Tensorflow 2 / Torch DA-RNN

A Tensorflow 2 (Keras) and pytorch implementation of the Dual-Stage Attention-Based Recurrent Neural Network for Time Series Prediction

Paper: https://arxiv.org/abs/1704.02971

Run notebook demo

Install dependencies (It is recommended to use anaconda to manage environments):

make install

Run notebook:

cd notebook
jupyter lab

# Run `pytorch.ipynb`

Install

For Tensorflow 2

pip install da-rnn[keras]

For PyTorch

pip install da-rnn[torch]

Usage

For Tensorflow 2 (Still buggy for now)

from da_rnn.keras import DARNN

model = DARNN(T=10, m=128)

# Train
model.fit(
    train_ds,
    validation_data=val_ds,
    epochs=100,
    verbose=1
)

# Predict
y_hat = model(inputs)

For PyTorch (Tested. Works)

import torch
from poutyne import Model
from da_rnn.torch import DARNN

darnn = DARNN(n=50, T=10, m=128)
model = Model(darnn)

# Train
model.fit(
    train_ds,
    validation_data=val_ds,
    epochs=100,
    verbose=1
)

# Predict
with torch.no_grad():
    y_hat = model(inputs)

Python Docstring Notations

In docstrings of the methods of this project, we have the following notation convention:

variable_{subscript}__{superscript}

For example:

• y_T__i means , the i-th prediction value at time T.
• alpha_t__k means , the attention weight measuring the importance of the k-th input feature (driving series) at time t.

DARNN(T, m, p, y_dim=1)

DARNN(n, T, m, p, y_dim=1)

The naming of the following (hyper)parameters is consistent with the paper, except y_dim which is not mentioned in the paper.

• n (torch only) int input size, the number of features of a single driving series
• T int the length (time steps) of the window
• m int the number of the encoder hidden states
• p int the number of the decoder hidden states
• y_dim int=1 the prediction dimension. Defaults to 1.

Return the DA-RNN model instance.

Data Processing

Each feature item of the dataset should be of shape (batch_size, T, length_of_driving_series + y_dim)

And each label item of the dataset should be of shape (batch_size, y_dim)

TODO

• no hardcoding (1 for now) for prediction dimentionality

License

MIT