It was an amazing week in Mexico city at NAACL 2024! I am happy I met so many brilliant researchers and had so many meaningful conversations. I am grateful to be a part of such a community. It’s a shame I was unable to join all the talks and posters I wanted to attend (sadly cannot be in two places at the same time, I need a time-turner ha-ha). Also, special thanks to Amsterdam folks for keeping me company 😊 In this post I want to share my experience and paper highlights.
Disclaimer: this post is for my bachelor and master students at UvA, to give them introduction for PRACTICAL NMT from top to bottom. There is a gap between theory and practice, that is not always addressed by books.
Introduction Since you are an informatics or ai program’s student, let’s assume you know something about machine learning, or maybe even machine translation. You’ve heard about Transformer. You probably read The Illustrated Transformer blog post.
Memory usage is a common issue for large ML models. Especially in academia, we have to use resources wisely and make the most out of resources available. While working on my mixture model’s KL-objective, I have to make some less common optimization to reduce memory usage.
Setup Decoder outputs a large matrix \(O\) with dimensionality \((M \times B \times L \times D)\) where \(M\) is the number of clusters, \(B\) is a batch size, \(L\) is a sequence lengths and \(D\) is model output dimension.
Table of content Fairseq How To Easy Mode Not So Easy Mode Fairseq How To Before we start with extension, let’s try to understand how fairseq training works for seq2seq models. In this tutorial I will use only hydra-train module to make it possible load yaml configs.
Install fairseq Follow installation guide on github page.
Training with Hydra Suppose you want to train translation model using hydra training. The command syntax is the following: