NLE practical session for PAISS 2018

This code requires Python 3 and Pytorch 0.4. Follow the instructions below to install all the necessary dependencies.

First, download and install the appropriate version of miniconda following the instructions for MacOS or Linux.

Then run the following commands:

source $HOME/miniconda3/bin/activate #Activates your conda environment
conda install numpy matplotlib ipython scikit-learn
conda install pytorch torchvision faiss-cpu -c pytorch

On MacOS there’s a bug for faiss related to libomp (facebookresearch/faiss#485): run “brew install libomp” (see https://brew.sh/ to install brew) to resolve this bug.

Install anaconda on windows (launch .exe file downloaded from the conda website). It has to be python 3 (pytorch doesn’t support 2.7 on windows)

In the anaconda prompt, run:

conda create -n pytorch
activate pytorch
conda install pytorch-cpu -c pytorch
pip install torchvision --no-deps
conda install pillow

NOTE: The FAISS package is not supported on Windows. Participants with Windows machines must follow the product quantization exercise with their neighbours.

First, clone this repository:

cd $HOME/my_projects
git clone https://github.com/almazan/paiss.git

Then, you will need to download 4 files:

• oxbuild_images.tgz (1.8GB)
• gt_files_170407.tgz (280KB)
• features.tgz (579MB)
• models.tgz (328MB)

and store them in the appropriate paths.

Note: All paths in this section are relative to the root directory of this repository.

Oxford dataset

On Linux/MacOS, execute the following:

cd $HOME/my_projects/paiss
wget www.robots.ox.ac.uk/~vgg/data/oxbuildings/oxbuild_images.tgz -O images.tgz
mkdir -p data/oxford5k/jpg && tar -xzf images.tgz -C data/oxford5k/jpg
wget www.robots.ox.ac.uk/~vgg/data/oxbuildings/gt_files_170407.tgz -O gt_files.tgz
mkdir -p data/oxford5k/lab && tar -xzf gt_files.tgz -C data/oxford5k/lab

On Windows, perform the following:

• Download www.robots.ox.ac.uk/~vgg/data/oxbuildings/oxbuild_images.tgz
• create directory data/oxford5k/jpg/
• uncompress oxbuild_images.tgz and store in data/oxford5k/jpg/
• Download www.robots.ox.ac.uk/~vgg/data/oxbuildings/gt_files_170407.tgz
• create directory data/oxford5k/lab/
• uncompress gt_files_170407.tgz and store in data/oxford5k/lab/

Features and models

On Linux/MacOS, execute the following:

cd $HOME/my_projects/paiss
wget https://www.dropbox.com/s/gr404xlfr4021pw/features.tgz?dl=1 -O features.tgz
tar -xzf features.tgz -C data
wget https://www.dropbox.com/s/mr4risqu7t9neel/models.tgz?dl=1 -O models.tgz
tar -xzf models.tgz -C data

On Windows, perform the following:

• Download https://www.dropbox.com/s/gr404xlfr4021pw/features.tgz?dl=1
• Uncompress features.tgz and store in data/
• Download https://www.dropbox.com/s/mr4risqu7t9neel/models.tgz?dl=1
• Uncompress models.tgz and store in data/

Execute:

source $HOME/miniconda3/bin/activate
cd $HOME/my_projects/paiss
python demo.py --qidx 42 --topk 5

and you should see the following ouput:

To go through the practical session, we will use the session.py script. This scrip contains all the exercises and questions in the comments, and accepts the following arguments:

• --sect [str]: Selects the section of the session to be run. You can select one section between 1a to 1i, and 2a to 2g.
• --qidx [int]: Selects the query index that you want to use for the section. Each section has already some default indexes, so this argument is not strictly required.
• --hide-tsne: This flag desactivate the tsne computation. In some sections you may want to experiment with different queries and not to visualize the TSNE projection again (since this may take between 30 seconds to several minutes depending on your CPU :S)

We recommend running the session from iPython using the command %run, since this allows you to visualize the variables that were created during the execution (eg. you may want to check a given layer of the network). Here's an example:

source $HOME/miniconda3/bin/activate
cd $HOME/my_projects/paiss
ipython
> %run session.py --sect 1i --hide-tsne --qidx 23
> qfeats.shape
> qfeats[q_idx]

Which will show the following results:

Note: If you don't have iPython installed, you can install it using the following conda command: conda install ipython

Note: The code is automatically saving the plots generated in the folder results/

Link to the slides: PDF 30MB