Final project of the Embedded System course based on the realization of a neural network and its implementation on Raspberry Pi model 3 and Intel Movidius neural compute stick.

minimum version:

• python 3.5, 3.6 actually 3.7 is not supported:

absl-py==0.4.1
astor==0.7.1
cycler==0.10.0
gast==0.2.0
graphviz==0.9
grpcio==1.14.2
h5py==2.8.0
Keras==2.2.2
Keras-Applications==1.0.4
Keras-Preprocessing==1.0.2
kiwisolver==1.0.1
Markdown==2.6.11
matplotlib==2.2.3
numpy==1.14.5
opencv-python==3.4.3.18
pandas==0.23.4
Pillow==5.2.0
protobuf==3.6.1
pydot==1.2.4
pyparsing==2.2.0
PyQt5==5.11.3
PyQt5-sip==4.19.13
python-dateutil==2.7.3
pytz==2018.5
PyYAML==3.13
scikit-learn==0.19.2
scipy==1.1.0
seaborn==0.9.0
selenium==3.14.0
six==1.11.0
sklearn==0.0
tensorboard==1.10.0
tensorflow==1.10.1
termcolor==1.1.0
urllib3==1.23
Werkzeug==0.14.1  

The packages needed are enclosed in file “requirements.txt“, to install, type in the terminal:

pip3 install -r requirements.txt

NB: if you want run the script with CUDA is necessary install "tensorflow-gpu" by type:

pip3 install tensorflow-gpu

To clone the repository can use the following command

git clone https://github.com/frank1789/NeuralNetworks.git

To run on Host/Guest machine use:

cd NeuralNetworks
sh test.sh

To run on guest machine with Intel Movidius use:

cd NeuralNetworks
sh test.sh movidius

To run on Raspberry Pi 3 model B with Intel Movidius use:

cd NeuralNetworks
sh test.sh reduced-nocompile

remember the neural compute strick must be attached.

Since the development kit provided by intel works in Linux Ubuntu environment it is recommended to install and configure a virtual machine:

• download VirtualBox and VirtualBox Extension Pack
• download image iso Linux Ubuntu 16.04 LTS (higher OS version are not supported by movidius software)

Proceed with the installation of: VirtualBox, extension package, OS.

Once completed, we configure the virtual guest machine as shown in the following pictures.

To install NCSDK 2.x you can use the following command to clone the ncsdk2 branch

git clone -b ncsdk2 https://github.com/movidius/ncsdk.git

The provided Makefile helps with installation. Clone the repository and then run the following command to install the NCSDK:

make install

The Neural Compute SDK also includes examples. After cloning and running 'make install,' run the following command to install the examples:

make examples

This script allows you to organize a dataset, downloaded from the internet or made in-house, as a structure of folders containing sets for training, validation and testing of the neural network.

This structure is congenial for use with Keras specifically with the flow_from_directory method. Be aware of the fact that if the folders are empty, the result will be a reduced dataset because it will automatically skip. It is necessary to pass in argument:

• absolute path folder containing the raw dataset (-d);
• absolute path folder containing the raw test set (-t);
• integer value between 0 100 for dividing the dataset (-s).

python3 makeDataset.py -d ./data -t ./test -s 30

After this it is possible to begin to train the neural network through the script 'name' passing in argument:

python3 train.py -d ..data/train -v ../data/validate -e 10 -b 10 -is 224 224 -n vgg16

The following neural networks are available within the script

• VGG16 (lower case for script argument)
• VGG19 (lower case for script argument)
• InceptionV3 (aka 'inception' argument script)
• Xception (lower case for script argument)
• ResNet50 (aka 'resnet50' argument script)

Once you have trained neural network use model file for conversion in graph format to work on Neural Compute Stick, type command:

python3 keras2ncsgraph.py -k model.h5 -n converted_model

The report is available on this branch and can be downloaded from the command:

git clone -b report https://github.com/frank1789/NeuralNetworks.git

This project is licensed under the MIT License - see the LICENSE.md file for details

• Embedded system lab @UNITN
• HPC Cluster | ICTS - University of Trento - ICTS@unitn