I’ve been diving a lot deeper into some interesting neural network papers recently and I’m looking to try and implement some of the models detailed in the papers. In general, I know that many papers include the code or I can just google the code to implement the model but I want to push myself to start implementing from scratch more.
Could anyone offer some tips on how they got started or gained the skills to be able to implement a model effectively within a few hours? Any advice would be much appreciated!
I’ve found https://paperswithcode.com/ and github search & topics useful tools. In my experience implementing a paper from scratch is a fantastic way to gain a deeper understand of a paper. Don’t be discouraged, I can’t remember a single paper I’ve tried to implement that provided all the details necessary to implement.
Out of curiosity - What do you do, when you are missing some key information?
I think this is why implementing is such a useful learning tool. Papers I tried to implement I probably read cover-to-cover 10 or 20 times, as opposed just skimming abstract, method, results.
When missing key info, after searching the paper a few times:
That was all the tricks I had, keen to know any more.
You can also contact the authors to ask them. Most of us are not monsters and will happily talk about our work. :)
What do you do if you find a reference implementation? Just run it? Try to implement from scratch?
I tried to copy as much as possible from the paper, then fill in the blanks with how I would have solved it. I only had a partial solution but I still learnt heaps more than just reading the paper.