Codebase for "Time-series prediction" with RNN, GRU, LSTM and Attention

Authors: Jinsung Yoon Contact: [email protected]

This directory contains implementations of basic time-series prediction using RNN, GRU, LSTM or Attention methods. To run the pipeline, simply run python3 -m main_time_series_prediction.py.

Note: We recommend to do MinMax normalization on both input and output.

Stages of time-series prediction framework:

• Load dataset (Google stocks data)
• Train model: (1) RNN based: Simple RNN, GRU, LSTM (2) Attention based
• Evaluate the performance: MAE or MSE metrics

Command inputs:

• train_rate: training data ratio
• seq_len: sequence length
• task: classification or regression
• model_type: rnn, lstm, gru, or attention
• h_dim: hidden state dimensions
• n_layer: number of layers
• batch_size: the number of samples in each mini-batch
• epoch: the number of iterations
• learning_rate: learning rates
• metric_name: mse or mae

Example command

$ python3 main_time_series_prediction.py 
--train_rate 0.8 --seq_len 7 --task regression --model_type lstm
--h_dim 10 --n_layer 3 --batch_size 32 --epoch 100 --learning_rate 0.01
--metric_name mae

Outputs

• MAE or MSE performance of trained model