kaggle-statefarm

detecting distracted drivers

Set up

Note

• The following instructions have been tested on an AWS g2.2xlarge instance running ubuntu 14.04
• We will set up a notebook server listening to port 8888 for incoming https requests

Download and Install NVIDIA Driver

Read this article for details. In short run the following in order

# install dependencies
sudo apt-get update
sudo apt-get install build-essential linux-image-extra-virtual

# download and extract the nvidia installer
wget http://developer.download.nvidia.com/compute/cuda/6_5/rel/installers/cuda_6.5.14_linux_64.run
chmod +x cuda_6.5.14_linux_64.run
mkdir nvidia_installers
./cuda_6.5.14_linux_64.run -extract=`pwd`/nvidia_installers
cd nvidia_installers

# run nvidia installer and follow the instruction (press Enter to accept the default settings)
sudo ./NVIDIA-Linux-x86_64-340.29.run
sudo modprobe nvidia
sudo ./cuda-linux64-rel-6.5.14-18749181.run
sudo ./cuda-samples-linux-6.5.14-18745345.run

# verify the installation
cd /usr/local/cuda/samples/1_Utilities/deviceQuery
sudo make
./deviceQuery
# you should see "Result = PASS" at the end of the command line output

Install Docker

For detailed instructions see this article. In short run the following in order

sudo apt-get update
sudo apt-get install apt-transport-https ca-certificates
sudo apt-key adv --keyserver hkp://p80.pool.sks-keyservers.net:80 --recv-keys 58118E89F3A912897C070ADBF76221572C52609D
# http://askubuntu.com/questions/103643/cannot-echo-hello-x-txt-even-with-sudo
echo "deb https://apt.dockerproject.org/repo ubuntu-trusty main" | sudo tee /etc/apt/sources.list.d/docker.list
sudo apt-get update
sudo apt-get install linux-image-extra-$(uname -r) apparmor
sudo apt-get install docker-engine
sudo usermod -aG docker ubuntu
exit
ssh ...

Run the docker container

You can simply pull this docker container that has everything set, by running docker pull evau23/kaggle-statefarm

• run_jupyter.sh is shell script to help with running the notebook server in a docker container. You can find the original script here

Or build the container yourself by running

git clone https://github.com/alireza-a/kaggle-statefarm.git
cd ./kaggle-statefarm
docker build -t name:tag . # make sure to change the name and tag in the following steps

Now you can run the container by running the following

# http://tleyden.github.io/blog/2014/10/25/running-caffe-on-aws-gpu-instance-via-docker/
DOCKER_NVIDIA_DEVICES="--device /dev/nvidia0:/dev/nvidia0 --device /dev/nvidiactl:/dev/nvidiactl --device /dev/nvidia-uvm:/dev/nvidia-uvm"

docker run $DOCKER_NVIDIA_DEVICES -d -p 8888:8888 -e "PASSWORD=your_password" --name statefarm evau23/kaggle-statefarm bash /run_jupyter.sh

Fantastic, now you have a notebook server with all the dependencies running

Fine tuning Caffe_Net

The Caffe_net.ipynb has been adopted from the 02-fine-tuning.ipynb. In short we will do the followings

some set up

1. Import the required libraries
2. load pretrained weights, labels, and mean image
3. load StateFarm data

• download StateFarm data
  • make sure you have read the README file under data directory
• define train and valid sets

some definitions

1. define caffenet

• to define caffeNet architecture

2. define solver

• to initialize caffe solver for training the model

3. define kaggle_net

• to initialize caffenet for our dataset

4. define run_solver

• a wrapper to step the model

train the model

1. initialize and run the solver

• specify the number of iterations

2. visualize the loss and accuracy

extract probabilities

1. define helper functions

• to initialize a transformer
• to preprocess a batch of images
• to extract probabilities from the batch of images
• to write the extracted probabilities into a file

2. extract probabilities for each batch of images and write it to a file