Anjishnu
Anjishnu
CS PhD student. Working on Fairness and Interpretability methods in NLP.

Week Zero

Week Zero

Let’s light this candle.

Cyclone Amphan made landfall in my city on the 21st of May. Even though we are lucky enough to be far away from the coast and thus only see about half of the maximum strength of the cyclone, it was still the worst we have seen in the last 20 years or so. And that is a major statement, given that we see a couple of cyclones every year during this period.

I still don’t have a very stable internet connection and that has made working on the project slightly more difficult for me. But I figured out one way I could stay productive was to download the armadillo and PyTorch docs, so that I can go through stuff offline while trying to figure out the flow of the source code. The following is the command I used for this purpose.

1
2
wget --mirror --convert-links --adjust-extension --page-requisites
     --no-parent http://arma.sourceforge.net/docs.html

Now coming to the project I will be doing across the next 3 months or so, it is titled Improvisation and Implementation of ANN Modules. The idea is that the ANN module of mlpack which is quite good at its current state is lacking in some features that would allow more user flexibility, more widespread adoption and hopefully also more usability. I will be adding some new features to the module and improving some of the existing features.

One of the first things that I am looking to implement is a Unpooling layer. We already have a maxpooling layer in mlpack, which has to implement unpooling as part of the backpropagation through the layer.

Unpooling as a standalone layer would do the exact opposite of maxpooling. So, basically the forward function of Unpool would resemble the backward for maxpool. My main task in this week has been to understand the forward and backward functions of maxpool and figure out how to invert them correctly to work in the separate layer. Print statements have been very useful in this regard :)

Apart from implementing the unpooling layer, I figured I would need to add some accessor methods to the existing maxpool layer to get the pooling indices, because from the pseudo-code I have curently for the Unpool layer, I would need to pass in indices as a input parameter. This design would also follow the format of the equivalent PyTorch implementation of the layer and is thus expected to be user-friendly. I will add these accessors in a PR very soon, once I have the Unpool layer completely figured out and working locally.

Also, understanding the maxpooling layer gave me some insight for the next feature that I will be implementing after UnPooling layer, which is the Lp Pooling layer. I grouped these 2 together, because it felt like it would save time as they are related topics.

Now, coming to the fun part of this blog. I mentioned during the introductory meet on IRC, that I will be suggesting some songs throughout the summer and then at the end, I will put all of them in a Spotify playlist for easier access. Here’s the first song in the series.

Man Up by Hailee Steinfeld

Another interesting thing happened this week. An issue I created in torchdata (which is a useful helper library for writing data loaders in PyTorch) got the author of the library to finally shrug off some dust and release a new minor version which removes unnecessary dependencies on Python versions and thereby allows the library to be used in Google Colab easily.

Also, watching a spaceship launch live, is probably one of the most breath-taking moments ever. I am so glad they made it up safe!

Until next time!
XOXO