Summary
Although we learned about the basic blocks of DL in the previous chapter, we put all of that into action while we used those blocks in common modeling patterns using PyTorch.
We saw how standard sequence models such as RNN, LSTM, and GRU can be used for time series prediction, and then we moved on to another paradigm of models, called Seq2Seq models. Here, we talked about how we can mix and match encoders and decoders to get the model we want. Encoders and decoders can be arbitrarily complex. Although we looked at simple encoders and decoders, it is very much possible to have something like a combination of a convolution block and an LSTM block working together for the encoder. Last but not least, we talked about teacher forcing and how it can help models train and converge faster and also with some performance boost.
In the next chapter, we will be tackling a subject that has captured a lot of attention (pun intended) in the past few years: attention and transformers...
