SEGNET-FACADE

Requirements (or Recommendations)

• An NVIDIA graphics card / GPU with a fair amount of RAM. I use an 11Gb card for training and inference, you may be able to get by with less. During inference, my system reports that a bit over 4Gb is used, during training it is closer to 8Gb. You may be able to change the code to use cpu mode but I have not run this code on a machine without NVIDA hardware.
• A system with a decent amount of RAM. I expect that 32Gb would be a good number; all of my machines have much more than that.

Jupyter Demo

I have put jupyter on the docker image so that you can follow along with a simple tutorial in scripts/process.ipynb. I will try to describe how to use it here:

1. Figure out which domain-name and port you want to use for a jupyter server; the default IP address is localhost, the default port is port 8888, but you may already be serving a jupyter notebook on that port (if you are not currently using jupyter, just remember that the port is 8888). Once you have settled on a port, let's save it to an environment variable:

   export MYIP=127.0.0.1
   export MYPORT=8888

   NOTE: If you are not on serving from a local machine, you will need the actual IP adress and not 127.0.0.1. You can use the ifconfig program and look for inet addr: in the output. You will potentially see many devices listed in the output of ifconfig, you most likely want to look at the output under eth0.

2. You will probably want to produce output to a folder on your computer, and you may also want to provide your own inputs via a folder on your computer. For now, I will asume that that the input and output will be placed under your /tmp directory. Let's set some environment variables for our input and output locations.

   mkdir -p /tmp/segnet-facade/output
   mkdir -p /tmp/segent-facade/input
   export MYINPUT=/tmp/segent-facade/input
   export MYOUTPUT=/tmp/segnet-facade/output

3. Now you can run a jupyter server using my docker image:

   nvidia-docker run -v "${MYOUTPUT}":/output -v "${MYINPUT}":/data -p ${MYPORT}:${MYPORT} \
       jfemiani/segnet-facade jupyter notebook --allow-root --ip=* --port ${MYPORT} \
       /opt/facades/scripts/process.ipynb

   NOTE If you are using a newer GPU than the old K40 I had, you will need an image with a newer version of cuda. I build a machine for cuda8 and tagged it cuda8-cudnn3, as shown below:

   nvidia-docker run -v "${MYOUTPUT}":/output -v "${MYINPUT}":/data -p ${MYPORT}:${MYPORT} \
      jfemiani/segnet-facade:cuda8-cudnn3 jupyter notebook --allow-root --ip=* --port ${MYPORT} \
      /opt/facades/scripts/process.ipynb

4. You should see a link on your terminal to http://localhost:8888/<a bunch of randomy text>. Copy and paste the link into your browser.

5. Run each line of the jupyter notebook in order to go through the segmentation process one step at a time. You may also edit those cells to process your own data or to use the CPU instead of the GPU. For example you can replace the line that says

   caffe.set_device(0)
   caffe.set_mode_gpu()

   with this:

   caffe.set_mode_cpu()

   You will need a good amount of RAM available (e.g. 32Gb).

Command Line Demo

To start the docker image you may also use:

nvidia-docker run -it \
    -v "/media/femianjc/My Book/facade-output/":/output \
    -v "/home/shared/Projects/Facades/src/data/":/data \
    jfemiani/segnet-facade

which will put you in a bash shell, with the code from this git repo under /opt/facades

NOTE: You should, of course, replace those paths with ones that make sense on your system.

Once you are in the docker image's shell, type:

cp -r /opt/facades/data/cock-congo /data
cp /opt/facades/scripts/jobs/one-file.txt /data

inference /data/one-file.txt

If everything works, this will copy an example input file into the /data folder and produce output (including figures and a YAML file with facade elements) in the /output folder.

Some explanations:

1. The -v tag allows you to map folders on your host computer with volumes in the docker image.
2. Most of the default-arguments to my scripts will assume that input imagery is under the /data folder. Usually you can change this by passing other paths or changing config files, but if you use a docker container you can also just map whatever folder has your files to the /data volume in the docker image.
3. My scripts, by default, write output to the /output folder.

The inference script just runs a python module that does the real work, you can use that directly to get more control of the output. Try:

python -m pyfacades --help

To get the most up-to-date help on how to use the script.

Installation

You may may use nvidia-docker to run this code. I have already built a machine and put it on Dockerhub jfemiani/segnet-facade . You should be able to use that once nvidia-docker is installed, but you may also wish to build your own image. To do that run

cd docker/gpu && source build-docker-image.sh

You may also consider the Dockerfile in docker/gpu/Dockerfile as instructions on how to configure your own machine. In this README I will assume you have followed the installation instructions for nvidia-docker which include putting your current user in 'dockers' group.

Development

I am using PyCharm to develop this code on an Ubuntu Machine with a Tesla K40 GPU.