[mir_caffe] Remove unused BatchNorm op (#6461)
It will be reimplemented taking into account different variations between frameworks.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[nnc] Remove unused BatchNorm op (#6460)
It will be reimplemented taking into account different variations between frameworks.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[nnc] Reformat using .clang-format (#6455)
Reformat sources to common style.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir] Delete ElementwiseOp (#6418)
It was replaced by several binary elementwise operations.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[nnc] Delete uses of ElementwiseOp (#6417)
It was replaced by several binary elementwise operations.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir] Remove BiasAdd and Scale ops (#6416)
These were replaced by more general binary elementwise operations.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
이성재/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 9 Aug 2019 10:23:06 +0000 (19:23 +0900)]
Update `freeze_graph.sh` script (#6345)
* Change to use the 'tensorflow' module installed on the system by default.
Signed-off-by: Sung-Jae Lee <sj925.lee@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 9 Aug 2019 09:59:06 +0000 (18:59 +0900)]
Remove loco-exporter (#6419)
loco-exporter was left for Circle, but it is likely that Circle exporter
follows the design of (revised) exo-tflite.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
[nnc] Remove remainings of BiasAdd and Scale ops (#6415)
These are leftovers after switching to equivalents elementwise operations.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
Remove mirunner component (#6414)
Remove `mirunner` PoC, there is no need for it anymore.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir_caffe2] Switch to binary elementwise operations (#6410)
Switch to new binary elementwise operations.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir2loco] Switch to binary elementwise operations (#6413)
Switch to new binary elementwise operations.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir_onnx] Switch to binary elementwise operations (#6412)
Switch to new binary elementwise operations.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir_tflite] Switch to binary elementwise operations (#6411)
Switch to new binary elementwise operations.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir_caffe] Switch to binary elementwise operations (#6409)
Switch to new binary elementwise operations.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 9 Aug 2019 08:07:49 +0000 (17:07 +0900)]
[res] DethwiseConv2D network with padding and strides (#6446)
A simple network that has DepthwiseConv2dNative with 'SAME' padding and
strides.
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 9 Aug 2019 07:09:02 +0000 (16:09 +0900)]
[mocotest-tf] enable test of fusion of binary op (#6443)
This will enable test of fusion of binary op with Conv2D, BiasAdd
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 9 Aug 2019 07:06:02 +0000 (16:06 +0900)]
[neurun] Do not wrap vector with unique_ptr (#6431)
As `std::vector` has move constructor, without wrapping it with
`unique_ptr` works efficiently as well.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 9 Aug 2019 06:51:24 +0000 (15:51 +0900)]
Add layout as a member of SubGraph into KernelGenerators (#6283)
This commit adds layout as a member of SubGraph into KernelGenerators.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 9 Aug 2019 06:06:44 +0000 (15:06 +0900)]
[moco-tf] Use precedingOp in FuseBinaryIntoPreceding (#6400)
This commit will introduce precedingOp instead of each operation variable
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 9 Aug 2019 06:06:29 +0000 (15:06 +0900)]
[loco] support DepthwiseConv2D of TypeForwardAlgorithm (#6404)
This commit is to support DepthwiseConv2D and DepthwiseFilterEncode in
CanonicalTypeForwardAlgorithm.
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 9 Aug 2019 06:06:12 +0000 (15:06 +0900)]
[moco-tf] Knob for ResolveConstantShape (#6439)
This commit introduces knob for ResolveConstantShape. Default behavior
is 'on'
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 9 Aug 2019 05:59:17 +0000 (14:59 +0900)]
[moco-tf] delete temporary loco_tobe::BiasShape (#6440)
This will delete temporary loco_tobe::BiasShape that was changed to loco::BiasShape
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 9 Aug 2019 05:30:29 +0000 (14:30 +0900)]
[res] DepthwiseConv2D network with one placeholder and one constant (#6427)
A simple network that has DepthwiseConv2dNative with input(Placeholder)
and filter(Const).
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 9 Aug 2019 05:28:49 +0000 (14:28 +0900)]
[moco-tf] Use loco::BiasShape (#6437)
This will change to use loco::BiasShape from temporary loco_tobe::BiasShape
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 9 Aug 2019 05:06:59 +0000 (14:06 +0900)]
[moco-tf] Knob for ResolveReshapeWildcardDim (#6436)
This commit introduces knob for ResolveReshapeWildcardDim. Default
behavior is 'on'
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 9 Aug 2019 05:04:57 +0000 (14:04 +0900)]
[neurun] Rename compiler::Element (#6430)
Move `compiler::Element` to `compiler::Linear::Element`. This removes
`Element.h` and make it to be an inner class of `Linear`.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 9 Aug 2019 03:38:58 +0000 (12:38 +0900)]
[moco-tf] Check attributes for ConcatV2 validate (#6428)
This will add attribute existence check for ConcatV2 node validation
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 9 Aug 2019 01:59:26 +0000 (10:59 +0900)]
[locop] Remove unused class in FormattedGraph.cpp (#6420)
CanonicalNodeSummaryBuilder was introduced for FormattedTensorShape class,
but CanonicalNodeSummaryBuilder is now not in this file, and thus no code
now uses this class.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 9 Aug 2019 01:59:11 +0000 (10:59 +0900)]
[moco-tf] Enable ResolveConstantShape (#6423)
This commit enables ResolveConstantShape transform
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 9 Aug 2019 01:50:58 +0000 (10:50 +0900)]
[neurun] Remove linear namespace (#6407)
As `namespace` linear has only few classes this commit moves those to
namespace `compiler`.
`compiler::Element` will be renamed to `compiler::LinearElement`.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 9 Aug 2019 01:18:12 +0000 (10:18 +0900)]
[locop] Introduce push_shape/pull_shape helper (#6398)
Let's replace tensor_shape with push_shape/pull_shape as the first step
to remove shape attribtues from Pull/Push node.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 9 Aug 2019 01:17:47 +0000 (10:17 +0900)]
[locop] Introduce FormattedTensorShape (#6399)
This commit introduces FormattedTensorShape class which dumps the
content of a given TensorShape into std::ostream.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 9 Aug 2019 01:16:05 +0000 (10:16 +0900)]
Revise cpu backend cmake (#6401)
- cpu backend, hi-perf-cpu backend
- Set strict build option: nnfw_common
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 9 Aug 2019 01:02:38 +0000 (10:02 +0900)]
[moco-tf] Turn off ImportAsTFConcatV2 (#6421)
This will turn off ImportAsTFConcatV2 knob before applying VariadicArityNode to TFConcatV2
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 9 Aug 2019 00:56:46 +0000 (09:56 +0900)]
[moco-tf] Import Shape operation (#6402)
This commit implements import stage for Shape operation
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
[nnc] Support for binary elementwise operations (#6408)
* Add implementations of binary elementwise operations to backends.
* Switch backend unittests to new ops.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir_onnx] Improve core structure of ONNX importer (#6297)
* Fix right initializer loading
* Fix output creation
* Other fixes
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
[mir] Move operations.lst.h into include/mir (#6405)
Move `include/mir/ops/operations.lst.h` -> `include/mir/Operations.inc` so that it can be found more easily.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 09:53:04 +0000 (18:53 +0900)]
Update skiplist for acl neon backend (#6396)
Update nnapi generated test for acl neon backend
- Enable validation test
- Enable trivial test except broadcast mul
- Enable depth-conv
- Enable comparison
- Enable remain fc tests
- Enable remain reshape tests
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 09:52:39 +0000 (18:52 +0900)]
Support async execution in NNAPI (#6395)
Implement Event class to support async execution
Use internal async execution function for startCompute and wait
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 09:51:29 +0000 (18:51 +0900)]
Revise acl cl backend cmake (#6391)
- Set strict build option: nnfw_common
- Remove needless link
- Update header include for acl common
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 08:55:46 +0000 (17:55 +0900)]
Rename gather implementation in arm-compute-ex (#6388)
Rename CLGather and CLGatherKernel to CLGatherEx and CLGatherExKernel because acl 19.05 already have same name layer & kernel
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 08:46:19 +0000 (17:46 +0900)]
Revise acl neon backend cmake (#6393)
- Set strict build option: nnfw_common
- Remove needless link
- Update header include for acl common
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 8 Aug 2019 07:58:52 +0000 (16:58 +0900)]
[moco-tf] relocate input method in Concat import (#6394)
This will relocate input method in Concat after class declaration to put anonymous codes together
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 07:43:00 +0000 (16:43 +0900)]
[moco-tf] Introduce ResolveConstantShape transform (#6379)
* [moco-tf] Introduce ResolveConstantShape transform
This commit introduces ResolveConstantShape transform, which is
responsible to replace determined TFShape node with TFConst.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
* Separate out condition check stage
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 07:42:30 +0000 (16:42 +0900)]
[loco] Update index on link (#6381)
Both "link" helpers now update the index of a corresponding node.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 07:42:02 +0000 (16:42 +0900)]
[moco-tf] TFShape shape inference (#6373)
This commit implements TFShape node's shape inference.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 07:41:42 +0000 (16:41 +0900)]
[exo-tflite] Use shared shape inference framework by default (#6390)
This commit makes exo-tflite to use the shape inference framework
implemented in loco core by default.
Please note that it is possible to turn off this feature via setting
EXOTFLITE_USE_LOCO_SHAPE_INFERENCE as 0.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 07:39:37 +0000 (16:39 +0900)]
[loco] Simplify ReLU Graph Testcase (#6392)
Let's simplify ReLU Graph Testcase using GraphBuilder.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 07:39:29 +0000 (16:39 +0900)]
[moco-tf] Revise sanity check for squeeze dim (#6370)
This commit would throw error for invalid squeeze dimension.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 07:23:31 +0000 (16:23 +0900)]
[res] Small network mocking the Epilogue (#6375)
This commit adds TF_SMALL_NET_0014, a mock Epilogue network using ReLU
instead of Softmax
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 07:04:05 +0000 (16:04 +0900)]
[neurun] Nit : fix OperandNode::BG_COLORS (#6384)
`BG_COLOR` is a static member of `Node` but it was refereced with a
subclass without a reason.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 8 Aug 2019 06:56:18 +0000 (15:56 +0900)]
[moco-tf] relocate MaxPool local class into anonymous (#6385)
This will relocate local class of MaxPool import into anonymous namespace
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 8 Aug 2019 06:55:53 +0000 (15:55 +0900)]
[moco-tf] relocate Mul local class into anonymous (#6386)
This will relocate local class of Mul import into anonymous namespace
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 8 Aug 2019 06:55:34 +0000 (15:55 +0900)]
[moco-tf] relocate Relu local class into anonymous (#6387)
This will relocate local class of Relu import into anonymous namespace
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 8 Aug 2019 06:52:39 +0000 (15:52 +0900)]
[moco-tf] relocate Relu6 local class into anonymous (#6389)
This will relocate local class of Relu6 import into anonymous namespace
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 05:42:38 +0000 (14:42 +0900)]
[loco] Simplify Identity Graph Testcase with Graph Builder (#6344)
* [loco] Simplify Identity Graph Testcase with Graph Builder
This commit simplifies Identity Graph Testcase with GraphBuilder.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Fix a typo
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 8 Aug 2019 05:40:14 +0000 (14:40 +0900)]
[moco-tf] relocate Identity local class into anonymous (#6382)
This will relocate local class of Identity import into anonymous namespace
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 8 Aug 2019 05:24:05 +0000 (14:24 +0900)]
[moco-tf] Switch on for ImportAsTFConcatV2 (#6380)
This will switch on for ImportAsTFConcatV2 in the knob
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 8 Aug 2019 04:30:28 +0000 (13:30 +0900)]
[moco-tf] Canonicalize ConcatV2 node (#6377)
This will enable Canonicalization of TFConcatV2 to canonical TensorConcat
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 04:19:50 +0000 (13:19 +0900)]
Fix acl backend test fail (#6376)
Fix acl cl backend test fail
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 04:19:19 +0000 (13:19 +0900)]
[loco] Add index getter (#6367)
This commit extends GraphInput and GraphOutput class with "index"
getter.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 04:12:52 +0000 (13:12 +0900)]
[mocotest-tf] Enable Squeeze_002 (#6369)
This commit would enable squeeze network test with negative squeeze
dimension.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 8 Aug 2019 04:03:59 +0000 (13:03 +0900)]
[moco-tf] relocate FusedBatchNorm local class into anonymous (#6374)
This will relocate local class of FusedBatchNorm import into anonymous namespace
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 04:00:26 +0000 (13:00 +0900)]
[moco-tf] TFShape summary (#6372)
This commit adds TFShape node summary.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 8 Aug 2019 03:54:17 +0000 (12:54 +0900)]
[moco-tf] Introduce ConcatV2Canonicalizer (#6360)
* [moco-tf] Introduce ConcatV2Canonicalizer
This will introduce ConcatV2Canonicalizer that converts TFConcatV2 to canonical TensorConcat node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* fix typo
* remove unused header
Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Thu, 8 Aug 2019 03:42:02 +0000 (06:42 +0300)]
Fix several warnings (#6242)
* Fix several warnings
Make class/struct definitions consistent( forward declararation with different type from definition )
Remove pessimising moves of temporary objects
Change `info` symbol in library_info.cc into non-static
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 03:40:39 +0000 (12:40 +0900)]
Replace strict build flag setting in libs/tflite (#6232)
Replace libs/tflite strict build setting cmake from using compile option to linking nnfw_common interface
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 02:45:27 +0000 (11:45 +0900)]
[res] TF Shape-Reshape network (#6343)
This commit adds tensorflow Shape test using Reshape node.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 02:44:47 +0000 (11:44 +0900)]
Replace strict build setting in libs (#6226)
Replace strict build setting in libs to use nnfw_common interface (except libs/tflite)
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 02:41:16 +0000 (11:41 +0900)]
[loco] Infer the shape of bias-taking nodes (#6332)
* [loco] Infer the shape of bias-taking nodes
CanonicalShapeInferenceRule is now able to infer the shape of
BiasEncode, FeatureBiasAdd, and TensorBiasAdd nodes.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Fix a typo
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 02:27:18 +0000 (11:27 +0900)]
[neurun] Enhance DotDumper backend color assign (#6368)
Enhance DotDumper backend color assignment. Each Backend will have its
unique color up to 8 of them. Before it was assigned based on hard-coded
Backend ID.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 01:52:17 +0000 (10:52 +0900)]
[moco-tf] Introduce TFShape IR (#6322)
This commit introduces TF-dialect TFShape IR for tensorflow 'Shape'
operation
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 01:47:06 +0000 (10:47 +0900)]
[loco] Shape inference over Conv2D-related nodes (#6349)
CanonicalShapeInferenceRule is now able to infer the shape of nodes
related with Conv2D (such as FilterEncode).
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 01:46:41 +0000 (10:46 +0900)]
[moco-tf] Resolve negative squeeze dimension (#6352)
This commit resolves negative squeeze dimension during shape inference
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 01:46:33 +0000 (10:46 +0900)]
[neurun] Favor `map::emplace` over `map::insert` (#6355)
Favor `map::emplace` over `map::insert` for every occurences in neurun.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 8 Aug 2019 01:46:11 +0000 (10:46 +0900)]
[moco-tf] relocate AvgPool local class into anonymous (#6366)
This will relocate local class of AvgPool import into anonymous namespace
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 8 Aug 2019 01:45:50 +0000 (10:45 +0900)]
[moco-tf] relocate Add local class into anonymous (#6365)
This will relocate local class of Add import into anonymous namespace
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 8 Aug 2019 01:45:17 +0000 (10:45 +0900)]
[moco-tf] Disable usage of nodes.lst in FixPadding (#6361)
This will disable using CanonicalNodes.lst and TFNodes.lst so that we don't have to modify FixPaddingTransform
- this is a temporary fix before deleting FixPaddingTransform related files
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 01:44:01 +0000 (10:44 +0900)]
[neurun] Refine DotDumper (#6347)
- Rename method `dumpIfNeeded` to `dump`
- Rename enum `OPTION` to `DotDumper::Level` and update comments
- Take out config fetching from class `DotDumper`
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 8 Aug 2019 01:42:30 +0000 (10:42 +0900)]
[moco-tf] Import as TFConcatV2 with a knob (#6348)
This will change import of ConcatV2 node as Canonical or TF-dialect with a knob.
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 01:20:17 +0000 (10:20 +0900)]
[loco] Hide deprecated "node" method (#6342)
Now, "node" method in GraphInput/GraphOutput is private.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 01:19:57 +0000 (10:19 +0900)]
[res] Simple TF Shape test case (#6341)
This commit adds simple Shape network for TF test.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 00:49:33 +0000 (09:49 +0900)]
Fix converting type of zero point in select_operator.py (#6353)
This commit fixes converting type of zero point in select_operator.py.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 00:45:46 +0000 (09:45 +0900)]
Revise acl common cmake (#6340)
- Add strict build option: nnfw_common
- Link arm_compute and arm_compute_ex library
- Include path acl_common for acl_common user (acl_cl, acl_neon)
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Aug 2019 00:45:27 +0000 (09:45 +0900)]
Merge neurun kernel and backend build (#6313)
- Merge neurun each backend's kernel and backend build
- Fix build error on cpu backend kernel by restrict build option
- Invalid type comparison
- Unused variable
- Unused function
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 8 Aug 2019 00:22:43 +0000 (09:22 +0900)]
[locop] FixedReshape node summary (#6354)
This commit adds FixedReshape node summary.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 7 Aug 2019 22:01:57 +0000 (07:01 +0900)]
[moco-tf] Use common calc_paddata in FixShape (#6346)
This will change calc_node_pad() and calc_annot_paddata() to use common calc_annot_paddata() in FixShapeTransform
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
[nnc] Replace BiasAdd and Scale with elementwise equivalents (#6358)
`BiasAdd` and `Scale` are restricted versions of equivalent Elementwise ops and are going to be removed.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 7 Aug 2019 12:23:12 +0000 (21:23 +0900)]
Remove TensorBuilder getter from IKernelGenerator (#6331)
Now that `Backend` holds `TensorBuilder` directly, no need to have this
interface from `IkernelGenerator` any longer.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
[mir] Add binary elementwise operations (#6351)
first step of future replacement of single `ElementwiseOp` class with set of individual operations.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 7 Aug 2019 09:56:12 +0000 (18:56 +0900)]
[moco-tf] Introduce TFConcatV2 IR (#6321)
This will introduce TFConcatV2 IR and required codes for TensorFlow Concat node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 7 Aug 2019 09:44:42 +0000 (18:44 +0900)]
Support asynchronous execution (#6314)
- Introduce methods for start execution & wait finish in Execution class to support asynchronous execution
- Add test using introduced methods
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 7 Aug 2019 09:21:33 +0000 (18:21 +0900)]
[tf2tflite] copying customop.conf into cmake build/compiler/tf2tflite (#6329)
* [tf2tflite] copying customop.conf into cmake build/compiler/tf2tflite
This makes cmake to copy customop.conf into cmake build/compiler/tf2tflite.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* indent, 'xxx_customop.conf' -> 'xxx.customop.conf'
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 7 Aug 2019 09:21:13 +0000 (18:21 +0900)]
[exo-tflite] Implement FilterShape converter (#6334)
This commit implement a helper which converts FilterShape as exo-tflite
internal ShapeDescription.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 7 Aug 2019 09:14:16 +0000 (18:14 +0900)]
[moco-tf] Remove redundant codes in FuseBinaryIntoPreceding (#6319)
* [moco-tf] Remove redundant codes in FuseBinaryIntoPreceding
There are some redundant codes in `FuseBinaryIntoPreceding.cpp`
This commit will remove them.
- Exactly only one of variables(`xc` and `yc`) is `nullptr` because `is_only_one_valid` already checks.
- Then `addparam` and `mulparam` cannot be `nullptr`.
- After removing above condition, there can be only one case that `bchanged` can be returned, that is when `biasadd` is `nullptr`. However when `biasadd` is `nullptr`, bchanged is always `false`.
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
* apply comments
* add missing codes
* fix comments
* apply comments
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 7 Aug 2019 09:10:30 +0000 (18:10 +0900)]
[moco-tf] Common method for calculation of PadData (#6338)
This will introduce common method for calculation of PadData that exist in calc_node_pad() and calc_annot_paddata() of FixShapeTransform
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 7 Aug 2019 08:22:09 +0000 (17:22 +0900)]
[loco] Introduce GraphBuilder helper class (#6308)
* [loco] Introduce GraphBuilder helper class
This commit introduces GraphBuilder helper class which faciliates
testcase implementation.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Update comment
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 7 Aug 2019 08:18:58 +0000 (17:18 +0900)]
[loco] TensorConcat Shape Inference (#6315)
* [loco] TensorConcat Shape Inference
CanonicalShapeInferenceRule is now able to infer the shape of TensorConcat
nodes.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Remove invalid comments