n beats

nbeats
해석 가능한 시계열 예측을위한 신경 기반 확장 분석

Tensorflow/Pytorch 구현 | 종이 | 결과

NBEATS의 일반적이고 해석 가능한 레이어의 출력

설치

두 백엔드를 동시에 설치할 수 있습니다.

PYPI에서

Tensorflow/Keras 백엔드 설치 : pip install nbeats-keras

Pytorch 백엔드 설치 : pip install nbeats-pytorch

출처에서

설치는 makefile을 기반으로합니다.

Keras로 n-beats를 설치하도록 명령 : make install-keras

Pytorch로 N-Beat를 설치하도록 명령 : make install-pytorch

GPU에서 실행하십시오

이 트릭은 최근 버전의 텐서 플로에서 더 이상 필요하지 않습니다. GPU의 활용을 강제하기 위해 (Keras 백엔드와 함께) 실행 : pip uninstall -y tensorflow && pip install tensorflow-gpu .

예

다음은 두 가지 백엔드에 익숙해지는 예입니다. Keras 백엔드 만 현재 input_dim>1 지원합니다.

 import warnings

import numpy as np

from nbeats_keras . model import NBeatsNet as NBeatsKeras
from nbeats_pytorch . model import NBeatsNet as NBeatsPytorch

warnings . filterwarnings ( action = 'ignore' , message = 'Setting attributes' )


def main ():
    # https://keras.io/layers/recurrent/
    # At the moment only Keras supports input_dim > 1. In the original paper, input_dim=1.
    num_samples , time_steps , input_dim , output_dim = 50_000 , 10 , 1 , 1

    # This example is for both Keras and Pytorch. In practice, choose the one you prefer.
    for BackendType in [ NBeatsKeras , NBeatsPytorch ]:
        # NOTE: If you choose the Keras backend with input_dim>1, you have 
        # to set the value here too (in the constructor).
        backend = BackendType (
            backcast_length = time_steps , forecast_length = output_dim ,
            stack_types = ( NBeatsKeras . GENERIC_BLOCK , NBeatsKeras . GENERIC_BLOCK ),
            nb_blocks_per_stack = 2 , thetas_dim = ( 4 , 4 ), share_weights_in_stack = True ,
            hidden_layer_units = 64
        )

        # Definition of the objective function and the optimizer.
        backend . compile ( loss = 'mae' , optimizer = 'adam' )

        # Definition of the data. The problem to solve is to find f such as | f(x) - y | -> 0.
        # where f = np.mean.
        x = np . random . uniform ( size = ( num_samples , time_steps , input_dim ))
        y = np . mean ( x , axis = 1 , keepdims = True )

        # Split data into training and testing datasets.
        c = num_samples // 10
        x_train , y_train , x_test , y_test = x [ c :], y [ c :], x [: c ], y [: c ]
        test_size = len ( x_test )

        # Train the model.
        print ( 'Training...' )
        backend . fit ( x_train , y_train , validation_data = ( x_test , y_test ), epochs = 20 , batch_size = 128 )

        # Save the model for later.
        backend . save ( 'n_beats_model.h5' )

        # Predict on the testing set (forecast).
        predictions_forecast = backend . predict ( x_test )
        np . testing . assert_equal ( predictions_forecast . shape , ( test_size , backend . forecast_length , output_dim ))

        # Predict on the testing set (backcast).
        predictions_backcast = backend . predict ( x_test , return_backcast = True )
        np . testing . assert_equal ( predictions_backcast . shape , ( test_size , backend . backcast_length , output_dim ))

        # Load the model.
        model_2 = BackendType . load ( 'n_beats_model.h5' )

        np . testing . assert_almost_equal ( predictions_forecast , model_2 . predict ( x_test ))


if __name__ == '__main__' :
    main ()

더 많은 예제를 찾아보십시오. Jupyter 노트가 포함되어 있습니다.

Jupyter Notebook : nbeats.ipynb : make run-jupyter .

소환

 @misc{NBeatsPRemy,
  author = {Philippe Remy},
  title = {N-BEATS: Neural basis expansion analysis for interpretable time series forecasting},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {url{https://github.com/philipperemy/n-beats}},
}

기고자

감사합니다!