matthieuvion / lstm-gpu-jupyter

LSTM (PyTorch) , time series prediction: a benchmark case w/ and w/o stationarity + scaling

Benchmark - time series prediction LSTM PyTorch

• Noteboook updated from this source article, with a famous dataset
• What was added :
  • clarifies steps / order making data stationary / normalization (a common mistake/lack of in a lot of online ressources)
  • Make data stationary (first order diff). Not mandatory with a lstm but very common, recommended and usually improves performance
  • More structure / pipelined : transform functions (normalize / inverse diff via cumsum etc.), steps / notes on arrays shape etc.
  • Normalization (minmax)
  • Input size (lookback) directly set to 4 (vs. 1, then 4)
  • benchmark/viz models accuracy with and without statio. + scaling

Run:
Personally, running it in a dockerized, gpu/cuda-enabled pytorch/notebook environment (cf. github/gpu-jupyter)