Oral Presentation at the 19th International Society for Music Information Retrieval Conference

- lmn

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:

The Jukedeck Team at ISMIR 2018 – (from left-to-right) Ben, Reinier, Gabriele, Matt, me, Katerina and Marco.

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.

The MIRG at ISMIR 2018: (from left-to-right, back-to-front) Shahar, me, Daniel, Tillman, Andreas, Radha and Reinier.

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.

Tillman, me and Hendrik at the conference venue.

Paper Accepted at ISMIR 2018

- lmn

A paper I submitted with my colleagues from Jukedeck was accepted at the 19th International Society for Music Information Retrieval Conference. A big congratulations to my co-authors Gabriele, Katerina, Matt, Samer, Marco, Ed and Kevin. It’s been a pleasure working with you all on this project and it’s a well-deserved recognition of the work itself!

More details to come soon!

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.

Tensorflow Tip: Pretrain and Retrain

- lmn

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.

Completed the Course “Machine Learning with Big Data” offered by UCSD on Coursera

- lmn

I successfully completed this course with a 98.9% mark. This course was relatively more focused than the others so far. The machine learning theory that was covered in it was very basic and good for beginners so I skimmed through it fairly quickly. Nevertheless, it was a good refresher of models such as Naive Bayes, Decision Trees and k-Means Clustering. What I found particularly useful was the introduction to the KNIME and Spark ML frameworks and the exercises where one had to apply these ML models to some example datasets.

I think this course and the last one were more hands-on and what I was looking for when I first started this module with a greater focus on ML in the context of Big Data.

And here’s the certificate that I was awarded on completing the course.

Completed the Course “Big Data Integration and Processing” offered by UCSD on Coursera

- lmn

I successfully completed this course with a 97.7% mark. This course was once again broad and touched upon some big data technologies through a series of lectures, assignments and hands-on exercises. The focus was mainly on querying JSON data using MongoDB, analysing data using Pandas, and programming in Spark (Spark SQL, Spark Streaming, Spark MLLIB and Spark GraphX). All these were things I was curious about and it was great that they introduced these in the course. There were also an exercise on analysing tweets using both MongoDB and Spark. They had one section on something called Splunk which I thought was a waste of time but I guess they have to keep their sponsors happy.

This specialisation so far (I’m halfway through) has been fairly introductory and lacking depth. It’s been good to the extent that I feel like I’m aware of all these different technologies and would be able to know where to start if I was to use them for some specific application. As I expected, this course was more hands-on which was great!

And here’s the certificate that I was awarded on completing the course.

New Guitar Video – Man in the Box by Alice in Chains!

- lmn

This is one of the first AiC songs I heard that got me into the band. While I had the broken finger, I also took some time off the fast licks and finger-intensive playing to learn some wah coordination and this was a great song to begin with. I used a Vox wah pedal here that I bought years ago. A bit squeaky but it worked alright. There’s still a few rough edges in this final recording and this is the result of the time I was willing to spend on perfecting it. Another Rush song coming up afte this!

The Tensorflow Datasets API for Sequence Data (Code Examples)

- lmn

This post was originally meant to be an entire tutorial (with a link to the GitHub repository) on how to use the Tensorflow Datasets API and how this contrasts with the placeholder approach for passing data into Tensorflow graphs that is generally more widely used. Unfortunately, I’m unable to set aside the time for writing about it in detail as I had originally intended, and thus I’m sharing the code with a few notes to help one make use of it.

First off, here is the link to the GitHub repository. It contains two main scripts – placeholder_vs_iterators.py and generator_vs_tfrecord.py. The first script implements three ways in which data can be passed into the Tensorflow graph. Note that in all cases this is sequence data. The first is the standard placeholder approach that most are familiar with. The second uses iterators and the third uses feedable iterators respectively to input data to the graph. The latter two are what I gathered to be the new methods to pass data into the graphs that the new Tensorflow Datasets API introduces. The script can be invoked with an integer command-line argument (1, 2 or 3) that chooses between the three approaches.

The second script is generator_vs_tfrecord.py. Having played around a bit with the two new data input approaches – iterators and feedable iterators, I decided to stick with the former in examining three different ways in which one can iterate through data while passing it into the graph during training. The first takes unbatched sequences via a generator function and applies certain standard preprocessing steps to it (zero-padding, batching, etc.) to it before using the data to train the model. The second approach begins with data that has been zero-padded and batched and passes that to the model via a generator function during training. The final approach first creates a Tensorflow Record file following the SequenceExample Protocol Buffer and reads sequences from this file, zero-pads and batches them before passing them to the graph during training. The third approach is what I would consider the most dependent on the Tensorflow Datasets API whereas the other two are to a greater extent reliant on Numpy. This script is also invoked with an integer command-line argument (1, 2 or 3) that chooses between the three approaches.

So there you have it! The rest of the code contains global constants for batch size, sequence length, etc. that can be changed as needed, basic training loops and a simple LSTM model to get the above examples working. I found, when I first got started with using this API around the time Tensorflow 1.4 was released, that there were few fully working examples that use it (and this is often what one is looking for when getting started with it) so I decided to share this code. One can realise more complex data pipelines using this handy API by building on these examples.

I refer the readers to the following useful links to understand more about the API, and Google Protocol Buffers while going through the code:

The Tensorflow Datasets API Blog Post

The Official Tensorflow Documentation

A blog post that helped me understand the SequenceExample Protocol Buffer format better.

A useful StackOverflow post on fixed and variable length features.

The definition of the Example and SequenceExample ProtoBufs

New Guitar Video – Jacob’s Ladder by Rush!

- lmn

I just finished recording and uploading the video of me playing this piece on YouTube, and I’m happy to share the link here! I’ve been learning to play it for a while now, and it was one of the pieces I could actually play as my broken little-finger was recovering after a minor bicycle accident. It’s so much fun to play with the 11/8 (or 11/16, whatever) time sections early on and the 13/8 (or 13/16, whatever) time sections towards the end of the piece! It’s fairly straightforward otherwise, but really requires some attention due to the repetitions that one tends to get carried away with. I love the solo composed by Alex Lifeson on this one, as I do most of his very clever guitar work! And as much as I could perhaps play it better, I’ll go easy on myself and settle with what I have in this video.

I’ve been learning to play another awesome song by Rush, that I’ll hopefully record and upload soon, stay tuned!

Completed Andrew Ng’s “Convolutional Neural Networks” course on Coursera

- lmn

I successfully completed this course with a 100.0% mark. Unlike the other two courses I had done as a part of this Deep Learning specialisation, there was much to learn for me in this one. I had only skimmed over a couple of papers on conv. nets in the past and hadn’t really implemented any aspects of this class of models except helping out colleagues in fixing bugs in their code. So I was stoked to do this course. And I was not disappointed. Andrew Ng designs and delivers his lectures very well and this course was no exception. The programming assignments and quizzes were engaging and moderately challenging. The idea of 1D, 2D and 3D convolutions was explained clearly and in sufficient depth in the lectures. They also covered some state-of-the-art convolutional architectures such as VGG Net, Inception Net, Network-in-Network and also applications such as Object and Face Recognition and Neural Style Transfer net, to all of which convolutional networks are a cornerstone. The reading list for the course was also very useful and interesting. All in all, a great resource in my opinion for someone interested in this topic! And as usual, here’s the certificate I received on completing this course.

Completed the Course “Big Data Modeling and Management Systems” offered by UCSD on Coursera

- lmn

I successfully completed this course with a 100.0% mark. It was quite broad and covered a range of topics somewhat superficially, from Relational Databases, their relation to Big Data Management Systems, the various alternatives that exist for processing different types of big data. As with the first course, there were a lot of new names to grasp and connections to be made between the things they represented. The assignments were straightforward and involved running a few specific command-line tools and spreadsheet commands to process data and carry out some basic analysis just to get a feel for data tables and how one might go about extracting information from them. The final assignment involved completing an incomplete relational database design for a game. In my opinion, its goals could have been more precise, its connection to the course material more clear, and being a peer-graded assignment the evaluation criteria more well-defined. Quite a few learners seem to have lost out due to someone else not being able to evaluate their assignment properly due to the latter shortcoming. And as usual, here’s the certificate that I was awarded on completing the course.

It looks like the upcoming courses in this specialisation contain more practical and hands-on exercises, so looking forward to that in the coming weeks!