Having been curious about Functional Programming for a while now, and tried incorporating features of the paradigm into my own work with Python, I decided to give the first (Part A) of the three-part Programming Languages course module on Coursera. The module is meant to systematically introduce one to various theoretical concepts of programming languages, while having a special focus on Functional Programming. This first course (Part A), which I recently completed with a score of 98%, illustrated said concepts with the help of Standard ML – a Functional-style language.
It was excellently designed course, and also quite challenging. Apart from spending time on introducing the very basics of SML early on, it covered some very interesting concepts such as Pattern Matching, Function Closures, Partials, Currying and Mutual Recurstion. The programming assignments really made sure you understood what was covered in the course material, and the course-handouts were thorough and clear. There was also a strong focus on the matter of programming style, with the instructor commenting on what he considered good/poor programming style while covering the various concepts. We were marked on the style of our submissions too.
It’s just been confirmed that four of us from Moodagent – Reinier de Valk, Pierre Lafitte, Tomas Gajarsky and I, will be attending ISMIR 2019 in Delft (The Netherlands). This year, two of my colleagues from Moodagent will be presenting their work at ISMIR:
It’s been about four months since I wrote here about leaving Jukedeck. So after a nice long break, I’m very pleased to share that I’ll be joining Danish music streaming startup Moodagent on the 17th of July, 2019. While the streaming service itself is new and hasn’t been launched yet, the company Moodagent A/S that owns it has been around for nearly two decades having built several products around their core technology for analysing musical content. You may have even come across their first music app on your Nokia phone back in the day! You can read all about them on Wikipedia, and find out more about the Moodagent streaming service on their website. I hear they’ll be launching it very soon!
I’ll be working in the Machine Learning team of the company as Senior Research Scientist on the design and development of their content organisation and music recommendation systems. I really look forward to the new beginning in Copenhagen and to learning a lot of new things from working on an area that’s still quite new to me. And also travelling around beautiful Europe!
I finally decided to get myself familiar with pandas while working on a recent side-project related to recommender systems. When I got started with it, I was still stubborn that I could achieve most things I needed to do in relation to data pre-processing with Python modules like tools like glob, json, numpy and scipy. True as that may be, I found myself spending way too much time writing routines to process the data itself and not getting anywhere close to working on the actual project. This was very reminiscent of the time a few years ago when I got immersed in writing code to manually compute gradients for various neural network architectures while getting nowhere in developing a music prediction model before finally deciding to make my life easier with theano! And so, this seemed like the perfect time to get started with learning pandas.
In the past I’ve found that, especially when it comes to learning useful features of new modules in Python, a hands-on and practical approach is much better than reviewing documentation and learning various features of a module without much of an application context, so I started looking around for such tutorial introductions to pandas. In the process I came across two invaluable resources that I thought I’d highlight here in this blog post. These really aren’t much, but gave me a surprisingly thorough (and quick) start to employ pandas in my own project.
Kaggle Learn has a bunch of very well-organised and basic introductory Micro-courses on various Data Science topics from Machine Learning, to Data IO and Visualisation. I get started with the Pandas Micro-course which proved to be the ideal starting point for someone like me that had never used the module previously. This can be followed up with some of the other micro-courses, such as the one on data visualisation or embeddings which help one understand various concepts better through application. In fact, it’s what I’m planning to do as well!
Pandas Exercises on GitHub
So the Pandas Micro-course was a great starting point, but still left me wanting more practice on the topic as I still didn’t feel totally fluent. It was then that I stumbled upon a fantastic compilation of Pandas exercises on GitHub by Guilherme Samora. So I cloned the repository, loaded these exercises up on Jupyter Notebook and got down to solving them one after another! This really did help with getting more fluent with the rich set of tools that Pandas has to offer.
By the time I was done with Guilherme’s exercises (only a couple of days after starting with the Kaggle micro-course), I felt ready to apply my newly acquired pandas skills to my own project, and to discover more about the module through it. There certainly were plenty more resources that a quick Google search returned, but none appealed as much to me at a first glance, as the two I finally went with.
I’m sure I have only scratched the surface when it comes to useful pandas learning resources, and I’m very curious to hear about those that others have found useful, and why, so that I can look them up as well! So do feel free to share them in the comments below.
As some of you might already know, I have been volunteering with a few of my peers in India to promote awareness about Music Technology in the country through the Music Tech Community – India initiative. Upon my suggestion, during the past months we had agreed upon and planned to begin a new blog post series that would contain interviews with individuals engaged with Music Technology in India, or elsewhere but who are from India. We hope that readers of this blog post series will have much to learn from the experiences of these individuals and that this will help them gain valuable insights into the field and inspire them to shape their own careers in the future.
I’m very pleased to announce today that we just published the first post in this series on the website! It is an interview with an active member of the community and a researcher applying Information Retrieval techniques to Indian classical music – Ajay Srinivasamurthy. During the weeks that preceded the publication of the post, we got in touch with Ajay who kindly offered to take part in this initiative. You can read what Ajay had to say during the interview in the blog post.
I believe this is a great start, and I look forward to more of such interesting chats in the future!
Now that I’m no longer working at Jukedeck, I happen to have plenty of free time on my hands! I’ve been spending this time travelling, catching up on my reading list, helping out with activities of the Music Tech Community – India and making music among other things. In an effort to satisfy a long-standing curiosity, I signed up for the Recommender Systems specialisation being offered on Coursera by University of Minnesota, and recently completed it. It comprised of four courses:
Introduction to Recommender Systems: Non-personalised and Content-based (certificate)
Recommender Systems: Evaluation and Metrics (certificate)
Matrix Factorisation and Advanced Techniques (certificate)
It took me about a month to complete all four courses at a fairly liesurely pace given how much time I had at my disposal while not working. This was a very well-taught specialisation with some of the best-designed Courses I’ve done on Coursera so far. It covered a wide range of topics that offered a comprehensive overview of a vast area of research. Solving the assignments by hand was a new, but very engaging experience that really allowed me to focus on what actually happens at a very basic level under-the-hood in such systems. It was all done by implementing the various formulae for content-based filtering, item-item collaborative filtering, user-user collaborative filtering (including matrix factorisation methods) in spreadsheets. There was an Honours Track in each course that focused on implementing the various types of recommender systems and related concepts that I decided not to pursue, as all the programming was in Java. I decided I would follow the courses up with my own implementation projects in Python as that’s something of greater interest to me. So now I’m looking for little projects to get me going.
I would definitely recommend this specialisation to anyone interested in Recommender Systems. It has left me with a very good understanding of the basics and a fair idea of the various directions in which I can pursue things in more detail. Not to mention, a tonne of references to read up on which I look forward to doing along with implementing some of the algorithms in the coming weeks.
This is just a quick post to let everyone know that I have decided to leave Jukedeck. It’s been a unique and fascinating journey the past three or so years with a flexible and forward-thinking company, and a stimulating work environment. I couldn’t have asked for a more apt transition into employment after my PhD than the one that led me to Jukedeck and I’m really grateful for all that I have learned here, the people I’ve had the opportunity to work with and everything the company has done for me during this period. This also means that I’m no longer going to be living or working in the UK, and my wife Nina and I have some new and exciting plans for the future that I’m really looking forward to.
There have also been some interesting developments in regards to where I’ll be going and what I’ll be doing next now that my tenure at Jukedeck has come to an end. I’ll post updates here on my blog as and when things take shape in the coming months.
A few months following the acceptance of our paper at ISMIR 2018, I attended the conference in Paris with several of my colleagues from Jukedeck. We had a fairly large presence there dwarfed (as far as I can tell) only by a larger one from Spotify. The conference was organised very well and everything went-off smoothly. It was great to be back in the beautiful city after my last visit nearly 8 years ago!
I was particularly pleased by the new format for presenting accepted papers at this ISMIR wherein each paper was given both oral and poster presentation slots thus removing the traditional distinction between papers that exists in conferences. In the case of our paper on StructureNet, I made the oral presentation and my colleagues and co-authors – Gabriele and Marco – made the poster presentation. Fortunately, this year ISMIR was streamed live and the videos were later stored on YouTube so I’m able to share the video of my presentation with you. It’s only a 4-minute presentation so do check it out! And it appeared to me each time I passed our poster by that it received a lot of attention, and this was of course great! I, with help from members of my team, also prepared a blog post on StructureNet which was published recently on Jukedeck R & D Team’s Medium page. I urge you to give it a read if you’re curious what the paper is all about. Here’s a picture of the Jukedeck team at ISMIR:
I also signed up to play in this year’s ISMIR jam session organised by Uri Nieto from Pandora! If I remember correctly, it’s something that started in 2014 and has been getting more popular by the year. As anticipated, the jam session was a success and a lot of fun, with music ranging from AI-composed folk tunes to Jazz, Blues, Rock and Heavy Metal. I played two songs with my fellow attendees – Blackest Eyes by Porcupine Tree and Plush by Stone Temple Pilots. My friend Juanjo shared a recording of the first song with me in which I played bass.
As always, ISMIR this year provided a great opportunity to make new acquaintances, and meet old friends and colleagues. As it turns out quite a few of my friends from the Music Informatics Research Group (MIRG) at City, University of London showed up this time and it was great to catch up with them.
And to top it all off, my master thesis supervisor Hendrik Purwins managed to make it to the conference on the last day giving me the opportunity to get this one selfie with Tillman (my PhD thesis supervisor) and him.
Edit (13-06-2018): ISMIR has officially announced the list of accepted papers, so I’m sharing the details of our accepted paper too!
Medeot, G., Cherla, S., Kosta, K., McVicar, M., Abdallah, S., Selvi, M., Newton-Rex, E., and Webster, K., StructureNet: Inducing Structure in Generated Melodies. In: Proc. International Society for Music Information Retrieval Conference (ISMIR 2018). Paris, France.
I recently ran into a situation where I had to initially train a neural network first on one dataset, save it and then load it up later to train it on a different dataset (or using a different training procedure). I implemented this in Tensorflow and thought I’d share a stripped down version of the script here as it could serve as an instructive example on the use of Tensorflow sessions. Note that this is not necessarily the best way of doing this, and it might indeed be simpler to load the original graph and train that graph itself by making its parameters trainable, or something else like that.
The script can be found here. In the first stage of this script (the pre-training stage) there is only a single graph which contains the randomly initialised and trained model. One might as well avoid explicitly defining a graph as Tensorflow’s default graph will be used for this purpose. This model (together with its parameters) is saved to a file and then loaded for the second re-training stage. In this second stage, there are two graphs. The first graph is loaded from the saved file and contains the pre-trained model whose parameters are the ones whose values we wish to assign to those of the second model before training the latter on a different dataset. The parameters of the second model are randomly initialised prior to this assignment step. In order for the assignment to work, I found it necessary to assign parameters across graphs and this could be done by saving the parameters of the first model as numpy tensors and assigning the values of these numpy tensors to the right parameters of the second model.