NNtrainer

NNtrainer is Software Framework for Training Neural Network Models on Devices.

Overview

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.

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).

Maintainer

Reviewers

Components

Supported Layers

This component defines Layers which consist of Neural Network Model. Layers has own properties to be set.

Keyword	Layer Name	Description
conv2d	Convolution 2D	Convolution 2-Dimentional Layer
pooling2d	Pooling 2D	Pooling 2-Dimentional Layer. Support average / max / global average / global max pooing
flatten	Flatten	Flatten Layer
fully_connected	Fully Connected	Fully Connected Layer
input	Input	Input Layer. This is not always requied.
batch_normalization	Batch Normalization Layer	Batch Normalization Layer.
loss layer	loss layer	hidden from users
activation	activaiton layer	set by layer property

Supported Optimizers

NNTrainer Provides

Keyward	Optimizer Name	Description
sgd	Stochastic Gradient Decent	-
adam	Adaptive Moment Estimation	-

Supported Loss

NNTrainer provides

Keyward	Loss Name	Description
mse	Mean squared Error	-
cross	Cross Entropy - sigmoid	if activation last layer is sigmoid
cross	Cross Entropy - softmax	if activation last layer is softmax

Supported Activations

NNTrainer provides

Keyward	Loss Name	Description
tanh	tanh function	set as layer property
sigmoid	sigmoid function	set as layer property
relu	relu function	set as layer propery
softmax	softmax function	set as layer propery
weight_initializer	Weight Initialization	Xavier(Normal/Uniform), LeCun(Normal/Uniform), HE(Normal/Unifor)
weight_regularizer	weight decay ( L2Norm only )	needs set weight_regularizer_param & type
learnig_rate_decay	learning rate decay	need to set step

Tensor

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. Currently we supports lazy calculation mode to reduce copy of tensors during calcuation.

Keyward	Description
4D Tensor	B, C, H, W
Add/sub/mul/div	-
sum, average, argmax	-
Dot, Transpose	-
normalization, standardization	-
save, read	-

Others

NNTrainer provides

Keyward	Loss Name	Description
weight_initializer	Weight Initialization	Xavier(Normal/Uniform), LeCun(Normal/Uniform), HE(Normal/Unifor)
weight_regularizer	weight decay ( L2Norm only )	needs set weight_regularizer_constant & type
learnig_rate_decay	learning rate decay	need to set step

APIs

Currently we privde C APIs for Tizen. C++ API also provides soon.

Examples for NNTrainer

Custom Shortcut Application

This is demo application which enable user defined custom shortcut on galaxy watch.

MNIST Example

This is example to train mnist dataset. It consists two convolution 2d layer, 2 pooling 2d layer, flatten layer and fully connected layer.

Reinforcement Learning Example

This is reinforcement learning example with cartpole game. It is using deepq alogrightm.

Classification for cifar 10

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.

Tizen CAPI Example

~~This is for demonstrate c api for tizen. It is same transfer learing but written with tizen c api.~~ Deleted instead moved to a test

KNN Example

This is Transfer learning example with cifar 10 data set. TFlite is used for feature extractor and compared with KNN

Logistic Regression Example

This is simple logistic regression example using nntrainer.

Getting Started

Prerequisites

The following dependencies are needed to compile / build / run.

gcc/g++ (>= 4.9, std=c++14 is used)
meson (>= 0.50.0)
blas library (CBLAS) (for CPU Acceleration, libopenblas is used for now)
cuda, cudart, cublas (should match the version) (GPU Acceleration on PC)
tensorflow-lite (>= 1.4.0)
libjsoncpp ( >= 0.6.0) (openAI Environment on PC)
libcurl3 (>= 7.47) (openAI Environment on PC)
libiniparser
libgtest (for testing)

Give It a Go Build with Docker

You can use docker image to easily set up and try building.

To run the docker

$ docker pull lunapocket/nntrainer-build-env:ubuntu-18.04
$ docker run --rm -it  lunapocket/nntrainer-build-env:ubuntu-18.04

Inside docker...

$ cd /root/nntrainer
$ git pull # If you want to build with latest sources.

You can try build from now on without concerning about Prerequisites.

How to Build

Download the source file by cloning the github repository.

$ git clone https://github.com/nnstreamer/nntrainer

After completing download the sources, you can find the several directories and files as below.

$ cd nntrainer

$ ls -1
api
Applications
debian
doc
jni
LICENSE
meson.build
meson_options.txt
nntrainer
nntrainer.pc.in
packaging
README.md
test

$ git log --oneline
f1a3a05 (HEAD -> master, origin/master, origin/HEAD) Add more badges
37032a1 Add Unit Test Cases for Neural Network Initialization
181a003 lower case for layer type.
1eb399b Update clang-format
87f1de7 Add Unit Test for Neural Network
cd5c36e Add Coverage Test badge for nntrainer
...

You can find the source code of the core library in nntrainer/src. In order to build them, use meson

$ meson build
The Meson build system
Version: 0.50.1
Source dir: /home/wook/Work/NNS/nntrainer
Build dir: /home/wook/Work/NNS/nntrainer/build
Build type: native build
Project name: nntrainer
Project version: 0.0.1
Native C compiler: cc (gcc 7.5.0 "cc (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0")
Native C++ compiler: c++ (gcc 7.5.0 "c++ (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0")
Build machine cpu family: x86_64
Build machine cpu: x86_64
...
Build targets in project: 11
Found ninja-1.8.2 at /usr/bin/ninja

$ ninja -C build
ninja: Entering directory `build'
[41/41] Linking target test/unittest/unittest_nntrainer_internal.

After completion of the build, the shared library, 'libnntrainer.so' and the static library, 'libnntrainer.a' will be placed in build/nntrainer.

$ ls build/nntrainer -1
d48ed23@@nntrainer@sha
d48ed23@@nntrainer@sta
libnntrainer.a
libnntrainer.so

In order to install them with related header files to your system, use the 'install' sub-command.

$ ninja -C build install

Then, you will find the libnntrainer.so and related .h files in /usr/local/lib and /usr/local/include directories.

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,

$ ls build/Applications/Training/jni/ -1
e189c96@@nntrainer_training@exe
nntrainer_training

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.

How to Test

Unittest Meson build enable-test set to true

$ echo $(pwd)
(project root)

$ meson build -Denable-test=true
The Meson build system
Version: 0.54.3
...
Configuring capi-nntrainer.pc using configuration
Run-time dependency GTest found: YES (building self)
Build targets in project: 17

Found ninja-1.10.0.git.kitware.jobserver-1 at /home/jlee/.local/bin/ninja

$ ninja -C build test
[79/79] Running all tests.
 1/12 unittest_tizen_capi            OK             8.86s
 2/12 unittest_tizen_capi_layer      OK             0.05s
 3/12 unittest_tizen_capi_optimizer  OK             0.01s
 4/12 unittest_tizen_capi_dataset    OK             0.03s
 5/12 unittest_nntrainer_activations OK             0.03s
 6/12 unittest_nntrainer_internal    OK             0.23s
 7/12 unittest_nntrainer_layers      OK             0.22s
 8/12 unittest_nntrainer_lazy_tensor OK             0.04s
 9/12 unittest_nntrainer_tensor      OK             0.04s
10/12 unittest_util_func             OK             0.05s
11/12 unittest_databuffer_file       OK             0.12s
12/12 unittest_nntrainer_modelfile   OK             2.22s

Ok:                 12
Expected Fail:      0
Fail:               0
Unexpected Pass:    0
Skipped:            0
Timeout:            0

if you want to run particular test

$ meson -C build test $(test name)

Sample app test NNTrainer provides extensive sample app running test.

Meson build with enable-app set to true

$ echo $(pwd)
(project root)

$ meson build -Denable-app=true
The Meson build system
Version: 0.54.3
...
Configuring capi-nntrainer.pc using configuration
Run-time dependency GTest found: YES (building self)
Build targets in project: 17

Found ninja-1.10.0.git.kitware.jobserver-1 at /home/jlee/.local/bin/ninja

$ ninja -C build test
...
 1/21 app_classification             OK             3.59s
 2/21 app_classification_func        OK             42.77s
 3/21 app_knn                        OK             4.81s
 4/21 app_logistic                   OK             14.11s
 5/21 app_DeepQ                      OK             30.30s
 6/21 app_training                   OK             38.36s
 7/21 app_classification_capi_ini    OK             32.65s
 8/21 app_classification_capi_file   OK             32.04s
 9/21 app_classification_capi_func   OK             29.13s
 ...

if you want to run particular example only

$ meson -C build test $(test name) #app_classification_capi_func

Running Examples

Training

After build, run with following arguments Make sure to put last '/' for the resources directory.

$./path/to/example ./path/to/settings.ini ./path/to/resource/directory/

To run the 'Training', for example, do as follows.

$ pwd
./nntrainer
$ LD_LIBRARY_PATH=./build/nntrainer ./build/Applications/Training/jni/nntrainer_training ./Applications/Training/res/Training.ini ./Applications/Training/res/
../../res/happy/happy1.bmp
../../res/happy/happy2.bmp
../../res/happy/happy3.bmp
../../res/happy/happy4.bmp
../../res/happy/happy5.bmp
../../res/sad/sad1.bmp
../../res/sad/sad2.bmp

...

Open Source License

The nntrainer is an open source project released under the terms of the Apache License version 2.0.

anyj0527/nntrainer

NNtrainer

Overview

Maintainer

Reviewers

Components

Supported Layers

Supported Optimizers

Supported Loss

Supported Activations

Tensor

Others

APIs

Examples for NNTrainer

Custom Shortcut Application

MNIST Example

Reinforcement Learning Example

Classification for cifar 10

Tizen CAPI Example

KNN Example

Logistic Regression Example

Getting Started

Prerequisites

Give It a Go Build with Docker

How to Build

How to Test

Running Examples

Open Source License