Home Research
Category:

Research

Data ScienceMachine LearningProgrammingResearchTips/Bugs

Compiling C++ Code Using Caffe

by Jack Simpson
written by Jack Simpson

As part of my PhD project, I have been writing a program in C++ to track hundreds of bees that I have tagged and to identify the pattern on the tags. Initially, I had thought that recognising the tags would be rather simplistic – I could threshold out the tags that are reflective under IR light, and then measure the shape of the object within the tag. Unfortunately, the tags lost their reflectiveness after a short period of time and become smeared (amongst other issues), and thus I had to turn to machine learning to try to find a solution to my recognition problem.

To cut a long story short, the deep learning framework Caffe helped me to solve this problem after a lot of frustration (on the positive side I did learn a lot about machine learning in the process). However, there is one thing that I found surprisingly frustrating when using this framework: compiling your own standalone C++ program that linked to Caffe. For anyone in the same situation as me, this is how to achieve this (keeping in mind I was not using GPUs):

 

  1. At the top of your .cpp program which is linking to the Caffe library, you will need to make the following definition:
    #define CPU_ONLY
  2. Now if we try to compile anything now, Caffe will make this complaint:
    caffe/proto/caffe.pb.h: No such file or directory
    Some of the header files are missing from the Caffe include directory. Thus, you’ll need to generate them with these commands from within the Caffe root directory:
    protoc src/caffe/proto/caffe.proto --cpp_out=.
    mkdir include/caffe/proto
    mv src/caffe/proto/caffe.pb.h include/caffe/proto
  3. Finally, I copied libcaffe.so into /usr/lib and the caffe directory containing the header libraries ($caffe_root/include/caffe) into the /usr/include directory. To compile this on a Mac (after installing OpenBLAS with Homebrew), I just had to run:
    g++ classification.cpp -lcaffe -lglog -lopencv_core -lopencv_highgui -lopencv_imgproc -I /usr/local/Cellar/openblas/0.2.14_1/include -L /usr/local/Cellar/openblas/0.2.14_1/lib -o classifier
  4. Alternatively, you could do what I did on my Linux machine and instead of copying header files, I just linked directly to those directories when I compiled:
    g++ classification.cpp -lcaffe -lglog -lopencv_core -lopencv_highgui -lopencv_imgproc -I ~/caffe/include -L ~/caffe/build/lib -I /usr/local/Cellar/openblas/0.2.14_1/include -L /usr/local/Cellar/openblas/0.2.14_1/lib -o classifier
0 comment
0 FacebookTwitterPinterestEmail
Image ProcessingResearchTips/Bugs

Be careful with FFmpeg metadata

by Jack Simpson
written by Jack Simpson

Recently I used the program FFmpeg to automatically extract metadata about the time and date some videos I’ve been using for my research were created. FFmpeg is a really useful tool for manipulating videos and images at the command-line. You can do things like change the format, quality or length of your video files with relative ease. The image processing library OpenCV even uses FFmpeg under the hood for opening videos you want to process. You can also use FFmpeg to extract metadata about your video files, and that is what I was using it for. However, I soon discovered that the metadata I retrieved with the program wasn’t completely accurate. Video file creation times were off by a couple of hours. Normally that doesn’t matter much, however if you’re breaking your data down hour-by-hour over a 2 week period, having these kinds of discrepancies really throws your analysis out. It was a great reminder to always thoroughly check the output of your analysis pipelines and to never assume that a tool will work perfectly all the time.

0 comment
0 FacebookTwitterPinterestEmail
Data ScienceResearch

iPython Notebook Presentation

by Jack Simpson
written by Jack Simpson

I have a talk today about how I’ve been using the iPython Notebook when prototyping different image processing techniques and their parameters. The presentation involved a bit of background on how to use the notebook and markdown, before I demonstrated how you could use the notebook to rapidly test out the best way to process the images I’ve been working on. In this case the images were taken from infrared footage of bees, and I wanted to separate the tags out from the background. I found simple thresholding worked well at first while the tags were very reflective, but once they became a bit more dull, that ceased to work effectively.

You can find the notebook I used for my talk here, GitHub renders the notebooks in the browser. I have cleared all the output images the file generated because the file was getting a bit large to view online.

0 comment
0 FacebookTwitterPinterestEmail
ResearchResources

Interesting Articles on Programming, Organisation and Bioinformatics

by Jack Simpson
written by Jack Simpson

Programming

Organisation

Bioinformatics

0 comment
0 FacebookTwitterPinterestEmail
Image ProcessingOpenCVResearch

Tracking bees using local features

by Jack Simpson
written by Jack Simpson

The project I am working on at present is focused on building an automated system for tracking the movement of honeybees in an observation hive, filmed using a grayscale camera and infrared light. My recent attempt at this has involved extracting features (or keypoints) from the frames and then describing them using the SIFT (Scale Invariant Feature Transform) algorithm, implemented in the OpenCV imaging library. There are a number of other (potentially faster) feature detection algorithms available (SURF, ORB, etc), however as SIFT has been traditionally regarded as the most robust, I thought I would start with it. SIFT works first by extracting what it deems to be robust keypoints/features. For each of these, it will then compute a unique feature descriptor based on the 16×16 pixel area of the feature. By extracting and describing these features, you can try to compare images as shown below.

Continue Reading
0 comment
0 FacebookTwitterPinterestEmail
ResearchThoughts

Essential Free Tools for Research

by Jack Simpson
written by Jack Simpson

I commenced my PhD in bioinformatics at the Australian National University about a month ago and thought I’d share some of the tools I’ve found absolutely essential. All of these tools are free to use (some do have paid plans, however at this point I haven’t had the need to sign up for them).

  1. Evernote + Evernote Web Clipper: Once I started to really get into using this tool, I couldn’t understand how I’d lived without it before. Now I can clip snippets from web pages for later reference, write notes from a seminar I attended or upload a Python script and have this all in one easily searchable (and tag-able) location. There’s also a plethora of phone/tablet apps which work really well.
  2. Wunderlist: While I tried to use Evernote for keeping to-do lists, I found that this simple yet powerful program (with support for PC, Mac, phones and tablets) worked extremely well for keeping track of my tasks.
  3. Dropbox: I do all my work within the Dropbox folder on my computer. This way, when I move between computers everything is synced and backed-up. While I have tried Google Drive before, the lack of support for Linux as well as issues with syncing have led me back to using Dropbox.
  4. Zotero: Zotero is a referencing manager with a browser plug-in which is fantastic at importing the bibliographic details and pdf of the paper from the web. It can also sync between computers (although you only get 300 mb of free sync space). While I have used the alternative referencing manager Mendeley before (which has 2gb free sync space and a brilliant program which enables you to annotate your pdfs), the Mendeley web import tool was so bad at importing the bibliographic details and pdfs of papers that I went over to Zotero.
  5. VirtualBox: If you find yourself having to use a lot of Linux-only tools (or you like the fact that you can install entire programs with multiple dependencies merely with the command ‘sudo apt-get’) but don’t really want to set up anything permanent , then VirtualBox could be the way to go. Its a free program that allows you to run a whole operating system on your computer within a window just like any other program.
  6. Coursera/EdX/Udacity: These websites offer free courses from great unis in everything from computer science to genomics.
  7. Stack Overflow/Biostars: Great websites for asking questions relating to programming or bioinformatics. If you can’t find anything after searching Google, then this is the place to post your question.
  8. Twitter: Although this can be a bit of a time-waster, I find that by following other researchers in my field and research organisations I can hear about new papers or tutorials.
  9. Google Plus: G+ may not be as popular as Facebook, but I find most of my programming friends tend to be quite active on it and if you join one of the programming communities it can be a great place to learn new things or ask questions.
  10. Google Scholar Alerts: Sign up to receive emails whenever the keywords you select are published in a paper.
0 comment
0 FacebookTwitterPinterestEmail
ProgrammingResearchTutorial

BioPython Tutorial

by Jack Simpson
written by Jack Simpson

This tutorial is a brief overview of what you can achieve using the Python BioPython module. Although I’m hoping to write up some more articles on this site for beginners when time permits, this post will assume that you have experience programming in Python and have a bit of an understanding of basic biological concepts such as DNA, restriction enzymes etc. If you’re still interested once you finish reading, feel free to consult the BioPython documentation, it will help give you a bit of an idea of how massive (and awesome) this module really is.

So to start I’ll show you how to install the BioPython module. While on Linux systems it can be as simple as typing ‘sudo apt-get install python-biopython’ or going to the Software Center, you can manually install a module by going to PyPI, downloading and extracting the file, opening the command-line or terminal and navigating into the root directory of the folder you just extracted and running the command ‘setup.py install’.

You will need to download two modules to install BioPython, each of which are hosted on their own site. The first is SciPy and the second is BioPython. Once you have installed these you’re ready to get into using BioPython.

Continue Reading
0 comment
0 FacebookTwitterPinterestEmail
ProgrammingResearchTutorial

Pyral Project

by Jack Simpson
written by Jack Simpson

Pyral (Python + Viral) was the name of a project I worked on in Dr Joanne Macdonald’s lab between September 2012 – January 2013 (although I am still providing tech support for the code and helping manage the server to this date). Throughout this time I wrote a lot of Perl and Python code to run on the university’s Linux server. The aim of these programs were as follows:

  • Download all the viral ref-seq genomes from GenBank;
  • BLAST a sequence of interest and retrieve all similar files;
  • Concatenate all sequences into one file that was run through CD-HIT;
  • Analyse the CD-HIT output, returning a file with the cluster numbers that sequences of interest may be found in;
  • Find variable length conserved regions of DNA within a designated cluster;
  • Ensure conserved region of DNA is completely dissimilar to that found in other virus clusters.
    Continue Reading
0 comment
0 FacebookTwitterPinterestEmail
Newer Posts
Older Posts