윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 21 Oct 2019 06:30:47 +0000 (15:30 +0900)]
[exo] turning on a knob for FeatureBiasAdd conversion (#8334)
This turns on a knob for FeatureBiasAdd conversion.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 21 Oct 2019 06:30:36 +0000 (15:30 +0900)]
[exo] Turning on knob to enabling fusion (of add/sub, relu/relu6) by default (#8333)
This turns on knobs to enabling fusion (of add/sub, relu/relu6) by default
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 21 Oct 2019 06:30:26 +0000 (15:30 +0900)]
[loco] DepthwiseFilterDecode canonical IR (#8332)
This commit introduces DepthwiseFilterDecode loco canonical IR
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 21 Oct 2019 06:26:13 +0000 (15:26 +0900)]
[tf2tflite] added NET_0035-37 (CBR related test) into test.lst (#8343)
This adds NET00_35, NET_0036, NET_0037 (CBR related test) into test.lst
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 21 Oct 2019 06:24:42 +0000 (15:24 +0900)]
[exo] DepthwiseConv2D TFL dialect IR (#8320)
* [exo] DepthwiseConv2D TFL dialect IR
This commit introduces TFLDepthwiseConv2D, a TFL dialect IR for
DepthwiseConv2D. It also introduces its summary.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
* Add default value and test
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 21 Oct 2019 06:21:57 +0000 (15:21 +0900)]
[moco-integration-test] adding NET_0035, NET_0037 into test.lst (#8346)
This adds NET_0035, NET_0037 into test.lst.
0036 fails. An issue will be created.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 21 Oct 2019 05:59:21 +0000 (14:59 +0900)]
[nnfw-api] Merge custom op API implementation (#8340)
Move custom op API implementation into neurun/api
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 21 Oct 2019 05:45:08 +0000 (14:45 +0900)]
Update BuildGraph parameter (#8337)
Current BuildGraph get subgraph parameter, not interpreter
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 21 Oct 2019 05:26:59 +0000 (14:26 +0900)]
[moco] Import Nodes (#8336)
This will add Import GraphBuilder Nodes to import TensorFlow NodeDef
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 21 Oct 2019 05:26:24 +0000 (14:26 +0900)]
[loco] PemutingDecoder for DepthwiseFilter (#8331)
* [loco] PemutingDecoder for DepthwiseFilter
This commit introduces PemutingDecoder for DepthwiseFilter
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
* Remove unrelated
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 21 Oct 2019 04:48:58 +0000 (13:48 +0900)]
[moco-tf] use mio-tf (#8328)
* [moco-tf] use mio-tf
This will let moco-tf to use mio-tf so that one TensorFlow protobuf library produced and used
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* check mio_tf cond
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 21 Oct 2019 04:46:59 +0000 (13:46 +0900)]
[moco] Import requires (#8339)
This will update Import requires file
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 21 Oct 2019 03:13:46 +0000 (12:13 +0900)]
[logo] Update SimplifyDomainConversionPass comment (#8323)
This commit updates SimplifyDomainConversionPass comment to cover its
whole functionality
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 21 Oct 2019 03:12:43 +0000 (12:12 +0900)]
[exo] Update make_filter_* comment (#8322)
This commit fixes typo and updates following helpers in GraphBlock.h:
- make_filter_encode
- make_filter_decode
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 21 Oct 2019 03:11:05 +0000 (12:11 +0900)]
[plier-tf] use mio-tf (#8329)
* [plier-tf] use mio-tf
This will let plier-tf to use mio-tf so that one TensorFlow protobuf library is produced
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* check mio_tf cond
* match cmake if endif
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 21 Oct 2019 03:09:08 +0000 (12:09 +0900)]
[moco] Convert and TestHelper (#8330)
This will introduce Convert for Import and TestHelper for Import testing
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 21 Oct 2019 02:36:52 +0000 (11:36 +0900)]
[nnpkg, circle] Define instance_norm in circle_schema.fbs (#8262)
It introduces instance_norm operator in circle schema.
Also, it updates circle_schema_generated.h for circle_loader.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 21 Oct 2019 01:59:19 +0000 (10:59 +0900)]
[moco-tf] Redirect GraphBuilder to moco (#8327)
This will update GraphBuilder and GraphBuilderContext to redirect to moco for migration
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 21 Oct 2019 01:45:22 +0000 (10:45 +0900)]
Relocate compute libraries (#8307)
Rename srcn to ncnn
Relocate compute libraries into compute/ (armcompute-ex, cker, ncnn)
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 21 Oct 2019 01:43:12 +0000 (10:43 +0900)]
[loco] DepthwiseFilterDecoder (#8319)
This commit introduces DepthwiseFilterDecoder class.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 21 Oct 2019 01:27:27 +0000 (10:27 +0900)]
[moco] Introduce TFPlaceholder IR (#8313)
* [moco] Introduce TFPlaceholder IR
This will introduce TFPlaceholder IR for Placeholder node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* add todo comment
이춘석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 21 Oct 2019 01:01:10 +0000 (10:01 +0900)]
Support Logistic in base loader (#7942)
- Add logistic operation on base loader
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Sun, 20 Oct 2019 21:24:46 +0000 (06:24 +0900)]
[loco] Fix typo for DepthwiseFilterEncoder comment (#8318)
This commit fixes typo in DepthwiseFilterEncode comment
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Sun, 20 Oct 2019 21:19:23 +0000 (06:19 +0900)]
[moco] Include moco import (#8316)
This will update CMake to include import as sub project
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
[mir] Refine operations constructors. (#8283)
Removed obsolete operations constructors in [mir].
Signed-off-by: Dmitry Gusev <d.gusev@partner.samsung.com>
[mir] Remove Operation::Input class (#8294)
Remove `Operation::Input` class and add lightweight `Use` class.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir] Remove unused functions (#8324)
Remove `replaceWithInputNode` and `replaceInputNodes` functions which were only used in unittests.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir] Introduce new fused bias constructors (#8186)
* Conv2DOp ctor (and fix copy)
* DepthwiseConv2DOp ctor (and fix copy)
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 18 Oct 2019 09:37:30 +0000 (18:37 +0900)]
[base-loader] Change SubNode to MulNode in loadMul (#8311)
It fixes wrong code in loadMul.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 18 Oct 2019 07:24:39 +0000 (16:24 +0900)]
[moco-tf] fix PadCanonicalizer (#8298)
* [moco-tf] fix PadCanonicalizer
This commit fixs PadCanonicalizer in moco-tf.
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
* add comment.
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 18 Oct 2019 07:11:38 +0000 (16:11 +0900)]
[moco] Import: Introduce GraphBuilder (#8306)
This will introduce moco-import target with GraphBuilder and related classes copied from moco-tf
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 18 Oct 2019 07:11:17 +0000 (16:11 +0900)]
[moco-tf] Refactor TF node graph builder test part 2 (#8297)
This will refactor test NOT to use Importer module of node graph builder from MaxPool to Tanh
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 18 Oct 2019 07:10:48 +0000 (16:10 +0900)]
[moco-tf] Refactor TF node graph builder test part 1 (#8296)
This will refactor test NOT to use Importer module of node graph builder from Add to FusedBatchNorm
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 18 Oct 2019 05:08:20 +0000 (14:08 +0900)]
[exo] replace PhaseStrategy of TestHelper.h from Saturate to Restart (#8304)
This replaces PhaseStrategy of TestHelper.h from Saturate to Restart
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 18 Oct 2019 05:07:29 +0000 (14:07 +0900)]
[res/TensorFlowTests] NET_0037 test case (#8302)
* [res/TensorFlowTests] NET_0037 test case
This adds NET_0037, Conv2D + ReLU + Add. By exo, Conv2D and ReLU are fused but Add isn't.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* simple description
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 18 Oct 2019 04:28:40 +0000 (13:28 +0900)]
[nnfw-gbs] Merge build require and option setting (#8232)
- Remove python3 build require
- Require libarmcl-devel and python on aarch target only
- Merge build option setting for coverage build, test build and general build
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 18 Oct 2019 04:26:00 +0000 (13:26 +0900)]
[exo] adding FuseReluPass into TFLOptimize.cpp (#8280)
This adds FuseReluPass into TFLOptimize.cpp. When a knob is on, this pass will run.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 18 Oct 2019 04:23:12 +0000 (13:23 +0900)]
[exo] rewriting a function to handle Mixin (#8281)
Instread of trying to cast to a lot of type, code was re-written to try one-time cast to Mixin.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 18 Oct 2019 04:19:49 +0000 (13:19 +0900)]
[lib/tflite] Update operation register (#8278)
Update register.h/cpp to match tflite v1.13.1
Support new operations introduce on v1.13.1
Support more upstream custom operation
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 18 Oct 2019 03:27:40 +0000 (12:27 +0900)]
[exo] FuseConv2DAddSubPass checks Conv2D's fused activation function (#8268)
* [exo] FuseConv2DAddSubPass checks Conv2D's fused activation function
This enables FuseConv2DAddSubPass to check Conv2D's fused activation function
and perform fusion considering the status of fused activation function.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* remove unused var
* test case
* one more test case
* float act_func(...) -> void act_func(..)
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 18 Oct 2019 02:20:46 +0000 (11:20 +0900)]
[mio-tf] TensorFlow model library (#8274)
* [mio-tf] TensorFlow model library
This will introduce mio-tf as a separate project for accessing TensorFlow protobuf model file
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* merge cmake and apply comments
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 18 Oct 2019 02:09:37 +0000 (11:09 +0900)]
[neurun] Backend priority by OP_BACKEND_ALLOPS (#8124)
- Try backend loading from default backend setting, but not fail if backend is not exist
- Fail if there is no loaded backend
- Reorder to load acl(cl, neon) first
- If OP_BACKEND_ALLOPS is empty, use first loaded backend
- Set default OP_BACKEND_ALLOPS emptyy
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 18 Oct 2019 00:48:40 +0000 (09:48 +0900)]
Rename moco-onnx as oneco (#8251)
Let's use moco only for TensorFlow Frontend modules.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 18 Oct 2019 00:42:46 +0000 (09:42 +0900)]
[tf2tflite] Set shape information in ModelSignature (#8290)
This commit will enable setting shape information in `ModelSignature`
at `tf2tflite`
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 18 Oct 2019 00:42:38 +0000 (09:42 +0900)]
[nnkit-mocotf] Set shape information in ModelSignature (#8289)
This commit will enable setting shape information in `ModelSignature`
at `nnkit-mocotf`
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 17 Oct 2019 23:33:31 +0000 (08:33 +0900)]
[res] Tensorflow test file for Relu( Add( Conv2D, Scalar )) (#8282)
This adds Conv2D + Add + ReLU where Add's Const operand is a scalar (size 1 tensor).
To check exo handles this OK.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 17 Oct 2019 23:19:27 +0000 (08:19 +0900)]
[moco-integration-test] Enable MergeConcatPass tests (#8288)
This commit will enable `MergeConcatPass` related tests
in `moco-integration-test`
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 17 Oct 2019 22:16:06 +0000 (07:16 +0900)]
[tf2tflite] Enable MergeConcatPass tests (#8287)
This commit will enable `MergeConcatPass` related tests
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 17 Oct 2019 22:15:42 +0000 (07:15 +0900)]
[exo] Enable ConvertTensorConcat knob (#8286)
This commit will enable `ConvertTensorConcat` knob in `exo`
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
[ncc/interpreter] Fix Reshape operation for quantization (#8246)
* Made Reshape independent from DataType
* Fixed setting output quantization
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
[nnc] Fix constant folding optimization for quantization (#8241)
* Fixed transpose for different DataType
* Added qunatization setting on optimization
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
[nnc] Fix ModelAnalyzer.linearization test (#8291)
The test relied on specific order of graph traversal.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 17 Oct 2019 10:15:42 +0000 (19:15 +0900)]
[moco] Update input accessor (#8279)
This will update to use correct input accessor for MaxPool and FusedBatchNorm node in moco
and remove temporary added accessors
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 17 Oct 2019 10:15:17 +0000 (19:15 +0900)]
[exo] Add tests for TFLConcatenate (#8273)
This commit will add tests for `TFLConcatenate` in `exo`
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 17 Oct 2019 10:14:33 +0000 (19:14 +0900)]
[moco-tf] Introduce TFNodeBuildTester and apply BiasAdd test (#8272)
This will introduce TFNodeBuildTester test helper class for testing node's GraphBuilder class.
And applied BiasAdd test code is also shown to show how it works.
This change purpose is not to use Import class and to test only the GraphBuilder that
converts TensorFlow NodeDef to TF dialect node.
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 17 Oct 2019 10:14:15 +0000 (19:14 +0900)]
[exo] Introduce MergeConcatNodePass (#8264)
* [exo] Introduce MergeConcatNodePass
This commit will introduce `MergeConcatNodePass`,
which fold multiple `TFLConcatenation` nodes with same axis
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
* apply comments
* add FusedActFunc check
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 17 Oct 2019 10:04:16 +0000 (19:04 +0900)]
[locomotiv] Introduce TensorConstantPad (#8027)
* [locomotiv] Introduce TensorConstantPad
This commit introduces TensorConstantPad to locomotiv.
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
* change algorithm of pad interpretation.
* make constant value more general.
* change assert to validate.
* modify test.
* add assert for pad_data.
* constant input can be all operator.
* fix some wrong valiate and message.
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 17 Oct 2019 09:58:56 +0000 (18:58 +0900)]
[moco-tf] Remove Shape Type inference header (#8270)
This will remove deprecated Shape and Type inference header use that of moco
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 17 Oct 2019 09:40:50 +0000 (18:40 +0900)]
[moco-tf] Update input accessor (#8276)
This will update to use correct input accessor for MaxPool and FusedBatchNorm node and related variables
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 17 Oct 2019 08:27:09 +0000 (17:27 +0900)]
[tool] Support SquaredDifferene in select operator (#8263)
- Support SquaredDifferene in select operator
- Add comment: not supporting options on schema v1.13.1
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 17 Oct 2019 07:38:31 +0000 (16:38 +0900)]
[exo] Register TransposedConv2DConverter (#8271)
This commit registers TransposedConv2DConverter to convert stage.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 17 Oct 2019 07:23:06 +0000 (16:23 +0900)]
[exo] adds loop to fuse the nodes for FuseReluPass (#8266)
This completes FuseReluPass.cpp by add loops that performs fusing selected nodes.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 17 Oct 2019 06:49:36 +0000 (15:49 +0900)]
[plier-tf] Introduce parse_nodedef (#8261)
This will introduce parse_nodedef() method to parse tensorflow::NodeDef object
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 17 Oct 2019 06:49:26 +0000 (15:49 +0900)]
[exo] Introduce TFLConcatenation exporter (#8259)
This commit will introduce `TFLConcatenation` exporter in `exo`
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 17 Oct 2019 06:35:25 +0000 (15:35 +0900)]
[exo] Convert canonical TransposedConv2D (#8267)
This commit introduces convert stage for canonical TransposedConv2D.
This conversion would apply to TransposedConv2D to make TFL-dialect
TFLTransposeConv with additional auxiliary nodes.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 17 Oct 2019 06:26:13 +0000 (15:26 +0900)]
[exo] Export TFLTransposeConv (#8260)
This commit introduces export stage of TFLTransposeConv operator
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 17 Oct 2019 06:25:28 +0000 (15:25 +0900)]
[moco-tf] Redirect to moco Service (#8265)
This will update Shape and Type inference class to be redirected to moco Service and remove unused source files
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 17 Oct 2019 05:28:38 +0000 (14:28 +0900)]
[exo] TransposeConv type & shape inference (#8258)
This commit introduces type and shape inferences of TFLTransposeConv
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 17 Oct 2019 04:41:16 +0000 (13:41 +0900)]
[nnpkg-spec] Add instance_norm and revise structure and level (#8148)
It updates nnpackage specification to define `instance_norm` operator.
Also, it updates the document structure since it begin to have additional
operator. I move operators to same level of model for user to find them easily.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 17 Oct 2019 04:28:44 +0000 (13:28 +0900)]
[exo] adding TFLRelu6Converter into convert.cpp (#8253)
This adds TFLRelu6Converter into convert.cpp with related knob on.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 17 Oct 2019 04:18:11 +0000 (13:18 +0900)]
[exo] TransposeConv IR (#8237)
* [exo] TransposeConv IR
This commit introduces TFLTransposeConv IR, TFL-dialect of loco node
representing TFlite TransposedConv operator.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
* Use camel case member name
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 17 Oct 2019 03:19:23 +0000 (12:19 +0900)]
[res] Add simple network of placeholder with unknown dimension. (#8229)
It is a simple network of placeholder with unknown dimension.
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 17 Oct 2019 03:18:38 +0000 (12:18 +0900)]
[exo] refactoring using new function (`create_fused_bias()`) (#8230)
This replaces body with a statement that calls extracted function. (`create_fused_bias()`)
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 17 Oct 2019 02:15:45 +0000 (11:15 +0900)]
[lib/tflite] Update nnapi delegate (#8238)
Update nnapi delegate to match with v1.13.1
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
여지환/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 17 Oct 2019 01:50:06 +0000 (10:50 +0900)]
Rename srcn backend shape (#8120)
Rename srcn backend Shape class into TensorDescriptor
- Use same class name as cpu backend
- Rename variables according to new class name
Signed-off-by: JiHwan Yeo <jihwan.yeo@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 17 Oct 2019 01:07:01 +0000 (10:07 +0900)]
[moco] Introduce Shape and Type inference Service (#8228)
* [moco] Introduce Shape and Type inference Service
This will introduce Shape and Type inference Service from copied from moco-tf
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* make compiler happy
* another for compiler
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 17 Oct 2019 00:22:20 +0000 (09:22 +0900)]
[neurun] Introduce build option for default config (#8167)
- Introduce ENVVAR_NEURUN_CONFIG option: use default neurun config using environment (default: on)
- Default off on tizen, and introduce gbs build option "test_build" to enable environment variable setting
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
[nnc] Explicitly declare some variables as `Operation::Input *` (#8247)
This is needed for transition period.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir-tflite-importer] Fixed a bug (wrong Deconv2d constructir). (#8248)
Setting attributes values is corrected.
Signed-off-by: Dmitry Gusev <d.gusev@partner.samsung.com>
[mir-caffe-importer] Fixed a bug in CaffeOpCreator (#8243)
Setting attributes values is corrected.
Signed-off-by: Dmitry Gusev <d.gusev@partner.samsung.com>
[nnc/interpreter] Fix Transpose for quantization (#8242)
* Fixed quantization for output
* Fixed work with every DataType
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
[mir2loco] Replace getInputProducer with getInput (#8245)
Replace calls to `getInputProducer` with equivalent `getInput`.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[nnc] Replace getInputProducer with getInput (#8244)
Replace calls to `getInputProducer` with equivalent `getInput`.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir] Make getInput the same as getInputProducer (#8236)
Make `getInput` the same as `getInputProducer`. The latter will be removed.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir-onnx-importer] Use new operations interface in importer. (#8182)
* Using new operations interface in importer.
Changes due to new operations constructors signatures.
Signed-off-by: Dmitry Gusev <d.gusev@partner.samsung.com>
[mir2loco] Switch to `getInputProducer` (#8190)
Replace calls to `getInput()->getProducer()` with calls to `getInputProducer()`.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[nnc] Switch to `getInputProducer` method (#8189)
Replace calls to `getInput()->getProducer()` with calls to `getInputProducer()`.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 16 Oct 2019 10:40:57 +0000 (19:40 +0900)]
[angkor] Overload operator== (#8233)
* [angkor] Overload operator==
This commit overloads operator== to index class.
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
* apply commnet.
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 16 Oct 2019 09:50:16 +0000 (18:50 +0900)]
[angkor] Overload operator+ (#8210)
* [angkor] Overload operator+
This commit overloads operator+ to Index class.
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
* add out_of_range thorw exception.
* remove unnecessary header.
* change error type.
* change operator+ from member function to non-member.
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 16 Oct 2019 09:09:17 +0000 (18:09 +0900)]
locomotiv is static library (#8212)
This commit changes locomotiv from shared to static library.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 16 Oct 2019 08:27:19 +0000 (17:27 +0900)]
[moco-tf] Remove VariadicArityNode (#8227)
This will remove not used anymore VariadicArityNode and its test
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 16 Oct 2019 08:23:47 +0000 (17:23 +0900)]
[moco-tf] Remove IR tests (#8225)
This will remove IR test codes that already moved to moco-lang
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 16 Oct 2019 08:03:58 +0000 (17:03 +0900)]
[exo] Introduce TFLConcatenation converter (#8221)
This commit will introduce `TFLConcatenation` converter in `exo`
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 16 Oct 2019 08:03:47 +0000 (17:03 +0900)]
[exo] Refactoring: Extract code from big method (Fuse) (#8179)
A piece of code is extracted as a method (`create_fused_bias()`) from big method (`Fuse`).
For diff difficulty, old code will be replace with one single call in next PR.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 16 Oct 2019 07:53:35 +0000 (16:53 +0900)]
[exo] TFLMixin for classes that have fused activation function (#8223)
This adds TFLMixin class for TFL nodes that have fused activation function
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 16 Oct 2019 07:51:49 +0000 (16:51 +0900)]
[nnfw-api] use circle loader for circle model (#8222)
It uses model_types from MANIFEST to determine loader.
It also added missing namespace neurun for circle_loader.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 16 Oct 2019 07:51:33 +0000 (16:51 +0900)]
[nnpkg-test] Remove conv test (#8220)
Remove fail conv test: neurun cannot run convolution with input weight
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 16 Oct 2019 07:48:34 +0000 (16:48 +0900)]
[exo] TODO rigorous padding conversion (#8219)
This commit adds TODO comments for rigorous padding conversion
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 16 Oct 2019 07:48:10 +0000 (16:48 +0900)]
[exo] Introducing FuseReluPass (#8176)
* [exo] Introducing FuseReluPass
This introduces FuseReluPass, which fuse relu/relu6 into 10 TensorFlow Lite ops.
This commit adds code that finds candidate subgraph.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* remove unused var (to avoid CI error)
* Update compiler/exo/src/Pass/FuseReluPass.h
Co-Authored-By: hyunsik-yoon <hyunsik.yoon@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 16 Oct 2019 07:39:12 +0000 (16:39 +0900)]
[moco-tf] Revise IR to using moco::IR (#8218)
This will revise moco::tf IR and base classes to alias of moco IR and base classes
Signed-off-by: SaeHie Park <saehie.park@samsung.com>