Portfolio

Image Matting with fastai & PyTorch

---

An online challenge proposed to apply image matting to the CelebA dataset. I tried my hand at it by re-implementing a paper by Adobe.

Context
Yearly, I work my way through the fast.ai online MOOC, which covers applied Deep Learning topics using their pyTorch-based library, fastai. Having finished the 2019 edition, I wanted to sink my teeth into something that was more than a toy problem. An online challenge proposed to apply image matting to the CelebA dataset, and that seemed like an ideal small-scale project.

My experience
While reading up on the challenge of applying image matting, I came across a 2017 paper published by Adobe, titled Deep Image Matting. Their approach to image matting was based on a U-net with a simple convolutional network for refining the edges. I knew that fastai provided some form of support for U-nets, so I re-read the paper a couple of times and set to work.

Over the next few weeks, I prepared a suitable dataset, which I also uploaded on Kaggle. In parallel, I spent most of my time re-implementing the Adobe paper by modifying the fastai-provided U-net, which often required me to drop back to the underlying pyTorch code.

My final results were decent enough to warrant a pretty long blogpost and got me accepted into a Fellowship program, though I did not end up taking that opportunity. It was a great way to get a gentle introduction to pyTorch too!

End-to-end Testing for AngularJS

---

Front-end testing was done manually when I joined my first company. I wanted to make it more robust and save us time.

Context
At my first gig out of university, I was a Key Accounts Manager for an enterprise SaaS startup, 1plusX. When I joined, new releases of the User Interface often were paired with an email to the company asking everyone to do some testing. While this was a good first attempt at Quality Assurance, I felt we would benefit from more consistent testing.

My experience
As our interface was written using Google’s AngularJS, I figured the best framework for end-to-end testing would be Protractor, which is Angular’s dedicated testing framework. This was my first foray into writing tests of any kind, but the concept felt fairly natural. There are quite a few blogposts explaining how to get started with Protractor, and their documentation is pretty great. I ended up writing approximately 40 tests, saving the company quite some time on repetitive tasks. I also wrote a Medium post about my experiences with Protractor.

Jekyll, Docker and rsync

---

I wanted to create a hub for my online presence that was flexible and easy to deploy.

Context
Looking for a new position, especially in the tech industry, often requires a bit of showing what you get up to when left alone with a computer. I needed a way to showcase that, as well as a hub for my online presence.

My experience
I’ve got a preference for simplicity, especially in a case like this. The showcase is most likely to be static, is not terribly performance-driven, and should just work. This is the ideal use-case for a static site generator, with one of the main ones being Jekyll. I also wanted to keep the deployment of the site as simple as possible. I ended up grabbin a docker image of jekyll so I could easily build the site without needing the entire ruby environment. This is as simple as docker run -it --rm -v $PWD:/srv/jekyll jekyll/jekyll:latest jekyll build. While I initially considered deploying using git hooks, I soon felt this was violating my “keep it stupid simple” principle. I opted instead to go with rsync to deploy the site to my personal VPS. As has become habit by now, I of course wrote a blog post documenting the story.

Django-based Annotation Web-app

---

I was experimenting a lot with Natural Entity Recognition and wanted a simple tool for helping in annotation.

Context
My first foray into machine learning-related topics was the topic of Natural Language Processing, specifically Named Entity Recognition (NER). While toying around with the various frameworks, mostly RASA NLU, I realised that having a nice dataset was an essential step in getting started.
Inspired by prodigy but disappointed at the pricing model, I decided to develop my own, extremely basic version.

My experience
Classifr is a dead-simple web-app that takes in a .csv of predictions from an algorithm, and displays those back to the user in a simple interface to confirm or reject the prediction. More information can be found here. A simple demo is included.
I have used Django plenty of times before in many different side projects, but this was the first, simple enough project with decent code quality where I felt confident in publishing it. It was also the first Django projects where I decided to write unit tests. I of course relied on Twitter’s Bootstrap 4 for tastefully sane defaults on the front-end.

Jupyter, fast.ai & Kaggle

---

I'm convinced that 'you master a concept if you can explain it'. I was struggling with multi-class image classification, so I created an explanatory notebook that steps through it.

Context
I’ve been following the fast.ai MOOC with absolute delight. I had already worked through Andrew Ng’s excellent Machine Learning coursera course, as well as the 2017 version of fast.ai, but the revised 2018 material with the new library really got me into the topic.
The second and third lessons of the Practical Deep Learning for Coders class cover multi-class image classification in the context of the Kaggle Planet competition. While the notebook provided in the lecture is good, it is a bit of a scratchpad, which makes it difficult to bring all concepts together into a coherent whole.

My experience
To challenge and prove to myself that I could master the concept, I decided to write a Jupyter notebook that I wished I would have found when I started the competition. Jupyter is a delightful way to write up knowledge, especially if you kick things up a gear with the excellent nbextensions package. I uploaded the result to my GitHub here.

Python scripting

---

If a default bash utility can't give me what I need, Python is my instant go-to scripting language of choice.

Context
It is not unusual for me to find myself in an environment where I need to automate a repetitive task. Sometimes it’s moving files around, sometimes it’s about quickly digging into a web API. Python, especially with its huge library support and easy to grasp syntax, is a delight to use. It is perfectly suited for my prototyping mindset: quickly get something working.
I am not particularily engaged in the war between Python 2 or 3. I use Python3 by default, as it is the most recent and has nicer support for encodings, but I don’t mind using 2 either, though it is always a bit of a mental shift (print statements for example).

My experience
It is hard to summarise all the one-off Python scripts I have written for various applications. I can give an example of the most recent one though. I had uploaded all of my pictures that I had scraped off Facebook with a Chrome extension to my NAS, but the dates were incorrect. Thankfully, I had written the scraping in such a way that the name of the image file reflected the date it was uploaded to Facebook. Now all I had to do was write the date to the Exif data of the picture, and the NAS would pick it up.
To that end, I wrote a script using the piexif and PIL packages for Python, which parsed the filenames, and if necessary, injected them into the image’s Exif data. After that I simply mounted the NAS on my laptop and ran the script. I doubt the approach would’ve been scalable if I had thousands of pictures (many network requests, would probably have been faster to run it on the NAS itself), but this perfectly suited my needs. I wrote about the script on my Medium blog.

Internal RFI Database

---

Answering repetitive Requests for Information ate up valuable time of our engineers. I centralised questionnaires to help Sales get more work done autonomously.

Context
The first company I joined was a B2B company, which meant answering RFIs (essentially business questionnaires) was a frequent occurrence. Due to the highly specific nature of the questions (concerning security practices, data processing timelines etc), Sales often had to involve us or responsible engineers to get the right answers.

My experience
As the nature of these questionnaires is often repetitive, there was always a high chance that a similar question had already been answered, but perhaps for a different prospect in a different sales region. However, the questionnaires were scattered across our file sharing systems with no reliable way to search them and retrieve the answers.

I wanted to centralise all these questionnaires while providing a search experience that was better than the usual Ctrl/Cmd+f in a really large spreadsheet. I had previously looked at Algolia (if you browse HackerNews it’s hard not to know them) and I knew this could be a way.

I wanted to keep the setup and maintenance simple, so I grabbed the default Jekyll template and mixed in the great jekyll-algolia and InstantSearch.js by following their relatively good documentation on it. Within a day I had launched a search engine that used could be filtered based on language, when it was answered and what company it was for!