DEEP PI

An example of running deep neural-net image classifier on Raspberry PI

Installing torch on Raspbian

If you are using minimal version of Raspbian, you will need to install several packages first:

sudo apt-get install -y build-essential gcc g++ curl  cmake libreadline-dev libjpeg-dev libpng-dev ncurses-dev imagemagick gfortran libopenblas-base libopenblas-dev

The clone torch distribution:

git clone https://github.com/torch/distro.git ~/torch --recursive

And start building (takes several hours on Raspberry PI B+:

cd ~/torch
./install.sh

If you encounter following error: ...In function ‘THByteVector_vectorDispatchInit’: /home/pi/torch/pkg/torch/lib/TH/generic/simd/simd.h:64:3: error: impossible constraint in ‘asm’ ...
it means that you are building on a cpu without NEON extension (the kind Raspberry PI Version A & B have). You will need to checkout latest version of torch and disable submodule update command in install.sh script ( comment out line 45 in ~/torch/install.sh ) and then update torch torch:

cd ~/torch/pkg/torch/
git checkout master
git pull

and run ./install.sh script again.

After ./install.sh is finished - it will ask if you want to update .bashrc to include call to initialize torch environment every time you login. If you don't want it, you will have to execute command . ~/torch/install/bin/torch-activate before you will be able to lauch th.

Alternative way to install torch on Raspbian, using precompiled blob

I created an archive of torch installation compiled for Raspberry PI B+ , running Raspbian 8 You can download it here : https://github.com/vfonov/deep-pi/releases/download/v1/torch_intstall_raspbian_arm6l_20161218.tar.gz Copy file to /home/pi, then run tar zxf torch_intstall_raspbian_arm6l_20161218.tar.gz - this will create torch subdirectory that will include only precompiled binaries. To activate it add . torch/install/bin/torch-activate in the end of the ~/.bashrc file.

Running MNIST digit classifier from torch demos

You can install various torch example from https://github.com/torch/demos, here is an output from MNIST digit classieifer training session:

pi:~/src/demos/train-a-digit-classifier $ th train-on-mnist.lua 
<torch> set nb of threads to 4	
<mnist> using model:	
nn.Sequential {
  [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> (9) -> (10) -> output]
  (1): nn.SpatialConvolutionMM(1 -> 32, 5x5)
  (2): nn.Tanh
  (3): nn.SpatialMaxPooling(3x3, 3,3, 1,1)
  (4): nn.SpatialConvolutionMM(32 -> 64, 5x5)
  (5): nn.Tanh
  (6): nn.SpatialMaxPooling(2x2, 2,2)
  (7): nn.Reshape(576)
  (8): nn.Linear(576 -> 200)
  (9): nn.Tanh
  (10): nn.Linear(200 -> 10)
}
<warning> only using 2000 samples to train quickly (use flag -full to use 60000 samples)	
<mnist> loading only 2000 examples	
<mnist> done	
<mnist> loading only 1000 examples	
<mnist> done	
<trainer> on training set:	
<trainer> online epoch # 1 [batchSize = 10]	
 [===================>.................... 471/2000 ....................................]  ETA: 2m20s | Step: 92ms      

Overall it is about 5 times slower then running the same example on a desktop with Core i5 @ 3.30GHz without using GPU.

Installing deep-pi

git clone https://github.com/vfonov/deep-pi 

After that you can launch download_net.sh script to download the pretrained NIN network ( based on https://gist.github.com/szagoruyko/0f5b4c5e2d2b18472854 ) to the /home/pi path. WARNING pretrained network is 33Mb file!

Running

To run on a single image: th test_single.lua <path to your image> To run continious classification using frames from camera ( I recommend using external USB camera) :

nohup th -ldisplay.start 8000 0.0.0.0 & 
th camera_interface.lua

Then open web browser and point to to location http://your.raspberry.ip:8000 - replace your.raspberry.ip with IP address that your Raspberry PI is configured to use.

Setup

Output