박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 15 Jul 2019 01:36:41 +0000 (10:36 +0900)]
[moco/tf] Revise ReLU6 unit test (#4247)
This will revise ReLU6 unit test to test only related properties
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 15 Jul 2019 01:36:22 +0000 (10:36 +0900)]
[moco/tf] Revise ReLU unit test (#4246)
This will revise ReLU unit test to test only related properties
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 15 Jul 2019 01:36:04 +0000 (10:36 +0900)]
[moco/tf] Revise MaxPool2D unit test (#4245)
This will revise MaxPool2D unit test to test only related properties
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 15 Jul 2019 01:35:23 +0000 (10:35 +0900)]
[moco/tf] Revise BiasAdd unit test (#4243)
This will revise BiasAdd unit test to test only related properties
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 15 Jul 2019 01:22:10 +0000 (10:22 +0900)]
[locoex/customop] Attr types and data structure (#4203)
* [locoex/customop] Attr types and data structure
This file contains attribute types and structs to hold the value of attr types.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* renaming: attrType() -> type()
* removed customop dir and namespace
* remvoe attrTypeTraits. This will be committed later
* Rename class name
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 15 Jul 2019 01:21:53 +0000 (10:21 +0900)]
[locoex/customop] Dialect for Customop (#4209)
* [locoex/customop] Dialect for Customop
This adds Dialects declaration for custom op.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* remove 'customop' dir and milestone. change class name.
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 15 Jul 2019 01:20:20 +0000 (10:20 +0900)]
[exo/tflite] Do NOT export Identity (#4237)
This commit revises TensorAllocation to allocate the same operand
to the input and output of "identity"-like operations, and revises
OperationExporter not to export such operations.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 15 Jul 2019 01:18:48 +0000 (10:18 +0900)]
Introduce TensorFlowProtoText 1.12 package (#4230)
* Introduce TensorFlowProtoText 1.12 package
This commit introduces TensorFlowProtoText 1.12 package, which provies
"proto_text" executable (with a different name) required for TensorFlow
build.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Be compatibile with CMake 3.1
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 15 Jul 2019 01:09:24 +0000 (10:09 +0900)]
Introduce and use import function for configuraiton loading (#4206)
This commit revises nncc & command scripts not to depend on
"NNCC_SCRIPT_PATH" variable.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 15 Jul 2019 01:05:56 +0000 (10:05 +0900)]
[moco/tf] Fix macro undef (#4210)
This will fix another wrong undef in FixPaddingTransform
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 15 Jul 2019 00:56:04 +0000 (09:56 +0900)]
[moco/tf] Introduce GraphBuilderSource interface (#4235)
This commit extracts GraphBuilderSource from GraphBuilderRegistry
which supports only lookup method.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 15 Jul 2019 00:35:23 +0000 (09:35 +0900)]
[moco_tf] helper of test to find nodes by type (#4241)
This will add find_nodes_bytype() for unit tests to find nodes by type
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Sun, 14 Jul 2019 23:33:55 +0000 (08:33 +0900)]
[moco/tf] Introduce find node by type test helper (#4226)
* [moco/tf] Introduce find node by type test helper
This will introduce find_node_bytype in test helper to find a node with a specific type
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* rename and include header
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Sun, 14 Jul 2019 23:28:34 +0000 (08:28 +0900)]
[moco/tf] Enable FeatureDecode-ReLU reorder by default (#4228)
It turns out that "FeatureDecode-ReLU" reorder transformation is
compatible with all the current implementation.
Let's enable this transformation by default.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 12 Jul 2019 07:57:58 +0000 (16:57 +0900)]
[exo/tflite] Support FeatureBiasAdd (#4229)
T/F Lite exporter now accepts FeatureBiasAdd node.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 12 Jul 2019 07:49:20 +0000 (16:49 +0900)]
Stop CMake immediately on GTest configure/build error (#4232)
Now, CMake stops immediately on GTest configure/build fails.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 12 Jul 2019 07:49:10 +0000 (16:49 +0900)]
Stop CMake immediately on FlatBuffer configure/build error (#4231)
Now, CMake configuration stops immediately if FlatBuffer configure/build
fails.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 12 Jul 2019 05:49:54 +0000 (14:49 +0900)]
[moco/tf] Import as TFConv2D (#4223)
This will enable import as TFConv2D
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 12 Jul 2019 05:32:04 +0000 (14:32 +0900)]
[exo/tflite] Export T/F Lite operation in exeuction order (#4222)
* [exo/tflite] Export T/F Lite operation in exeuction order
The current implementations generates T/F Lite operators following node
generation order which may differ from actual execution order.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Apply patch
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 12 Jul 2019 05:14:26 +0000 (14:14 +0900)]
[moco/tf] Selective Reorder (#4216)
This commit allows users to control whether to apply reorder transform
per operation.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 12 Jul 2019 05:14:15 +0000 (14:14 +0900)]
[exo/tflite] Remove commented function declarations (#4218)
These functions declarations are no longer used, but just commented
to reduce the amount of code diff.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 12 Jul 2019 05:13:09 +0000 (14:13 +0900)]
[exo/tflite] Export Tensors after Allocation (#4199)
This commit rewrites exportOpDefinedTensors as a 2-stage function. Now,
exportOpDefinedTensors first allocates a T/F Lite tensor first, and then
export allocated T/F Lite tensors.
The first phase iterates over loco nodes, but the second phase now
iterates over allocated T/F Lite tensors.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 12 Jul 2019 05:06:40 +0000 (14:06 +0900)]
[mocotest/tf] Use CMake include (#4214)
This commit revises how to add moco a T/F testcase. This change allows
users to comment out some testcases temporarily.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 12 Jul 2019 04:50:01 +0000 (13:50 +0900)]
[moco/tf] Move Transform.test.cpp (#4217)
Transform.cpp is in the top-level source directory, but Transform.test.cpp
is currently inside Transforms directory.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 12 Jul 2019 04:47:07 +0000 (13:47 +0900)]
[locoex/customop] Introducing VariadicArityNode (#4189)
* [locoex/customop] Introducing VariadicArityNode
In case of custom op, we don't know how many inputs custom op has.
Because of this, previous we need this (previous FixedArityNode cannot handle this situation)
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* unsigned to uint32_t, remove _arity member variable and duplicate assert()
* test case
* removed customop dir and namespace
* use unused var pull_02
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 12 Jul 2019 04:11:43 +0000 (13:11 +0900)]
Update git ignore file (#4173)
Cleanup .gitignore file to manage files generated by official tool only
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 12 Jul 2019 02:38:24 +0000 (11:38 +0900)]
[res] TF_SMALL_NET_0008 with Conv2D + Add (#4207)
This will introduce TF_SMALL_NET_0008 having Conv2d and Add node
This is to check FuseBinaryIntoPreceding transformation that fuses Add into Conv2D + BiasAdd
where BiasAdd will be created with 0 bias if not exist
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 12 Jul 2019 00:37:14 +0000 (09:37 +0900)]
[locop] Introduce CanonicalNodeSummaryBuilder (#4205)
* [locop] Introduce CanonicalNodeSummaryBuilder
This commit introduces CanonicalNodeSummaryBuilder, which allows users
to embed CanonicalNodeSummryBuilder into their own NodeSummaryBuilder.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Update example
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 12 Jul 2019 00:14:53 +0000 (09:14 +0900)]
Stop CMake immediately on Protobuf configure/build failure (#4215)
The current implementation just returns with STATUS message.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 11 Jul 2019 09:49:03 +0000 (18:49 +0900)]
[locop] Add license notice (#4213)
There is no license notice in FormattedGraph.h and ForamttedGraph.cpp.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 11 Jul 2019 06:56:59 +0000 (15:56 +0900)]
[moco/tf] Show lines in transformation log (#4190)
This will show lines in PhaseRunner for better readability in between transforms and phase
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 11 Jul 2019 06:56:39 +0000 (15:56 +0900)]
[res] TF_SMALL_NET_0007 with Conv2D+Mul (#4200)
This will introduce TF_SMALL_NET_0007 having Conv2D and Mul node
This is to check FuseBinaryIntoPreceding transformation that fuses Mul into Conv2D kernel and be removed
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 11 Jul 2019 06:56:11 +0000 (15:56 +0900)]
[moco/tf] TFConv2D to Conv2D Canonicalizer (#4202)
This will implement TFConv2D to Conv2D Canonicalizer
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 11 Jul 2019 06:55:44 +0000 (15:55 +0900)]
[moco/tf] Fix macro undef (#4201)
This will fix macro undef to right name
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 11 Jul 2019 06:17:50 +0000 (15:17 +0900)]
[locop] Extract canonical_node_desc from node_desc (#4198)
This commit extracts canonical_node_desc helper from node_desc in order
to make it easy to reuse.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 11 Jul 2019 04:59:37 +0000 (13:59 +0900)]
[tf2tflite] Enable Placeholder_000 and Multiple_IO_002 (#4196)
Recent changes on moco-tf and exo-tflite resolve Placeholder_000 and
Multiple_IO_002 test failure.
- moco-tf now eliminates Forward by default, which resolves
Placeholder_000 failure.
- exo-tflite now honor the input/output order, which resolves
Multiple_IO_002 failure.
This commit updates tf2tflite test framework to check Multiple_IO_002
and Placeholder_000 by default to prevent regression.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 11 Jul 2019 03:47:45 +0000 (12:47 +0900)]
[moco/tf] Use PadData and StrideData for TFConv2d (#4163)
* [moco/tf] Use PadData and StrideData for TFConv2d
This will fix TFConv2D related FixPadding and FixShare transformations to use PadData and ShapeData to simplify Canonicalization
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* check data_layout
* fix log message
* fix log message
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 11 Jul 2019 01:37:44 +0000 (10:37 +0900)]
[moco/tf] Add fix for shape and padding (#4188)
This will add fix for shape and padding after TF transformations that can add new nodes with empty shape and padding
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 11 Jul 2019 01:35:28 +0000 (10:35 +0900)]
[moco/tf] Remove unused Canonicalization files (#4191)
This will remove unused Canonicalization files that are replaced by the files in Canonicalization folder
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 11 Jul 2019 01:21:33 +0000 (10:21 +0900)]
[exo/tflite] Hide ShapeContext (#4187)
This commit makes ShapeContext visible only to ShapeInference module,
and revises all the other modules to use "ShapeInferece::get".
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 11 Jul 2019 00:33:39 +0000 (09:33 +0900)]
Add Multiple_IO_002 TensorFlow test (#4181)
This commit adds Multiple_IO_002 test. This testcase allows us to
check what happens if there is a mismatch between node declataion order
and input order.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 10 Jul 2019 22:36:12 +0000 (07:36 +0900)]
[moco/tf] Add ClearAnnotTransform items (#4186)
This will add PadData and StrideData to ClearAnnotTransform transform
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 10 Jul 2019 08:34:58 +0000 (17:34 +0900)]
[loco] Remove make_dimension(uint32_t) (#4176)
* [loco] Remove make_dimension(uint32_t)
This commit removes make_dimension(uint32_t) helper from loco, and
updates relevant code.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Remove implementation
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 10 Jul 2019 08:33:29 +0000 (17:33 +0900)]
[moco/tf] Enable Stable Optimizations by default (#4183)
This commit enables the following optimizations by default.
- RemoveDeadNode
- SimplifyDomainConversion
- RemoveForwardNode
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 10 Jul 2019 08:33:12 +0000 (17:33 +0900)]
[exo/tflite] Honor Graph Input/Output order (#4184)
TFLExporter now creates TensorFlow Lite Graph Input/Output using
graph-level specification.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 10 Jul 2019 08:30:53 +0000 (17:30 +0900)]
[moco/tf] Knob for Canonicalize Conv2D (#4185)
This will add a knob for Canonicalize Conv2D node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 10 Jul 2019 08:12:30 +0000 (17:12 +0900)]
[loco] Graph::NodeContext::create with multiple parameters (#4182)
To support loco node with constructors that takes parameter, Graph::NodeContext::create now takes parameters.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 10 Jul 2019 05:56:50 +0000 (14:56 +0900)]
[locoex/customop] Change directory structure (#4153)
* [locoex/customop] Change directory structure
Moved already-merged file into the following directory structure
-- locoex-customop -+- include -- locoex -- customop -- header files (no sub-directory)
|
+- src - source files (no sub-directory)
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* rename subdir(namespace) customop to node
* renaming directory and node name
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 10 Jul 2019 05:55:39 +0000 (14:55 +0900)]
[moco] Remove empty directories (#4177)
Most of implementations are now in moco-tf and moco-onnx.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 10 Jul 2019 05:37:58 +0000 (14:37 +0900)]
[moco/tf] Introduce PadData (#4160)
* [moco/tf] Introduce PadData
This will introduce PadData used as annotation for TFConv2D pad
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* fix guard comment
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 10 Jul 2019 04:06:01 +0000 (13:06 +0900)]
[exo.tflite] Manage TFLTensorIndex as annotation (#4162)
The current implementation manages TFLTensorIndex via SubGraphContext,
which necessitates some-level of indirection.
This commit makes exo-tflite to annotate each node with its tensor
index.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 10 Jul 2019 04:05:48 +0000 (13:05 +0900)]
[moco/onnx] Remove make_dimension(uint32_t) usage (#4170)
This commit revises moco-onnx not to invoke deprecated make_dimension(uint32_t).
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 10 Jul 2019 04:05:38 +0000 (13:05 +0900)]
[locomotiv] Remove make_dimension(uint32_t) usage (#4171)
This commit revises locomotiv not to invoke deprecated
make_dimension(uint32_t).
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 10 Jul 2019 04:05:27 +0000 (13:05 +0900)]
[moco/tf] Remove make_dimension(uint32_t) usage (#4172)
This commit revises moco-tf not to invoke deprecated
make_dimension(uint32_t).
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 10 Jul 2019 03:59:31 +0000 (12:59 +0900)]
[locop] Extensible Node Formatting (#4156)
This commit introduces NodeSummaryBuilder & NodeSummaryBuilderFactory
interfaces.
These interfaces allow users to inject their own Node Formatting
implementation.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 10 Jul 2019 03:59:12 +0000 (12:59 +0900)]
[exo.tflite] Restrict TypeContext's visibility (#4159)
TypeInference is the only module that accesses TypeContext.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 10 Jul 2019 01:33:20 +0000 (10:33 +0900)]
[tf2tflite] Enable tests (#4154)
This commit enables additional tests for tf2tlite. Here is the list of tests to
be enabled:
- AvgPool_000
- AvgPool_001
- Concat_002
- Conv2D_000
- MaxPool_000
- Relu_000
- TF_SMALL_NET_0002
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 10 Jul 2019 00:54:01 +0000 (09:54 +0900)]
[moco/tf] Reorder Knob list by function (#4168)
* [moco/tf] Reorder Knob list by function
This will reorder Knob list items by the function
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* remove new one
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 10 Jul 2019 00:01:28 +0000 (09:01 +0900)]
Add Google NSync CMake package (#4161)
* Add Google NSync CMake package
This commit adds Google NSync CMake package.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Remove duplicated set
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 9 Jul 2019 22:07:06 +0000 (07:07 +0900)]
[moco/tf] Introduce StrideData (#4158)
This will introduce StrideData used as annotation for TFConv2D stride
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 9 Jul 2019 06:17:10 +0000 (15:17 +0900)]
[moco/tf] Introduce FuseBinaryIntoPreceding (#4150)
This will introduce empty FuseBinaryIntoPreceding transformation for fusing binary(i.e. mul, add) node to preceding node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 9 Jul 2019 04:48:31 +0000 (13:48 +0900)]
[exo/tflite] Extract SubGraphContext from SerializedModelData (#4151)
This commit extracts SubGraphContext from SerializedModelData.
This information related with T/F Lite SubGraph is necessary to
serialize graph, but not a part of serialized data.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 9 Jul 2019 04:42:36 +0000 (13:42 +0900)]
[locop] Expose NodeDesc class (#4152)
* [locop] Expose NodeDesc class
This commit exposes NodeDesc class as a public class as a step to
support user-defined node pretty-printer.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Remove unnecessary comment
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 9 Jul 2019 01:34:59 +0000 (10:34 +0900)]
[moco/tf] Enable Canonicalizer (#4148)
This will enable two existing two canonicalizer in transformation phase
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 9 Jul 2019 00:30:15 +0000 (09:30 +0900)]
[exo.tflite] Properly propagate type information (#4144)
The current implementation of type inference fails to properly propagate
type information.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 8 Jul 2019 22:07:52 +0000 (07:07 +0900)]
[moco/tf] Introduce CanonicalizeBiasAdd (#4139)
* [moco/tf] Introduce CanonicalizeBiasAdd
This will introduce CanonicalizeBiasAdd transformation that transforms Tensorflow dialect BiasAdd to Canonical dialect BiasAdd.
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* rename and move folder
* fix use return
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 8 Jul 2019 22:03:44 +0000 (07:03 +0900)]
[moco/tf] Processing the schama file for model metainfo file (#4137)
This commit adds schema file and CMakeLists for processing model metainfo file.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 8 Jul 2019 09:36:48 +0000 (18:36 +0900)]
[moco/tf] Fix to use return in Conv2DCanonicalizer (#4142)
This will fix to use return value in Conv2DCanonicalizer
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 8 Jul 2019 09:36:23 +0000 (09:36 +0000)]
[moco/tf] Introduce ReLU6 (#4131)
* [moco/tf] Introduce ReLU6
This commit introduce ReLU6 operation to moco.
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
* This commit apply for comments.
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 8 Jul 2019 09:17:55 +0000 (18:17 +0900)]
[moco/tf] Reorder FeatureDecode-TensorBiasAdd (#4127)
"ReorderDecode" is now able to rewrite "FeatureDecode-TensorBiasAdd" as
"FeatureBiasAdd-FeatureDecode".
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 8 Jul 2019 08:23:50 +0000 (17:23 +0900)]
[locop] Introduce SymbolTable interface (#4128)
This commit introduces "SymbolTable" interface as a step to make
locop framework extensible.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 8 Jul 2019 07:21:49 +0000 (16:21 +0900)]
[locop] Default-contructible NodeDesc (#4134)
This commit makes (internal) NodeDesc default-constructible.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 8 Jul 2019 07:21:11 +0000 (16:21 +0900)]
[exo/tflite] Extract Shape Inference (#4138)
This commit extracts shape inference phase from tensor export phase.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 8 Jul 2019 07:14:58 +0000 (16:14 +0900)]
[moco/tf] Introduce CanonicalizeConv2D (#4140)
* [moco/tf] Introduce CanonicalizeConv2D
This will introduce emptyCanonicalizeConv2D that transforms TensorFlow dialect Conv2D to Canonical Conv2D
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* rename and move folder
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 8 Jul 2019 07:10:50 +0000 (16:10 +0900)]
[locoex/customop] Initial commit for Custom Op (#4136)
* [locoex/customop] Initial commit for Custom Op
This is an initial commit for Custom Op. Parents for Custom Op will be added after successive PRs.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* revised
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 8 Jul 2019 02:45:27 +0000 (11:45 +0900)]
[moco/tf] Introduce ResolveFusedBatchNorm (#4125)
This will introduce ResolveFusedBatchNorm transformation that decomposes FusedBatchNorm node into Add and Mul
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 8 Jul 2019 01:50:37 +0000 (10:50 +0900)]
[exo.tflite] Extract Type Inference as a separate phase (#4126)
This commit extracts (Data) Type Inference from exportOpDefinedTensor
phase.
This chanage allows users to infer types without exporting tensors.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 8 Jul 2019 01:11:39 +0000 (10:11 +0900)]
[plier/tf] Move more functions from Moco::tf::convert.h & cpp (#4122)
The following was moved into plier.tf from moco/../convert.h & cpp
- two functions
- test case
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 8 Jul 2019 01:06:31 +0000 (10:06 +0900)]
[moco/tf] name attribute for transforms (#4121)
This will add name() attribute for existing Transforms
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 8 Jul 2019 00:55:21 +0000 (09:55 +0900)]
[moco/tf] simplify get_data_layout (#4124)
This will simplify get_data_layout to call as_DataLayout as common part
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 8 Jul 2019 00:14:54 +0000 (09:14 +0900)]
Add "GoogleDoubleConversion" package (#4112)
* Add "GoogleDoubleConversion" package
This commit introduces "GoogleDoubleConversion" which allows users to
import system-wide Google DoubleConversion library in nncc style.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Remove tab
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 8 Jul 2019 00:14:32 +0000 (09:14 +0900)]
[exo/tflite] Use loco::output_nodes in exportOpDefinedTensors (#4118)
exportOpDefinedTensors currently uses its own implementation of
loco::output_nodes.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 8 Jul 2019 00:11:36 +0000 (09:11 +0900)]
[moco_tf] Fix BiasAdd test indent todo (#4120)
This will fix indentation of BiasAdd test from previous change
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 8 Jul 2019 00:08:21 +0000 (09:08 +0900)]
[moco/tf] Conv2D test support TFConv2D (#4083)
* [moco/tf] Conv2D test support TFConv2D
This will split Conv2D test to support TFConv2D by the knob
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* include knob
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Sun, 7 Jul 2019 23:29:51 +0000 (08:29 +0900)]
[moco/tf] Rename FusedBatchNorm Graph builder (#4119)
This will rename GraphBuilder for FusedBatchNorm with 'TF' prefix to match other GraphBuilders
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 5 Jul 2019 08:49:35 +0000 (17:49 +0900)]
[moco/tf] Add Canonicalization and Knob for TFBiasAdd (#4117)
* [moco/tf] Add Canonicalization and Knob for TFBiasAdd
This will Add Canonicalization class with TFBiassAdd node.
Add a Knob for canonicalization of TFBiasAdd to Canonical BiasAdd node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* fix note
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 5 Jul 2019 08:47:43 +0000 (17:47 +0900)]
[loco] Introduce FeatureBiasAdd node (#4110)
This commit introduces FeatureBiasAdd which adds a "bias" along
a FeatureMap's depth axis.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 5 Jul 2019 08:25:26 +0000 (17:25 +0900)]
[moco/tf] Import as TFBiasAdd by Knob (#4116)
This will add ImportAsTFBiasAdd knob and import as TFBiasAdd when turned on.
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 5 Jul 2019 07:51:28 +0000 (16:51 +0900)]
[moco/tf] Move helper alias to TFNodeDecl header (#4115)
This will move helper alias from TFConv to TFNodeDecl header for common usage
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 5 Jul 2019 07:33:24 +0000 (16:33 +0900)]
[moco/tf] Introduce TFBiasAdd (#4114)
This will introduce TFBiasAdd to import BiasAdd node as-is
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 5 Jul 2019 05:24:38 +0000 (14:24 +0900)]
[exo/tflite] Extract Type/Shape Context (#4107)
SerializedModelData has multiple responsibilities as of now.
This commit extracts TypeContext (which manages node-to-type mapping)
and ShapeContext (which manages node-to-shape mapping) as the first step
to internal design revision.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 5 Jul 2019 04:08:10 +0000 (13:08 +0900)]
Add TF_SMALL_NET_0006 test (#4103)
* Add TF_SMALL_NET_0006 test
This test checkes what happens when "Concat" is in between two Feature operations.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Update comment
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 5 Jul 2019 04:07:59 +0000 (13:07 +0900)]
[moco/tf] Set Graph Input/Output Data Type (#4104)
From now on, moco::tf::Importer generates a loco graph with
valid Input/Output dtype values.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 5 Jul 2019 04:07:49 +0000 (13:07 +0900)]
[tf2tflite] Refine how to add tests (#4105)
This commit refines tf2tflite test framework to use CMake's "include"
command instead of ListFile_Read function.
This change permits flexible test management. For example, now it is
possible to comment out some tests temporarily.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 5 Jul 2019 01:15:41 +0000 (10:15 +0900)]
[loco] Introduce TensorBiasAdd alias (#4102)
This change allows users to use a bit short "loco::TensorBiasAdd" instead
of "loco::BiasAdd<loco::Domain::Tensor>".
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 5 Jul 2019 01:04:45 +0000 (10:04 +0900)]
[mocotest/onnx] Generate test file only on change (#4106)
The current implementation always copies and generates test files. With
this commit, test files are copied and generated only when it is
necessary.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 4 Jul 2019 22:05:27 +0000 (07:05 +0900)]
[TFtest] Add TF_SMALL_NET_0004 test (#4095)
* [TFtest] Add TF_SMALL_NET_0004 test
This commit adds TF_SMALL_NET_0004 test which sequentially performs
DepthwiseConv2d - FusedBatchNorm
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
* unpack values
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 4 Jul 2019 10:27:03 +0000 (19:27 +0900)]
[loco] Add DataType attribute to Graph Input/Output (#4097)
This commit allows loco frontends to record the data type of graph-level
input/output in GraphInput/GraphOutput.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 4 Jul 2019 10:11:46 +0000 (19:11 +0900)]
[locomotiv] Introduce ReLU6 operation to locomotiv (#4096)
This commit introduce ReLU6 operation to locomotiv.
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 4 Jul 2019 09:22:37 +0000 (18:22 +0900)]
[mocotest/tf] Copy files only on change (#4094)
moco testframework now copies (and generates) test files only on change.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>