3 [![Code Coverage](http://nnsuite.mooo.com/nntrainer/ci/badge/codecoverage.svg)](http://nnsuite.mooo.com/nntrainer/ci/gcov_html/index.html)
4 ![GitHub repo size](https://img.shields.io/github/repo-size/nnstreamer/nntrainer)
5 ![GitHub issues](https://img.shields.io/github/issues/nnstreamer/nntrainer)
6 ![GitHub pull requests](https://img.shields.io/github/issues-pr/nnstreamer/nntrainer)
8 NNtrainer is Software Framework for Training Neural Network Models on Devices.
12 NNtrainer is an Open Source Project. The aim of the NNtrainer is to develop Software Framework to train neural network model on embedded devices which has relatively limited resources. Rather than training the whole layers, NNtrainer trains only one or a few layers added after the feature extractor.
14 Even though it trains part of the neural network models, NNtrainer requires quite a lot of functionalities to train from common neural network frameworks. By implementing them, it is good enough to run several examples which can help to understand how it works. There are k-NN, Neural Network, Logistic Regression and Reinforcement Learning with CartPole in Applications directory and some of them use Mobilenet V2 with tensorflow-lite as a feature extractor. All of them tested on Galaxy S8 with Android and PC (Ubuntu 16.04).
17 * [Jijoong Moon](https://github.com/jijoongmoon)
18 * [MyungJoo Ham](https://github.com/myungjoo)
19 * [Geunsik Lim](https://github.com/leemgs)
22 * [Sangjung Woo](https://github.com/again4you)
23 * [Wook Song](https://github.com/wooksong)
24 * [Jaeyun Jung](https://github.com/jaeyun-jung)
25 * [Hyoungjoo Ahn](https://github.com/helloahn)
26 * [Parichay Kapoor](https://github.com/kparichay)
27 * [Dongju Chae](https://github.com/dongju-chae)
28 * [Gichan Jang](https://github.com/gichan-jang)
29 * [Yongjoo Ahn](https://github.com/anyj0527)
30 * [Jihoon Lee](https://github.com/zhoonit)
36 This component defines Layers which consist of Neural Network Model. Layers has own properties to be set.
38 | Keyword | Layer Name | Description |
39 |:-------:|:---:|:---|
40 | conv2d | Convolution 2D |Convolution 2-Dimentional Layer |
41 | pooling2d | Pooling 2D |Pooling 2-Dimentional Layer. Support average / max / global average / global max pooing |
42 | flatten | Flatten | Flatten Layer |
43 | fully_connected | Fully Connected | Fully Connected Layer |
44 | input | Input | Input Layer. This is not always requied. |
45 | batch_normalization | Batch Normalization Layer | Batch Normalization Layer. |
46 | loss layer | loss layer | hidden from users |
47 | activation | activaiton layer | set by layer property |
49 ### Supported Optimizers
53 | Keyward | Optimizer Name | Description |
54 |:-------:|:---:|:---:|
55 | sgd | Stochastic Gradient Decent | - |
56 | adam | Adaptive Moment Estimation | - |
62 | Keyward | Loss Name | Description |
63 |:-------:|:---:|:---:|
64 | mse | Mean squared Error | - |
65 | cross | Cross Entropy - sigmoid | if activation last layer is sigmoid |
66 | cross | Cross Entropy - softmax | if activation last layer is softmax |
68 ### Supported Activations
72 | Keyward | Loss Name | Description |
73 |:-------:|:---:|:---|
74 | tanh | tanh function | set as layer property |
75 | sigmoid | sigmoid function | set as layer property |
76 | relu | relu function | set as layer propery |
77 | softmax | softmax function | set as layer propery |
78 | weight_initializer | Weight Initialization | Xavier(Normal/Uniform), LeCun(Normal/Uniform), HE(Normal/Unifor) |
79 | weight_regularizer | weight decay ( L2Norm only ) | needs set weight_regularizer_param & type |
80 | learnig_rate_decay | learning rate decay | need to set step |
84 Tensor is responsible for the calculation of Layer. It executes the addition, division, multiplication, dot production, averaging of Data and so on. In order to accelerate the calculation speed, CBLAS (C-Basic Linear Algebra: CPU) and CUBLAS (CUDA: Basic Linear Algebra) for PC (Especially NVIDIA GPU) for some of the operation. Later, these calculations will be optimized.
85 Currently we supports lazy calculation mode to reduce copy of tensors during calcuation.
87 | Keyward | Description |
89 | 4D Tensor | B, C, H, W|
90 | Add/sub/mul/div | - |
91 | sum, average, argmax | - |
92 | Dot, Transpose | - |
93 | normalization, standardization | - |
100 | Keyward | Loss Name | Description |
101 |:-------:|:---:|:---|
102 | weight_initializer | Weight Initialization | Xavier(Normal/Uniform), LeCun(Normal/Uniform), HE(Normal/Unifor) |
103 | weight_regularizer | weight decay ( L2Norm only ) | needs set weight_regularizer_constant & type |
104 | learnig_rate_decay | learning rate decay | need to set step |
107 Currently we privde [C APIs](https://github.com/nnstreamer/nntrainer/blob/master/api/capi/include/nntrainer.h) for Tizen. C++ API also provides soon.
110 ### Examples for NNTrainer
112 #### [Custom Shortcut Application](https://github.com/nnstreamer/nntrainer/tree/master/Applications/Tizen_native/CustomShortcut)
115 This is demo application which enable user defined custom shortcut on galaxy watch.
117 #### [MNIST Example](https://github.com/nnstreamer/nntrainer/tree/master/Applications/MNIST)
119 This is example to train mnist dataset. It consists two convolution 2d layer, 2 pooling 2d layer, flatten layer and fully connected layer.
121 #### [Reinforcement Learning Example](https://github.com/nnstreamer/nntrainer/tree/master/Applications/ReinforcementLearning/DeepQ)
123 This is reinforcement learning example with cartpole game. It is using deepq alogrightm.
125 #### [Classification for cifar 10](https://github.com/nnstreamer/nntrainer/tree/master/Applications/TransferLearning/CIFAR_Classification)
127 This is Transfer learning example with cifar 10 data set. TFlite is used for feature extractor and modify last layer (fully connected layer) of network.
129 #### ~Tizen CAPI Example~
131 ~This is for demonstrate c api for tizen. It is same transfer learing but written with tizen c api.~
132 Deleted instead moved to a [test](https://github.com/nnstreamer/nntrainer/blob/master/test/tizen_capi/unittest_tizen_capi.cpp)
134 #### [KNN Example](https://github.com/nnstreamer/nntrainer/tree/master/Applications/KNN)
136 This is Transfer learning example with cifar 10 data set. TFlite is used for feature extractor and compared with KNN
138 #### [Logistic Regression Example](https://github.com/nnstreamer/nntrainer/tree/master/Applications/LogisticRegression)
140 This is simple logistic regression example using nntrainer.
146 The following dependencies are needed to compile / build / run.
148 * gcc/g++ (>= 4.9, std=c++14 is used)
150 * blas library (CBLAS) (for CPU Acceleration, libopenblas is used for now)
151 * cuda, cudart, cublas (should match the version) (GPU Acceleration on PC)
152 * tensorflow-lite (>= 1.4.0)
153 * libjsoncpp ( >= 0.6.0) (openAI Environment on PC)
154 * libcurl3 (>= 7.47) (openAI Environment on PC)
156 * libgtest (for testing)
159 ### Give It a Go Build with Docker
161 You can use [docker image](https://hub.docker.com/r/lunapocket/nntrainer-build-env) to easily set up and try building.
166 $ docker pull lunapocket/nntrainer-build-env:ubuntu-18.04
167 $ docker run --rm -it lunapocket/nntrainer-build-env:ubuntu-18.04
174 $ git pull # If you want to build with latest sources.
177 You can try build from now on without concerning about Prerequisites.
181 Download the source file by cloning the github repository.
184 $ git clone https://github.com/nnstreamer/nntrainer
187 After completing download the sources, you can find the several directories and files as below.
208 f1a3a05 (HEAD -> master, origin/master, origin/HEAD) Add more badges
209 37032a1 Add Unit Test Cases for Neural Network Initialization
210 181a003 lower case for layer type.
211 1eb399b Update clang-format
212 87f1de7 Add Unit Test for Neural Network
213 cd5c36e Add Coverage Test badge for nntrainer
217 You can find the source code of the core library in nntrainer/src. In order to build them, use [meson](https://mesonbuild.com/)
220 The Meson build system
222 Source dir: /home/wook/Work/NNS/nntrainer
223 Build dir: /home/wook/Work/NNS/nntrainer/build
224 Build type: native build
225 Project name: nntrainer
226 Project version: 0.0.1
227 Native C compiler: cc (gcc 7.5.0 "cc (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0")
228 Native C++ compiler: c++ (gcc 7.5.0 "c++ (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0")
229 Build machine cpu family: x86_64
230 Build machine cpu: x86_64
232 Build targets in project: 11
233 Found ninja-1.8.2 at /usr/bin/ninja
236 ninja: Entering directory `build'
237 [41/41] Linking target test/unittest/unittest_nntrainer_internal.
240 After completion of the build, the shared library, 'libnntrainer.so' and the static library, 'libnntrainer.a' will be placed in build/nntrainer.
242 $ ls build/nntrainer -1
243 d48ed23@@nntrainer@sha
244 d48ed23@@nntrainer@sta
249 In order to install them with related header files to your system, use the 'install' sub-command.
251 $ ninja -C build install
253 Then, you will find the libnntrainer.so and related .h files in /usr/local/lib and /usr/local/include directories.
255 By default, the command ```ninja -C build`` generates the five example application binaries (Classification, k-NN, LogisticRegression, ReinforcementLearning, and Training) you could try in build/Applications. For 'Training' as an example case,
257 $ ls build/Applications/Training/jni/ -1
258 e189c96@@nntrainer_training@exe
262 In order to run such example binaries, Tensorflow-lite is a prerequisite. If you are trying to run on the Android, it will automatically download tensorflow (1.9.0) and compile as static library. Otherwise, you need to install it by yourself.
267 Meson build `enable-test` set to true
273 $ meson build -Denable-test=true
274 The Meson build system
277 Configuring capi-nntrainer.pc using configuration
278 Run-time dependency GTest found: YES (building self)
279 Build targets in project: 17
281 Found ninja-1.10.0.git.kitware.jobserver-1 at /home/jlee/.local/bin/ninja
283 $ ninja -C build test
284 [79/79] Running all tests.
285 1/12 unittest_tizen_capi OK 8.86s
286 2/12 unittest_tizen_capi_layer OK 0.05s
287 3/12 unittest_tizen_capi_optimizer OK 0.01s
288 4/12 unittest_tizen_capi_dataset OK 0.03s
289 5/12 unittest_nntrainer_activations OK 0.03s
290 6/12 unittest_nntrainer_internal OK 0.23s
291 7/12 unittest_nntrainer_layers OK 0.22s
292 8/12 unittest_nntrainer_lazy_tensor OK 0.04s
293 9/12 unittest_nntrainer_tensor OK 0.04s
294 10/12 unittest_util_func OK 0.05s
295 11/12 unittest_databuffer_file OK 0.12s
296 12/12 unittest_nntrainer_modelfile OK 2.22s
306 if you want to run particular test
309 $ meson -C build test $(test name)
313 NNTrainer provides extensive sample app running test.
315 Meson build with `enable-app` set to true
320 $ meson build -Denable-app=true
321 The Meson build system
324 Configuring capi-nntrainer.pc using configuration
325 Run-time dependency GTest found: YES (building self)
326 Build targets in project: 17
328 Found ninja-1.10.0.git.kitware.jobserver-1 at /home/jlee/.local/bin/ninja
330 $ ninja -C build test
332 1/21 app_classification OK 3.59s
333 2/21 app_classification_func OK 42.77s
334 3/21 app_knn OK 4.81s
335 4/21 app_logistic OK 14.11s
336 5/21 app_DeepQ OK 30.30s
337 6/21 app_training OK 38.36s
338 7/21 app_classification_capi_ini OK 32.65s
339 8/21 app_classification_capi_file OK 32.04s
340 9/21 app_classification_capi_func OK 29.13s
344 if you want to run particular example only
347 $ meson -C build test $(test name) #app_classification_capi_func
352 1. [Training](https://github.com/nnstreamer/nntrainer/blob/master/Applications/Training/README.md)
354 After build, run with following arguments
355 Make sure to put last '/' for the resources directory.
357 $./path/to/example ./path/to/settings.ini ./path/to/resource/directory/
360 To run the 'Training', for example, do as follows.
365 $ LD_LIBRARY_PATH=./build/nntrainer ./build/Applications/Training/jni/nntrainer_training ./Applications/Training/res/Training.ini ./Applications/Training/res/
366 ../../res/happy/happy1.bmp
367 ../../res/happy/happy2.bmp
368 ../../res/happy/happy3.bmp
369 ../../res/happy/happy4.bmp
370 ../../res/happy/happy5.bmp
371 ../../res/sad/sad1.bmp
372 ../../res/sad/sad2.bmp
378 ## Open Source License
380 The nntrainer is an open source project released under the terms of the Apache License version 2.0.