arxiv sanity preserver

This project is a web interface that attempts to tame the overwhelming flood of papers on Arxiv. It allows researchers to keep track of recent papers, search for papers, sort papers by similarity to any paper, see recent popular papers, to add papers to a personal library, and to get personalized recommendations of (new or old) Arxiv papers. This code is currently running live at www.arxiv-sanity.com/, where it's serving 15,000+ Arxiv papers from Machine Learning (cs.[CV|CL|LG|NE]/stat.ML) over the last ~3 years. I am looking for collaborators who wish to try to get this running for other parts of Arxiv as well (e.g. theory? physics?)

Code layout

There are two large parts of the code:

Indexing code. Uses Arxiv API to download the most recent papers in any categories you like, and then downloads all papers, extracts all text, creates tfidf vectors based on the content of each paper. This code is therefore concerned with the backend scraping and computation: building up a database of arxiv papers, calculating content vectors, creating thumbnails, computing SVMs for people, etc.

User interface. Then there is a web server (based on Flask/Tornado/sqlite) that allows searching through the database and filtering papers by similarity, etc.

Dependencies

Several: You will need numpy, feedparser (to process xml files), scikit learn (for tfidf vectorizer, training of SVM), flask (for serving the results), and tornado (if you want to run the flask server in production). Also dateutil, and scipy. And sqlite3 for database (accounts, library support, etc.). Most of these are easy to get through pip, e.g.:

$ virtualenv env                # optional: use virtualenv
$ source env/bin/activate       # optional: use virtualenv
$ pip install -r requirements.txt

You will also need ImageMagick and pdftotext, which you can install on Ubuntu as sudo apt-get install imagemagick poppler-utils.

Processing pipeline

I tried to keep the project code relatively clean, but I do encourage you to skim each script when you run it. There are a few magic numbers here and there. In order, the processing pipeline is:

1. Run fetch_papers.py to query arxiv API and create a file db.p that contains all information for each paper. This script is where you would modify the query, indicating which parts of arxiv you'd like to use. Note that if you're trying to pull too many papers arxiv will start to rate limit you. You may have to run the script multiple times, and I recommend using the arg --start_index to restart where you left off when you were last interrupted by arxiv.
2. Run download_pdf.py, which iterates over all papers in parsed pickle and downloads the papers into folder pdf
3. Run parse_pdf_to_text.py to export all text from pdfs to files in txt
4. Run thumb_pdf.py to export thumbnails of all pdfs to thumb
5. Run analyze.py to compute tfidf vectors for all documents based on bigrams. Saves a tfidf.p, tfidf_meta.p and sim_dict.p pickle files.
6. Run buildsvm.py to train SVMs for all users (if any), exports a pickle user_sim.p
7. Run the flask server with serve.py (and make sure to run sqlite3 as.db < schema.sql if this is the very first time ever you're starting arxiv-sanity, which initializes an empty database). Visit localhost:5000 and enjoy sane viewing of papers!

I have a simple shell script that runs these commands one by one, and every day I run this script to fetch new papers, incorporate them into the database, and recompute all tfidf vectors/classifiers. More details on this process below.

protip: numpy/BLAS: The script analyze.py does quite a lot of heavy lifting with numpy. I recommend that you carefully set up your numpy to use BLAS (e.g. OpenBLAS), otherwise the computations will take a long time. With ~15,000 papers and ~500 users the script runs in about half an hour on my current machine with a BLAS-linked numpy.

Running online

If you'd like to run this flask server online (e.g. AWS) run it as python serve.py --prod.

You also want to create a secret_key.txt file and fill it with random text (see top of serve.py).

Current workflow

Running the site live is not currently set up for a fully automatic plug and play operation. Instead it's a bit of a manual process and I thought I should document how I'm keeping this code alive right now. I have two machines: a local machine that does a lot of the updating and compute and a remote machine that hosts the site.

I have a script that performs the following update early morning on my local machine:

# pull the database (by default stored in as.db) from remote to local
rsync -v karpathy@REMOTE:/home/karpathy/arxiv-sanity-preserver/as.db as.db

# now perform the update and recomputation:
python fetch_papers.py
python download_pdfs.py
python parse_pdf_to_text.py
python thumb_pdf.py
python analyze.py
python buildsvm.py

# now rsync the results and new thumbnails from local to remote
rsync -v db.p tfidf_meta.p sim_dict.p user_sim.p karpathy@REMOTE:/home/karpathy/arxiv-sanity-preserver
rsync -vr static/thumbs karpathy@REMOTE:/home/karpathy/arxiv-sanity-preserver/static

Of course, I had to set up the ssh keys so that rsync/ssh commands can run without needing password. I think log on to the remote machine and restart the server. I run the server in a screen session, so I ssh to REMOTE, screen -r the screen session, and restart the server:

python serve.py --prod --port 80

The server will load the new files and begin hosting the site. Yes, currently the server has to be restarted, so the site goes down for about 15 seconds. There are several ways to make this cleaner in the future. Note that on some systems you can't use port 80 without sudo. Your two options are to use iptables to reroute ports, or less recommended: you can use setcap to elavate the permissions of your python interpreter that runs serve.py. In this case I'd recommend careful permissions and maybe virtualenv, etc.