galeos93 / disaster_tweets Goto Github PK

Now that our dataset is ready, we are able to train a model. For a baseline, nothing too elaborate is needed. Possible ideas:

• Vanilla LSTM
• CNN

Try to obtain same performance as https://www.kaggle.com/pinkaxe/pytorch-with-embeddingbag-layer, which is very similar to your approach. I don't obtain such high accuracy on validation set.

I think TweetDataset will not work directly if I feed it to DataLoader. This is so because of the different lenghts of the sentences. This fact makes their aggrupation on a batch impossible. I have to find out how to fix it.

For example, here, they use collate_fn=collate_batch to "flatten" the sentences into a giant unidimensional vector. I will see if there are alternatives to this.

We need to add tox so we can share a common environment easily.

We have to make an Exploratory Data Analysis on the data. As Andrej Karpathy wisely indicated in this post, we have to become one with the data. Understanding the data is the first thing that have to be done before starting coding.

Naturally, we need to feed our data to our model. Well, how we do that? Luckily PyTorch has a comprehensive guide on dataset and data_loaders we can follow.

Documentation

• https://pytorch.org/tutorials/beginner/basics/data_tutorial.html
• https://pytorch.org/tutorials/beginner/text_sentiment_ngrams_tutorial.html

I will need:

• A CNN model
• Add token to the vocabulary and give it an index
• A way to pad the sequences
• (...)