/PyTorch-RNN-Tutorial

PyTorch tutorial for using RNN and Encoder-Decoder RNN for time series forecasting

Primary LanguageJupyter NotebookApache License 2.0Apache-2.0

PyTorch tutorial on using RNNs and Encoder-Decoder RNNs for time series forcasting and hyperparameter tuning

Some blabber

This package resulted from my effort to write a simple PyTorch based ML package that uses recurrent neural networks (RNN) to predict a given time series data.

You must be wondering why you should bother with this package since there is a lot of stuff on the internet on this topic. Well, let me tell you that I have traveled the internet lanes and I was really frustrated by how scattered the information is in this context. It was a lot of effort to collect all the relevant parts from the internet and construct this package.

I had only a basic background in ML and zero knowledge of PyTorch (using Keras doesn't prepare you for PyTorch 😛) when I started writing this package. But that actually ended up being a blessing in disguise. Since I was starting from scratch, I was able to write the code in a way that was intuitive and easy to understand for people who are new to the subject.

So if you're feeling lost and frustrated, give this package a try. It might just help you understand not only RNNs, but PyTorch as well. And who knows, you might even have a little fun along the way.

Code Functionalities

  1. Many-to-One prediction using PyTorch's vanilla versions of RNN, LSTM, and GRU.
  2. Many-to-Many (or Seq2Seq) prediction using Encoder-Decoder architecture; base units could be RNN, LSTM, or GRU.
  3. Hyperparameter Tuning! It uses the Optuna library for that.
  4. Save PyTorch models, as well as reload and train them further.
  5. Works on any univariate and equispaced time series data.
  6. Can use GPUs.

Usage

Best way to figure out how to use this package is to check out the example notebooks available in the Notebooks folder.

I have also made a sample notebook available in Google Colab! Open In Colab

Code Structure

I have structured the code so that different operations are abstracted away in Python Classes. Here is a brief summary:

  • Model: Directory  -  contains classes which define the RNN models. RNN_Vanilla.py defines the Many-to-One RNN; the traditional kind. The EncDec.py file defines the Encode-Decoder class which uses the traditional RNN units as Encoder and Decoder modules, which are then combined together to provide a one-shot Many-to-Many prediction.

  • Notebooks: Directory - example notebooks which demonstrate how to use the code on a sample time series data consisting of multi frequency sin waves. It also contains a notebook which demonstrates how to perform hyperparameter tuning using Optuna.

  • Saved_models: Directory, empty - used to store the output from the Create_and_Train.py file.

  • Utils Directory - contains all the class files which do the data prep, training, testing, validation, and predicting.

    • Trainer.py contains the training loop, a test function to run the model on test data, as well as functions to make predictions.
    • SeqData.py file is used to create sequenced dataset, in torch tensors format, given a numpy 1D time series.
    • Create_and_Train.py is THE main file which creates a model (using the classes in the Model directory), runs the epoch loop, saves PyTorch models and train-test loss curves.

  • imports.py file is used by the notebooks present in the Notebooks folder.

  • requirements.txt file can be used in conjunction with pip to install the required packages.

Limitations

I haven't generalized the code to use multivariate time series data for the sake of simplicty. But, it is relatively easy to do. If interested, report in the repo's Issues section and we can collaborate!

Note

I also recommend checking out my colleague's implementation of rnn in pytorch.