Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Mon, 29 Jul 2019 12:43:24 +0000 (15:43 +0300)]
[neurun] Add custom kernel registry (#5597)
Add custom kernel registry used to keep track of custom kernels used by specific backend( see #5411 )
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
[nnc] Join `import` and `createIR` methods into `importModel` (#5984)
Methods `import` and `createIR` are always used together. Join them to simplify usage.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 29 Jul 2019 11:30:43 +0000 (20:30 +0900)]
Fix unmatched result of tconv in NCHW (#5968)
This commit fixes unmatched result of tconv in NCHW
- Change weights of tconv to OIHW in case of NCHW
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 29 Jul 2019 11:16:12 +0000 (20:16 +0900)]
[moco-tf] Summary for TFReshape node (#5979)
This commit adds summary for TFReshape node
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 29 Jul 2019 11:15:43 +0000 (20:15 +0900)]
[moco-tf] Enable AvgPoolCanonicalizer (#5980)
This will add AvgPoolCanonicalizer to Canonicalizer list
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 29 Jul 2019 11:13:41 +0000 (20:13 +0900)]
[loco] Introduce const Graph getters (#5974)
Now, Graph provides const "nodes", "inputs", "outputs" methods.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 29 Jul 2019 11:13:31 +0000 (20:13 +0900)]
[loco] Remove unused input_nodes helper (#5972)
* [loco] Remove unused input_nodes helper
This commit removes unused input_nodes helper. This helper is currently
unused, but its presence makes it difficult to decouple loco core from
loco canonical.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Remove the implementation
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 29 Jul 2019 11:13:19 +0000 (20:13 +0900)]
[loco] Input/Output Index from Push/Pull node (#5961)
* [loco] Input/Output Index from Push/Pull node
This commit revises Push/Pull node to store associated Graph-level
Input/Output index in it.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Fix typos
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 29 Jul 2019 06:21:05 +0000 (15:21 +0900)]
[moco-tf] Introduce TFMaxPool IR (#5975)
This will introduce TFMaxPool IR in TensorFlow Dialect for MaxPool node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 29 Jul 2019 06:12:12 +0000 (15:12 +0900)]
[neurun] Always notify ExecutionObserver (#5958)
Now, DataflowExecutor notifies the beginning and end of the whole
execution whenever any observer exists (even when profiling mode is
disabled).
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 29 Jul 2019 05:38:30 +0000 (14:38 +0900)]
Fix format error in moco-tf (#5970)
Fix format error in moco-tf's FixShapeTransform.cpp
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 29 Jul 2019 05:12:18 +0000 (05:12 +0000)]
[moco/tf] Introduce TFDepthwiseConv2D IR. (#5912)
* [moco/tf] Introduce TFDepthwiseConv2D IR.
This commit introduce TFDepthwiseConv2D IR.
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
* apply 'Native' keyword.
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 29 Jul 2019 04:25:16 +0000 (13:25 +0900)]
Remove sinks and sources member in ExecutorBase (#5885)
Remove sinks and sources member in ExecutorBase and source/sink methods return unique pointer
Move unused input checking into execute method
Remove unused parameter for source/sink
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
김용섭/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 29 Jul 2019 01:49:15 +0000 (10:49 +0900)]
[neurun] Remove unnecessary class and verbose code (#5930)
Remove unnecessary class(LinearAclMemoryManager) and verbose code
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 29 Jul 2019 01:42:29 +0000 (10:42 +0900)]
Unified dockerfile (#5925)
Merge infra/nncc/Dockerfile, infra/nnfw/docker/Dockerfile, and infra/nnfw/docker/Dockerfile_tizen into infra/docker/Dockerfile
- Base ubuntu version: 16.04
- Use apt-get
- Native/cross gcc/g++ version: 6.x
- Update nncc/nnfw docker build documents
- Move infra/nnfw/command/build-docker command into infra/command and change name to build-docker-image
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 29 Jul 2019 01:28:51 +0000 (10:28 +0900)]
Fix CRLF new line (#5959)
Fix CRLF new line checking error by 19_NN_Compiler_and_Return_SDD.rst
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 29 Jul 2019 01:21:01 +0000 (10:21 +0900)]
[neurun] Rename functions in TemplTensorBuilder (#5943)
Function `start_lifetime` and `end_lifetime` could be misleading sine
they may or may not do. So this commit renames them to `def_handler` and
`use_handler`, respectively.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
Nikita Sizov/AI Tools Lab /SRR/Professional/삼성전자 [Mon, 29 Jul 2019 01:07:40 +0000 (04:07 +0300)]
Mock ACL NEON nodes for further implementations (#5942)
* Add mock functions for ACL NEON nodes
Mock: Squeeze, StridedSlice, Transpose, Div, Exp, ReduceMax and Comparison
Signed-off-by: Sizov Nikita <n.sizov@samsung.com>
* Use unused variables in mocked functions
Signed-off-by: Sizov Nikita <n.sizov@samsung.com>
Remove .FORMATDENY files from properly formatted directories (#5945)
Remove.FORMATDENY files from mir_{caffe2,caffe,onnx,tflite}_importer directories which have been recently formatted.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 29 Jul 2019 01:00:31 +0000 (10:00 +0900)]
Support activation operations for acl neon (#5932)
This commit supports activation operations for acl neon.
- Abs, ReLU*, SQRT
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 29 Jul 2019 00:43:14 +0000 (09:43 +0900)]
[res/TensorFlowTests] Modify Custom Op user data (#5941)
This commit adds `dtype` of custom op, which will be accessed for TypeInference.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
DR1 RNN requirements (#5772)
* DR1 RNN requirements
Add requirements related to RNN networks
remove redundant requirements
Signed-off-by: Efimov Alexander <a.efimov@samsung.com>
Dr1 development document (#5646)
* Updated SDD document for DR1
- Added more tasks
- Changed formatiing of task table
- Responsible person is changed
* change development plan, fix responcibilities, remove NNCC repo mentions
* add control flow related item
* remove redundant tizen requirements
* relax control flow requirements
* fix naming of components
Signed-off-by: Efimov Alexander <a.efimov@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 29 Jul 2019 00:09:56 +0000 (09:09 +0900)]
[moco-tf] remove ModelMetainfo.proto (#5940)
This file, a schema for user input, is now moved into tf2tflite.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Sun, 28 Jul 2019 23:52:34 +0000 (08:52 +0900)]
[moco-tf] Implement shape inference for TFReshape (#5914)
This commit implements shape inference for TFReshape on fix_shape().
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Sun, 28 Jul 2019 23:37:41 +0000 (08:37 +0900)]
[loco] Support traversal to Graph from Node (#5910)
This commit extends Node class to provide "graph" method which returns
a pointer to associated Graph object.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
Extract importers out of nnc (#5900)
Move sources of Caffe, Caffe2, TensorFlow Lite, ONNX importers from `nnc` to top-level `compiler` directory.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
Павел Ильютченко/AI Tools Lab /SRR/Engineer/삼성전자 [Fri, 26 Jul 2019 11:25:42 +0000 (14:25 +0300)]
[mir2loco] Support Pooling operations (#5846)
* Support Pooling operation and tests for 2 ops
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
[nnc] Cleanup after separating importers (#5902)
Importers have been extracted into independent components. This is a follow-up cleanup.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
Reformat mir importers (#5904)
Reformat code according to coding style.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 26 Jul 2019 09:57:14 +0000 (18:57 +0900)]
[neurun] Fix load_model_from_file (#5928)
load_load_from_file should load model from {nnpackage path + model path}.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
김용섭/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 26 Jul 2019 07:32:40 +0000 (16:32 +0900)]
[neurun] Remove unnecessary methods of ITensorBuilder (#5923)
Remove unnecessary methods of ITensorBuilder, cpu/TensorBuilder and
TemplTensorBuilder
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
[neurun] Introduce OperationIndexMap as Operand has (#5918)
There are a lot of use `std::unordered_map<model::OperationIndex, T>`
So, introducing map alias like Operand has is logical
Signed-off-by: Dilshodzhon Poshshoev <d.poshshoev@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 26 Jul 2019 05:36:37 +0000 (14:36 +0900)]
[neurun] Generalize ExecutionObserver interface design (#5867)
The current design declares "uploadExecTime" which is a bit specific to
dedicated observer implementation.
This commit revises ExecutionObserver to hook on the whole model execution,
and ProfileObserver to invoke "uploadExecTime" when model execution
ends.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 26 Jul 2019 05:30:04 +0000 (14:30 +0900)]
[moco-tf] Apply plier-tf to AvgPoolCanonicalizer (#5921)
This commit applies plier-tf to AvgPoolCanonicalizer.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 26 Jul 2019 05:00:11 +0000 (14:00 +0900)]
[moco-tf] Applying plier-tf (#5891)
This commit applies `plier-tf` into `moco-tf`.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 26 Jul 2019 04:19:56 +0000 (13:19 +0900)]
Remove exceptional case in python style check (#5917)
Remove skip compiler directory check on python style check
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
김용섭/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 26 Jul 2019 04:18:21 +0000 (13:18 +0900)]
[neurun] Enable handling subtensors for notify{First|Last}Uses (#5915)
Enable handling subtensors for notify{First|Last}Uses by preVisit and
postVisit.
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 26 Jul 2019 03:46:34 +0000 (12:46 +0900)]
Format check for diff file with master (#5916)
This commit allow user to check format diff files with master via CHECK_DIFF_ONLY environment variable
Use this feature on pre-push githook
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 26 Jul 2019 03:00:17 +0000 (12:00 +0900)]
[neurun] Selective HI_PERF_CPU backend build (#5913)
This commit allows users to diable HI_PERF_CPU backend build via
BUILD_NEURUN_HI_PERF_CPU_BACKEND option.
Please note that this commit does not change the default behavior.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
김용섭/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 26 Jul 2019 02:45:07 +0000 (11:45 +0900)]
[neurun] Apply acl internal buffer manager into KernelGenerator (#5908)
Apply acl internal buffer manager into KernelGenerator on cl/neon
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
김용섭/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 26 Jul 2019 00:08:03 +0000 (09:08 +0900)]
[neurun] Append acl_memory_manager method into TemplTensorBuilder (#5887)
Append acl_memory_manager method into TemplTensorBuilder to be used for
MemoryManager's internal buffer manager
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
김용섭/On-Device Lab(SR)/Engineer/삼성전자 [Fri, 26 Jul 2019 00:07:53 +0000 (09:07 +0900)]
[neurun] Revise planTensors of Linear (#5882)
Revise planTensors of Linear on neurun
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 25 Jul 2019 23:38:44 +0000 (08:38 +0900)]
[mocotest-tf] Add TF_SMALL_NET_0005 test (#5888)
This test is consisted of `Reshape` and `Shape` operation.
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 25 Jul 2019 23:29:57 +0000 (08:29 +0900)]
[loco] Fix canonical node list (#5893)
This will fix canonical node list to match alphabetical order of foward
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 25 Jul 2019 08:59:11 +0000 (17:59 +0900)]
[tf2tflite] Enable Relu6_000 test (#5886)
tf2tflite test framework now checks Relu6_000 automatically.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 25 Jul 2019 08:39:50 +0000 (17:39 +0900)]
[moco-tf] Introduce AvgPool Canonicalizer (#5819)
This will introduce AvgPool node Canonicalizier
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
김용섭/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 25 Jul 2019 08:33:03 +0000 (17:33 +0900)]
[neurun] Append preVisit and postVisit to ITensorBuilder (#5879)
Append preVisit and postVisit to ITensorBuilder for #5642
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 25 Jul 2019 07:23:11 +0000 (16:23 +0900)]
[exo-tflite] Support ReLU6 (#5884)
TFLExporter is now able to export ReLU6 node as T/F Lite RELU6
operation.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 25 Jul 2019 06:41:24 +0000 (15:41 +0900)]
[moco-tf] Revise Indentity GraphBuilder (#5855)
This will revise Identity node GraphBuilder to either convert to TF or Canonical dialect by a knob
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 25 Jul 2019 06:21:58 +0000 (15:21 +0900)]
[plier/tf] Move more func from current version of moco-tf Convert.h (#5869)
* [plier/tf] Move more func from current version of moco-tf Convert.h
This commit moves more func from current version of moco-tf Convert.h.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* add require(loco). format checked
* error when trying to convert DT_BOOL to loco::DataType
* fix wrong modification
* fix wrong modification
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 25 Jul 2019 06:21:25 +0000 (15:21 +0900)]
[loco] Support ReLU6 type inference (#5877)
Now, type inference framework supports ReLU6 node.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 25 Jul 2019 06:04:20 +0000 (15:04 +0900)]
[plier/tf] remove macro that does not fit with plier-tf (#5871)
Remove a macro, which has no relation with TF and plier-tf.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 25 Jul 2019 05:54:39 +0000 (14:54 +0900)]
[res] Add test for endmost part of Inception V3 (#5834)
* [res] Add test for endmost part of Inception V3
This commit adds test case for 'the Epilogue', or the endmost part of
Inception V3 comprised of Squeeze, Reshape, Shape and Softmax.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
* Style consistency
김용섭/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 25 Jul 2019 05:30:28 +0000 (14:30 +0900)]
[neurun] Make visitor for Subgraph override in KernelGenerator (#5876)
Make visitor for Subgraph override in KernelGenerator for preVisit and
postVisit in the future
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 25 Jul 2019 05:30:20 +0000 (14:30 +0900)]
[moco-tf] Rename file and GraphBuilder name for MaxPool (#5851)
This will rename file and GraphBuilder name for MaxPool to follow TensorFlow GraphDef
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
Code refactoring Scheduler except makeRank method (#5822)
Split Scheduler::scheduleNode into smaller ones, created
const Graph *_graph, did const some temp variables
Signed-off-by: Dilshodzhon Poshshoev <d.poshshoev@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 25 Jul 2019 05:28:17 +0000 (14:28 +0900)]
[exo-tflite] Remove unused type inference code (#5859)
exo-tflite now reuses type inference framework that loco provides.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 25 Jul 2019 05:27:11 +0000 (14:27 +0900)]
[moco-tf] Introduce TFSqueeze dialect node (#5843)
* [moco-tf] Introduce TFSqueeze dialect node
This commit introduces TFSqueeze dialect node.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
* Use 'squeeze_dims' instead of 'axis'
Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Thu, 25 Jul 2019 05:25:16 +0000 (08:25 +0300)]
[neurun] Add custom operation node (#5598)
Used as a base node for all custom operations. For general design see at #5411
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 25 Jul 2019 05:01:30 +0000 (14:01 +0900)]
[neurun] Remove unused prepare method (#5850)
IFunction interface declares prepare method, but no code invokes this
method.
Let's remove this used prepare method.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 25 Jul 2019 04:48:13 +0000 (13:48 +0900)]
Remain unified format check script only (#5861)
Remove nncc/nnfw format check script
Remove all .FORMATCHECKED
Add .FORMATDENY into some directories to ignore format check
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 25 Jul 2019 04:26:56 +0000 (13:26 +0900)]
[moco-tf] Run for active nodes for FixPad and FixShape (#5862)
This will fix to run for active nodes from outputs for FixPadding and FixShape transformations
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
Decouple ONNX frontend (#5755)
Remove dependencies of `ONNXImporterImpl` on other `nnc` components.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
Decouple caffe2 frontend (#5754)
Remove dependencies of `Caffe2Importer` on other `nnc` components.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
Decouple caffe frontend (#5753)
Remove dependencies of `CaffeImporter` on other `nnc` components.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 25 Jul 2019 04:10:39 +0000 (13:10 +0900)]
Replace githook and install (#5860)
Replace githook to infra/git-hooks
Replace githook install command to infra/command
Update path in script
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
Decouple TFLite frontend (#5752)
Remove dependencies of `TfliteImporter` on other `nnc` components.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
Reformat mir component (#5847)
Reformat code according to coding style.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 25 Jul 2019 03:51:56 +0000 (12:51 +0900)]
[moco-tf] Introduce TFReshape dialect node (#5863)
This commit introduces TFReshape dialect node.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
Павел Ильютченко/AI Tools Lab /SRR/Engineer/삼성전자 [Thu, 25 Jul 2019 03:50:09 +0000 (06:50 +0300)]
[mir2loco] Introduce MIR to Loco transformer (#5751)
* Implemented core structure
* Supported 2 ops: Input and Output
* Unit test for input and output operations
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 25 Jul 2019 02:46:46 +0000 (11:46 +0900)]
[moco-tf] Introduce IdentityCanonicalizer (#5854)
* [moco-tf] Introduce IdentityCanonicalizer
This will introduce Identity node Canonicalizer
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* fix comment
* another fix comment
Unify PermuteLayer::runTempl and use Object::access (#5778)
For all types of PermuteNode::Type, runTempl does the same calculation.
The only difference is for rank==4. Unify it and move switch for type
into rank==4 case
Signed-off-by: Dilshodzhon Poshshoev <d.poshshoev@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 25 Jul 2019 02:04:15 +0000 (11:04 +0900)]
Fix format script bug (#5858)
Fix format script bug checking environment variable
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 25 Jul 2019 01:30:21 +0000 (10:30 +0900)]
Change tab size in format script (#5852)
Fix tab size to 2 in format script
Same tab size with other scripts in infra/nncc/command and infra/nnfw/command
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 25 Jul 2019 01:17:36 +0000 (10:17 +0900)]
[moco-tf] Fix test of fix shape and pad transform (#5848)
This will set outputs of test graph for FixShape and FixPadding transform
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 25 Jul 2019 00:59:31 +0000 (09:59 +0900)]
[moco-tf] Constant Folding Transform for Canonical Dialect (#5759)
* [moco-tf] Constant Folding Transform for Canonical Dialect
This commit adds constant folding transform for canonical dialect, plus a test and build files.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* revised
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 25 Jul 2019 00:30:17 +0000 (09:30 +0900)]
Update ExecutorFactory comments (#5844)
Update outdated comments in `ExecutorFactory.cc`.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 25 Jul 2019 00:21:09 +0000 (09:21 +0900)]
[moco-tf] Remove knob condition for Canonicalizer (#5833)
This will remove condition with knob for checking for specific TF dialect node
With this change there should be no TF dialects after Canonicalization
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 25 Jul 2019 00:14:10 +0000 (09:14 +0900)]
[moco-tf] Introduce TFIdentity IR (#5839)
This will introduce TFIdentity IR for TensorFlow Identity node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 25 Jul 2019 00:13:39 +0000 (09:13 +0900)]
[moco-tf] ShapeInferenceData equality operator (#5845)
This commit introduces equality operator for ShapeInferenceData.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 24 Jul 2019 23:46:22 +0000 (08:46 +0900)]
[moco-tf] Remove unused header (#5841)
"CompilerArgs.h", which is not used, was removed.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 24 Jul 2019 23:45:31 +0000 (08:45 +0900)]
[moco-tf] Fix transformation changed info (#5836)
This will fix to show transformation has change the graph where it was like
once changed is true then all other transformations in a phase will show as changed
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 24 Jul 2019 22:43:02 +0000 (07:43 +0900)]
[locoex/custom op] enhancing locoex custom op (#5810)
Now customop node can store dtype info. Two methods are defined as const.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
[neurun] Unify pull and push of Sink and Source (#5808)
* Unify pull and push of Sink and Source
All calculations are the same except the case rank==4
Signed-off-by: Dilshodzhon Poshshoev <d.poshshoev@samsung.com>
* Remove pushUnif and pullUnif from highest level base
Signed-off-by: Dilshodzhon Poshshoev <d.poshshoev@samsung.com>
* Remove "override"
Signed-off-by: Dilshodzhon Poshshoev <d.poshshoev@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 24 Jul 2019 11:18:47 +0000 (20:18 +0900)]
Remove setInput and setOutput in IExecutor (#5832)
Remove setInput/setOutput in IExecutor and ExecManager
Remove deprecated setInput/setOutput in ExecutorBase
Internally used setInput/setOutput in ExecutorBase: rename to source/sink and change to private
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 24 Jul 2019 11:15:04 +0000 (20:15 +0900)]
Remove PlanBuilder (#5838)
Now `PlanBuilder` is redundant class which just makes harder to read
code. So its behavior is moved to `createLinearExecutor`, just like
`createDataflowExecutor` does.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 24 Jul 2019 10:59:45 +0000 (19:59 +0900)]
Update format check script (#5829)
From nncc format checker
- Update clang-format tool checking
- Return false if git diff remain
- Support running format check from outside nnfw repo directory
Script update
- Merge tool check and file check function
- Move environment variable check for c++/python check
- Remove git submodule check
- Update file search condition to find c++ and python
- Use bracket for script variable
- Add function keyword
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 24 Jul 2019 10:56:08 +0000 (19:56 +0900)]
[moco-tf] setup_output_node for unittest (#5828)
* [moco-tf] setup_output_node for unittest
This will introduce setup_output_node() in test helper to add Push node as graph outputs
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* edit comment
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 24 Jul 2019 10:09:10 +0000 (19:09 +0900)]
Enable gtest for two live execution instance (#5831)
Enable gtest to test multiple live execution instance using one executor
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 24 Jul 2019 09:38:52 +0000 (18:38 +0900)]
[moco-tf] default name for TFNodeSummaryBuilder (#5817)
This will change to show default name for TFNodeSummaryBuilder and not to break while testing cause of unimplementation
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 24 Jul 2019 09:26:07 +0000 (18:26 +0900)]
Set thread count for tflite_run (#5830)
Set thread count for tflite_run by using THREAD environment variable
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 24 Jul 2019 08:33:44 +0000 (17:33 +0900)]
Support resize input in tflite_run (#5821)
Support resize input in tflite_run by using --ishapes argument
Example:
./Product/out/bin/tflite_run --ishapes 1 90 160 3 --tflite 131_x2_deconv.tflite
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 24 Jul 2019 07:29:48 +0000 (16:29 +0900)]
[locomotiv] Reshape node accepts tensor only (#5798)
Reshape loco node is defined to accept tensor only. This commit corrects
this fact on locomotiv.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 24 Jul 2019 07:09:47 +0000 (16:09 +0900)]
Rename model to graph which is a Graph (#5820)
`_model` was misnamed and could be confusing since we use "Model" and
"Graph" differently in our source so this commit renames it to `_graph`.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 24 Jul 2019 07:09:00 +0000 (16:09 +0900)]
[exo-tflite] Use loco type inference framework (#5807)
This commit revises the implementation of exo-tflite type inferene framework
by using loco type inference framework.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 24 Jul 2019 07:08:47 +0000 (16:08 +0900)]
[locop] Define << operator and dependent class in the same namespace (#5800)
The current implementation defines operator<< on loco::TensorShape
in anonymous namespace, which results in undefined behavior when it
is invoked from external libraries (such as pp).
This commit introduces FormattedTensorShape as a wrapper to mitigate
this undefined behavior.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
[neurun] Elimination data transfer assumption in Scheduler (#5818)
Eliminate multiplying data tranfer time by 2 and calculate
free time slot for permutation
Signed-off-by: Dilshodzhon Poshshoev <d.poshshoev@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 24 Jul 2019 05:34:52 +0000 (14:34 +0900)]
Apply coding style to compiler modules (#5812)
Apply coding style to
- cli
- moco-onnx
- nnop
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 24 Jul 2019 05:12:28 +0000 (14:12 +0900)]
Remove IExecutor's setInput and setOutput call (#5785)
* Remove IExecutor's setInput and setOutput call
Pass gathered input/output description when call IExecutor's executor() instead of call setInput and setOutput in IExecutor interface
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
* Handle optional input/output in ExeucotrBase