Run it preferably on GNU/Linux. Latest version available to you of Git and Docker (google them to find their installation instructions for your OS).

Clone this repository:

git clone

Run the build.sh, then start_jupyter, both with sudo rights if you haven't set your user so that you can use docker commands without being root.

sudo ./build.sh
sudo ./start_jupyter.sh

The first command log should end with a message like the following:

Successfully built 554f78aa2d91
Successfully tagged deep:latest

The second one should return something like:

52aac7bf9976a85dd9b98258dc350bbe790acb94152c33884f6b3885a21bf31a

You can check whether the container started properly by writing:

sudo docker ps -a

If everything is fine you should get a line like the following (what matters here is that the "STATUS" field is "UP"):

CONTAINER ID        IMAGE               COMMAND                  CREATED       
52aac7bf9976        deep                "/bin/sh -c 'jupyter…"   13 minutes ago
STATUS                    PORTS                    NAMES
Up 13 minutes             0.0.0.0:8887->8888/tcp   deep_jupyter

Once this is done, you should have a "data" folder in your cloned repository. Copy the test_images and train_images data folders into it. Then open a bash in the container :

sudo docker exec -it deep_jupyter /bin/bash

And run the following commands inside to build the image database :

cd data

/opt/caffe/build/tools/convert_imageset --shuffle --gray train_images/ posneg.txt train_lmdb

/opt/caffe/build/tools/convert_imageset --shuffle --gray test_images/ testposneg.txt test_lmdb

/opt/caffe/build/tools/compute_image_mean train_lmdb/ train_mean.binaryproto

/opt/caffe/build/tools/compute_image_mean test_lmdb/ test_mean.binaryproto

You can then close your terminal and open your favorite web browser to http://localhost:8887/ . This will open the Jupyter Notebook instance that is inside the container.

You will get to a login form. In the "Password or token" field, enter "demo". You will then be redirected to the usual Jupyter Notebook interface. Open the "notebooks" folder, and then the "01_CNN_Training.ipynb" if you want to train the neural network, "02_CNN_Usage.ipynb" if you want to use a pre-trained one or your own, or "03_CNN_Image_Parsing.ipynb" if you wish to use the neural network to detect faces in a given set of images.

To change the number of threads used by caffe during training, open the "./caffe-intel-4threads/Dockerfile" file, change the "ENV OMP_NUM_THREADS=4" value to whichever number of threads you wish to use and restart the build process (run build.sh and start_jupyter.sh after making sure you stop and removed your previous container with "sudo docker stop deep_jupyter && sudo docker rm deep_jupyter"). No need to rebuild the image database, since it was stored in a shared folder with the host machine.

• Stefan Duffner's website (where test data and latest version of the assignment can be found): http://www.duffner-net.de/

• Caffe Official Tutorial to get to know the tool: http://caffe.berkeleyvision.org/tutorial/

• Official Caffe Docker Image, for easy deployment: https://github.com/BVLC/caffe/tree/master/docker

• Caffe with Docker Tutorial, base for our own Dockerfile: https://towardsdatascience.com/playing-with-caffe-and-docker-to-build-deep-learning-models-99c9570ffc3d

• Deep Learning with Caffe tutorial, to get started: http://christopher5106.github.io/deep/learning/2015/09/04/Deep-learning-tutorial-on-Caffe-Technology.html

• Custom Face Images Dataset ; http://www.vision.caltech.edu/Image_Datasets/Caltech_10K_WebFaces/Caltech_WebFaces.tar (comes from this website)