BTSTN (version 1.0)

❖ Introduction

BTSTN is a Python package designed to model the metabolic dynamics from Partially-Observed Time Series based on the Bidrectional Time-series State Transfer Network.

❖ Framework

❖ System Requirements

The source code developed in Python 3.8 using PyTorch. The required python dependencies are given below.

torch==2.0.0
numpy>=1.23.5
pandas>=1.5.2
pypots==0.5

❖ Data Format

Batch	Time	Parameter 1	Parameter 2	...
A1	T1	2.1	5.0	...
A1	T2	1.4	4.3	...
A1	T3	3.8	0	...
B1	T1	2.1	5.0	...
B1	T3	1.4	4.3	...
B1	T4	3.8	0	...
...	...	...	...	...

❖ Read Your Data

DataReader are used to read your file as input for model training and forecasting.

import pandas as pd
from btstn import DataReader

# Load data
file_path = "xxx.csv"
data = DataReader(
    data=pd.read_csv(file_path),
    scaler="MinMax",
    unit = 1,
    dropna=False,
    sort=True,
    header=0
)

❖ Define Your Model

BTSTN provides a simple tool to define the structure of model and training parameters.

from btstn import BTSTN

# Initialize the model
tcnits = BTSTN(
    n_features=2,
    n_dims=16,
    g_inner=32,
    g_layers=2,
    d_inner=None,
    d_layers=None,
    dropout=0.1,
    d_dropout=0,
    activation={"fc_name":"tanh"},
    max_gap=10,
    batch_size=128,
    epoch=1000,
    patience=100,
    learning_rate=0.001,
    threshold=0,
    gpu_id=-1,
    num_workers=0,
    pin_memory=False,
    saving_path=None,
    saving_prefix=None
)

❖ Train Your Model

# fit the model
train_loss = tcnits.fit(data=data)

❖ Application

BTSTN supported the following tasks with the trained model: imputation and forecasting.

1. Imputation

BTSTN supported the Imputation Task (Interpolation and Extrapolation).

# Imputation
impdata = tcnits.impute(
    data=data,
    max_gap_f=None,
    batch_size_f=None,
    epoch_f=None,
    patience_f=None,
    learning_rate_f=None,
    inverse=True
)

2. Forecasting

BTSTN supported the Forecasting Task.

# Load data
mfile_path = "yyy.csv"
mdata = DataReader(
    data=pd.read_csv(mfile_path),
    scaler="MinMax",
    unit = 1,
    dropna=False,
    sort=True,
    header=0
)

# Monitor
tcnits.monitor(
    data=mdata,
    max_gap_f=10,
    batch_size_f=128,
    epoch_f=3000,
    patience_f=-1,
    learning_rate_f=0.001
)

# Forecast
foredata = tcnits.forecast(
    data=mdata,
    stime=[0],
    pred_step=99,
    multi_step=1,
    inverse=True
)