Awesome

Time sequence generation using PyTorch

This is an attempt to familiarize myself with PyTorch. In this example, the target to generate a sequence of continuous data (sine waves or mix of them) using LSTM

Updates

• 16/04/2017: When trying to generate a simple sine wave, the system flats out. It is unclear for me why this happens. The same happens with 2 and 3 sine-wave components.
• 18/04/2017: Thanks to the advice of Sean Robertson - https://discuss.pytorch.org/t/lstm-time-sequence-generation/1916/4 - to reduce the frequency of the sine-waves, I was finally able to generate signals. The 2 and 3 sine-wave components are working well (the 3 is a bit unstable).
• 19/04/2017: The method of teaching the model using only the ground truth is called Teacher Forcing.
• 20/04/2017: After further testing , I found my model works when the sine wave has a relatively high frequency (1/60 Hz or more). Lower frequency like (1/180 Hz) doesn't work.
  • With a sequence length of 100 timesteps, the model flats out when I use it for generation.
  • I tried to increase the sequence length till 500. The model no longer flats out, but the performance is poor. Probably the dependency is too long for the model to remember.
    • I need a way to be definite about this issue

TODO:

• Train the model on generation instead of prediction: training the model on its own output
  • Strangely, it doesn't lead to different results. With low frequencies, it doesn't work. With higher frequency, its performance is almost similar to the naive approach (where I train on the ground) truth.
• Try Bengio approach DAD Scheduled Sampling
  • Not optimistic though
• Re-package the experiment in order to be able to give it a set of configurations, and it will run them and store their results.