Data is weird.

This week was more productive for me than the others! I was able to add visualization to two of the models I worked with and play around with tensorboard for a bit. That said, I have about 50 tabs open.

I also started training two BoW models with the STS dataset. The first, a Naive Bayes Classifier, had about 75% accuracy. The hard part was/ is training the set via my tensorflow/Bow/neural net model. Shaping and massaging the data has taken more time than expected. Also, because I’m using two different data files, ensuring that the data aren’t very different from one another by mixing the test sentences and weaving the data from both sets.

I feel a bit behind because certain things (data manipulation) are taking far longer than I intended or thought. I’m redoing my fast.Ai courses because I think a second pass over NLP/ Embeddings will help me understand what’s going on as I try to train embeddings for the data. My current approach needs reworking as I’m adapting someone else’s code. Rather, I’ll be creating my own vocabulary from the words and inducing my own embeddings.

Next week’s goal is to train the net and start incorporating elements of transformers on the data. I’ve wholly abandoned reading several papers a week – though I understand them, I need a more thorough hands on approach. Research is hard and while my mentor is practically begging me to bother him with questions, I’ve always been stuck in my own head and seeking to find answers myself. But then, why waste time? I’d rather look foolish and green but get the work done, than spend time hacking away by myself, frustrated.