new one.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 10 Sep 2019 09:54:17 +0000 (18:54 +0900)]
[cop, neurun] Use cpu backend by default for custom op (#7346)
`neurun` uses `acl_cl` backends by default.
However, for custom op, cpu backend will be used by default.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 10 Sep 2019 09:49:19 +0000 (18:49 +0900)]
[exo-tflite] Modifying TODO in knob.lst (#7347)
* [exo-tflite] Modifying TODO in knob.lst
Modifying names in TODO from TFL names to canonical names.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* reordering eltwiseAdd
이춘석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 10 Sep 2019 09:36:59 +0000 (18:36 +0900)]
Implement backend manipulation API (#7335)
* Implement backend manipulation API
- Implement backend manipulation API
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
* Apply suggestion
* fix format
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 10 Sep 2019 09:17:40 +0000 (18:17 +0900)]
[nnpkg] update custom operator specification (#7337)
* [nnpkg] update custom operator specification
It updates custom operator specification.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
* Update custom operator specificaiton
- Remove `nncc`
- Change the sentence as requested by review
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
[neurun] Eliminate redundant fn in acl KernegGen (#7302)
No need to do extra move for most of the functions
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 10 Sep 2019 08:49:07 +0000 (17:49 +0900)]
[cop] Set isNnAPi=true for CustomNode (#7341)
It sets isNnAPI for CustomNode to enable backend selection.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 10 Sep 2019 08:30:32 +0000 (17:30 +0900)]
[locomotiv] Decouple NodeDataImpl and NodeAnnotation (#7332)
NodeDataImpl now inherits NodeData interface, which allows users to
annotate any NodeData implementation other than NodeDataImpl.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 10 Sep 2019 07:00:41 +0000 (16:00 +0900)]
[exo-tflite] Adding TFLAveragePool2D into OperationExporter (#7315)
TFLAveragePool2D was added into OperationExporter.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
이춘석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 10 Sep 2019 04:56:15 +0000 (13:56 +0900)]
API to manipulate backend setting (#7298)
Two functions(nnfw_set_default_backend, nnfw_set_op_backend) are
added.
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 10 Sep 2019 04:18:11 +0000 (13:18 +0900)]
Fix error during building custom op app for pacl (#7329)
Only nuerun supports custom op.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Tue, 10 Sep 2019 02:10:45 +0000 (05:10 +0300)]
[custom op] Add example application (#7313)
* [custom op] Add example application
Add example application using FillFrom kernel and FillFrom nnpackage
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
* fix wrong kernel name
Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Tue, 10 Sep 2019 02:10:01 +0000 (05:10 +0300)]
[custom op] Add example nnpackage (#7312)
Adds example model using FillFrom custom operation
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 10 Sep 2019 01:47:25 +0000 (10:47 +0900)]
[moco-tf] Use shape_inference_done for fix_shape COpCall (#7325)
This will update fix_shape for COpCall to use shape_inference_done helper function
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 10 Sep 2019 01:46:16 +0000 (10:46 +0900)]
[exo-tflite] Rename ConvertAdd (#7323)
This will rename ConvertAdd to loco IR name ConvertEltwiseAdd
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 10 Sep 2019 01:34:41 +0000 (10:34 +0900)]
[locomotiv] Support TensorBroadcast (#7309)
* [locomotiv] Support TensorBroadcast
This commit extends locomotiv to support TensorBroadcast nodes.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Fix a copy-and-paste error
* Check input shape
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 10 Sep 2019 01:32:40 +0000 (10:32 +0900)]
[loco] TransposedConv2D canonical IR (#7269)
* [loco] TransposedConv2D canonical IR
This commit introduces TransposedConv2D, loco canonical IR for
spatial transposed convolution.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
* Reorder alphabetically
* Edit comment
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 10 Sep 2019 00:03:58 +0000 (09:03 +0900)]
[loco] Refactor eltwise_binary_node_shape (#7296)
This commit relocates eltwise_binary_node_shape helper in order to
reduce the amount of code changes required for API v2 migration.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 10 Sep 2019 00:03:46 +0000 (09:03 +0900)]
[loco] Do not invoke global shape_known helper (#7297)
This commit revises ForwardShapeInferenceAlgorithm to invoke internal
"shape_known" helper instead of global shape_known helper.
This change reduces the amount of code changes for API v2 migration.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 10 Sep 2019 00:03:28 +0000 (09:03 +0900)]
[bino] Remove bitwise-or operator overloading (#7300)
Recently introduced fipe provides more general function pipe
implementation.
Let's use the function pipe implementation in fipe!
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 9 Sep 2019 23:57:28 +0000 (08:57 +0900)]
[loco] Implement Type/Shape Inference for TensorBroadcast (#7306)
This commit implements type/shape inference for TensorBroadcast nodes.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 23:39:18 +0000 (08:39 +0900)]
[exo-tflite] Moving Knob check of each converter into convert.cpp (#7303)
* [exo-tflite] Moving Knob check of each converter into convert.cpp
Knob check of each converter was moved into `convert.cpp`.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* remove unused header
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 23:38:56 +0000 (08:38 +0900)]
[moco-tf] Delete moved to logo (#7311)
This will delete transforms that has been moved to logo
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 23:38:45 +0000 (08:38 +0900)]
[exo-tflite] Putting all converts header file into one header file (#7305)
* [exo-tflite] Putting all converts header file into one header file
Now Conversion.h will include all converter's header file. This is to keep `convert.cpp` clean.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* move conversion/conversions.h to src
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 23:38:31 +0000 (08:38 +0900)]
[exo-tflite] adding `const` to TFLNodes.h (#7307)
This adds missing `const` to TFLNodes.h
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 23:34:45 +0000 (08:34 +0900)]
[exo-tflite] type inference of TFLAveragePool2D (#7314)
This commit adds type inference of TFLAveragePool2D.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 23:34:29 +0000 (08:34 +0900)]
[exo-tflite] change knob name for loco::AvgPool2D (#7316)
Previous name `ConvertAveragePool2D` followed TFL node name.
So, renamed to ConvertAvgPool2D just like converter's name follows canonical name.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Mon, 9 Sep 2019 11:26:23 +0000 (14:26 +0300)]
[custom op] Add FillFrom kernel example (#7310)
Add example for kernel developers
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Mon, 9 Sep 2019 10:34:18 +0000 (13:34 +0300)]
[custom op] Add basic custom op example structure (#7308)
* [custom op] Add basic custom op example structure
Add directories and CMakeLists for apps and kernels
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
* use add_subdirectories
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 9 Sep 2019 10:18:11 +0000 (19:18 +0900)]
[cli] Use GTest_AddTest (#7286)
Let's simplify CMakeLists.txt a bit with GTest_AddTest helper.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 10:14:25 +0000 (19:14 +0900)]
[moco-tf] Apply shape_inference_done (#7304)
This will apply to use shape_inference_done() for fix_shape nodes
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 10:00:51 +0000 (19:00 +0900)]
[logo] Unit test for ConstantFoldingPass (#7301)
* [logo] Unit test for ConstantFoldingPass
This will add unit test for ConstantFoldingPass copied from moco-tf ConstantFoldingTransform excluing import and log
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* apply comments
윤지영/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 9 Sep 2019 09:53:30 +0000 (18:53 +0900)]
[backend/srcn] Add transpose conv kernel (#7248)
* [backend/srcn] Add transpose conv kernel
Srcn library supports raw-major(NCHW) and col-major(NHWC) image layout.
However this commit only supports col-major image layout.
And Srcn library takes HWOI kernel filter layout when the image layout is NHWC.
But in neurun, it takes OHWI kernel filter layout when using the same image layout.
So it need to permute codes for TransposeConv weight.
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
* Change the Permute function to be limited to kernel layout
* Use original data array without copying to new address
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 9 Sep 2019 09:37:11 +0000 (18:37 +0900)]
[locomotiv] Use const if possible (#7264)
"annot_data" and "annot_domain" do not change the node itself.
This commit revies the signature of these helpers to express this
property explicitly.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>