ts-rnn 0.2.12

Package to forecast time series with recurrent neural network

Time_Series_Prediction_RNN

Requirements

ts_rnn requires the following to run:

• Python 3.7.3+

Installation

From pip

You could install the latest version from PyPi:

pip install ts-rnn

From Github

You could install the latest version directly from Github:

pip install https://github.com/LevPerla/Time_Series_Prediction_RNN/archive/master.zip

From source

Download the source code by cloning the repository or by pressing 'Download ZIP' on this page.

Install by navigating to the proper directory and running:

python setup.py install

Example

Example

Documentation

The full documentation haven't ready yet. I hope, it will show later.

Getting started

To import TS_RNN model run

from ts_rnn.model import TS_RNN

First of all, we need to set architecture of RNN in config in the way like this:

rnn_arch = {"layers": [
                        ["LSTM", {"units": 64,
                                  "return_sequences": False,
                                  "kernel_initializer": "glorot_uniform",
                                  "activation": "linear"}],
                        ["Dropout", {"rate": 0.2}],
                        ["Dense", {"activation": "linear"}]
                    ]}

WARNING: Last RNN block need to gave return_sequences: False, another - True

To set the architecture of RNN you can use some of this blocks:

# LSTM block
["LSTM", {Keras layer params}],
["GRU", {Keras layer params}],
["SimpleRNN", {Keras layer params}],
["Bidirectional", {Keras layer params}],
["Dropout", {Keras layer params}],
["Dense", {Keras layer params}]

TS_RNN class has 7 attributes:

1. n_lags - length of the input vector;
2. horizon - length of prediction horizon;
3. rnn_arch - description of the model's parameters in Python dictionary format;
4. strategy - prediction strategy: "Direct", "Recursive", "MiMo", "DirRec", "DirMo"
5. tuner - tupe of Keras.tuner: "RandomSearch", "BayesianOptimization", "Hyperband"
6. tuner_hp - keras_tuner.HyperParameters class
7. n_step_out - length of the output vector (Need to define only for DirMo strategy);
8. loss - Keras loss to train model;
9. optimizer - Keras optimizer to train model.
10. n_features - number of time series in the input (only for factors forecasting);
11. save_dir - dir to save logs

You can set model this way:

model = TS_RNN(rnn_arch=rnn_arch,  # dict with model architecture
               n_lags=12,  # length of the input vector
               horizon=TEST_LEN,  # length of prediction horizon
               strategy="MiMo",  # Prediction strategy from "Direct", "Recursive", "MiMo", "DirRec", "DirMo"
               loss="mae",  # Keras loss
               optimizer="adam",  # Keras optimizer
               n_features=X_train.shape[1]  # also you need to define this if use factors
               )

TS_RNN supports 5 methods:

1. fit - train the neural network;
2. predict - predict by the neural network by input;
3. forecast - predict by the neural network by last train values;
4. summary - print NNs architecture
5. save - save model files to dict

FIT

my_callbacks = [callbacks.EarlyStopping(patience=30, monitor='val_loss')]

model.fit(factors_train=factors_val,  # pd.DataFrame with factors time series
          target_train=target_val,  # pd.DataFrame or pd.Series with target time series
          factors_val=factors_val,  # pd.DataFrame with factors time series
          target_val=target_val,  # pd.DataFrame or pd.Series with target time series
          epochs=100,  # num epoch to train
          batch_size=12,  # batch_size
          callbacks=my_callbacks,  # Keras callbacks
          save_dir="../your_folder",  # folder to image save 
          verbose=2)  # verbose

PREDICT

predicted = model.predict(factors=factors_to_pred,
                          target=target_to_pred,
                          prediction_len=len(y_test))

FORECAST

predicted = model.forecast(prediction_len=HORIZON)

SUMMARY

model.summary()

SAVE

model.save(model_name='tsrnn_model', save_dir='path')

Also you may load TS_RNN model from folder

from ts_rnn.model import load_ts_rnn

model = load_ts_rnn(os.path.join('path', 'tsrnn_model'))

Simple example of usage:

Info: For better performance use MinMaxScaler and Deseasonalizer before fitting

from sklearn.model_selection import train_test_split
from ts_rnn.model import TS_RNN
import pandas as pd

HORIZON = 12

data_url = "https://raw.githubusercontent.com/LevPerla/Time_Series_Prediction_RNN/master/data/series_g.csv"
target = pd.read_csv(data_url, sep=";").series_g
target_train, target_test = train_test_split(target, test_size=HORIZON, shuffle=False)

model = TS_RNN(n_lags=12, horizon=HORIZON)
model.fit(target_train=target_train,
          target_val=target_test,
          epochs=40,
          batch_size=12,
          verbose=1)

model.summary()
predicted = model.predict(target=target_train[-model.n_lags:], prediction_len=HORIZON)