Successful model development can be surprisingly dependent on good engineering practices. Despite this, many model implementations scattered about Github are difficult to follow and hard to recreate locally. But what should a good model look like? I would propose that the gold standard for a model implemented on Github could be: The dependencies may be installed automatically, using a single command. I can build the model in a sandbox without polluting with my dev.

Keep It Simple Stupid: CNN Backbones Many modern CNNs (convolutional neural networks) use a backbone to extract feature maps as the first step in their architecture: for example, some flavour of Resnet50. Helpfully, many of these backbones are implemented for us in torchvision, even with pre-trained models made available. This means we can often write simple code that looks like the following: 1 2 3 4 5 6 7 8 9 10 11 12 from torchvision.

Back to post 1 Starting the the floor The first post in this series was rather destructive. The floor was torn out, and the deck trim was prised off. This post is more constructive, and covers the first few steps of building a new floor. Fortunately, it was possible to save enough of the original floor to use as a template. Working in this way was much simpler than trying to measure out the awkwardly curved inner profile of the hull!

First Steps In order to keep myself busy whilst locked down, I took on a project of restoring a rather dilapidated, though (in my opinion) charming, and pleasingly-proportioned keelboat. The boat was complete with both a mainsail and a jib and, despite being quite scruffy, actually was sailable. Despite this, the general condition of the boat was quite bad, and it was unlikely to deal well with strong gusts… The boat is a Hanseatic 19.

How did things get this bad?The Axioms of linking Every few months I end up in the a shady corner of the internet, with twenty tabs open searching for variations on ‘why the $%^&!@ does this not work’. Hours later I end up re-discovering the truth: using binary libraries in $CURRENT_YEAR still pretty much sucks. So let’s try and get something positive out of this, and at least write up the facts of the situation.

I have started trying to avoid the word “own” in the workplace. Admittedly, it’s a bit of a silly idea, but I do think it’s a worthwhile adjustment that can help improve our professional relationships. This little word can pop up quite a lot, “Jane owns this project”, “I own this document”, “Pete can you own this?”. It feels innocent enough, but I would argue that it can nudge our mindset in the wrong direction in the workplace.

When it comes to reviewing code that does any kind of file IO, there are a number of patterns that I find myself discouraging quite often. Interestingly, I have found that these patterns are rarely considered in discussions around code review and clean code. In this post, I will explore some of these patterns in more depth, and discuss why I believe it’s worth taking the time to do things slightly differently in many cases.

Abisso Chimera 2011 (trip report) Written as a trip report for the UBSS newsletter – Text in italics is my (attempt!) at translating Florence Ménétrier’s trip report into English. Participants: Fred, Ross, Flo for the Sophitaupes, the Faverjons (Marc, Anne et Maud) and five other Italians. This trip to Abisso Chimera (at 1006m, it is currently the 85th deepest cave in the world) was completed in October 2011. I found myself in the sad position of having finished my degree last year, and not feeling particularly keen on entering the world of work, I somehow managed to install myself in a Ph.

A few weeks ago, I posted to say that I had been sent a travel espresso machine by Handpresso and that I had the intention of taking it on expedition. Well I have now returned and despite several bouts of torrential rain, a path to base camp that was muddy enough to absorb an entire man and a herd of voracious cows that ate their way through an array of hiking clothes, I am pleased to report that the coffee available was second to none.