torch-timeseries 0.1.2

Timeseries Learning Library for PyTorch.

pytorch_timeseries

An all in one deep learning library that boost your timeseries research. Check the documentation for more detail.

Compared to previous libraries, pytorch_timeseries is

• dataset automatically downloaded
• easy to use and extend
• clear documentation
• highly customizable
• install and run!
• ..........

1. installation

pip install torch-timeseries

⚠️⚠️⚠️ Warning: We only support python version >= 3.8+

2. Running Implemented Experiments

Forecast

# running DLinear Forecast on dataset ETTh1 with seed = 3 
pytexp --model DLinear --task Forecast --dataset_type ETTh1 run 3
# running DLinear Forecast on dataset ETTh1 with seeds=[1,2,3]
pytexp --model DLinear --task Forecast --dataset_type ETTh1 runs '[1,2,3]'

Imputation

# running DLinear Imputation on dataset ETTh1 with seed = 3 
pytexp --model DLinear --task Imputation --dataset_type ETTh1 run 3
# running DLinear Imputation on dataset ETTh1 with seed = [1,2,3] 
pytexp --model DLinear --task Imputation --dataset_type ETTh1 runs '[1,2,3]'

UEAClassification

# running DLinear UEAClassification on dataset EthanolConcentration with seed = 3 
pytexp --model DLinear --task UEAClassification --dataset_type EthanolConcentration run 3
# running DLinear UEAClassification on dataset EthanolConcentration with seed = [1,2,3] 
pytexp --model DLinear --task UEAClassification --dataset_type EthanolConcentration runs '[1,2,3]'

AnomalyDetection

# running DLinear AnomalyDetection on dataset MSL with seed = [1,2,3] 
pytexp --model DLinear --task AnomalyDetection --dataset_type MSL run 3
# running DLinear AnomalyDetection on dataset MSL with seed = [1,2,3] 
pytexp --model DLinear --task AnomalyDetection --dataset_type MSL runs 3

Development Milestones

Implemented Datasets

Full list of datasets can be found at Documentation.

Datasets	Forecasting	Imputation	Anomaly	Classification
ETTh1	✅	✅
ETTh2	✅	✅
ETTm1	✅	✅
ETTm2	✅	✅
......And More	✅	✅	✅	✅

Implemented Tasks

• Forecast
• Classfication (for UEA datasets)
• Anomaly Detection
• Imputation
• You can fill this check box! (contribute to develop your own task!)

Implemented Models

Models	Forecasting	Imputation	Anomaly	Classification
Informer (2021)	✅	✅	✅	✅
Autoformer (2021)	✅	✅	✅	✅
FEDformer (2022)	✅	✅	✅	✅
DLinear (2022)	✅	✅	✅	✅
PatchTST (2022)	✅	✅	✅	✅

Customizing Your Own Pipeline

we provide examples of :

• forecast
• imputation
• anomaly detection
• UEA classfication

Detail of customize forecasting pipeline is as follows:

1 Forecasting

1.1 download dataset

The dataset will be downloaded automatically!!!!

from torch_timeseries.dataset import ETTh1
from torch_timeseries.dataloader import StandardScaler, SlidingWindow, SlidingWindowTS
from torch_timeseries.model import DLinear
from torch.nn import MSELoss, L1Loss
from torch.optim import Adam
dataset = ETTh1('./data')

1.2 setup scaler/dataloader

Once you setup a dataloader and pass a scaler into this dataloader, the scaler will be fitted on the training set.

scaler = StandardScaler()
dataloader = SlidingWindowTS(dataset, 
                        window=96,
                        horizon=1,
                        steps=336,
                        batch_size=32, 
                        train_ratio=0.7, 
                        val_ratio=0.2, 
                        scaler=scaler,
                        )

After this, you can access the train/val/test loader by dataloader.train_loader/val_loader/test_loader

1.3 training

model = DLinear(dataloader.window, dataloader.steps, dataset.num_features, individual= True)
optimizer = Adam(model.parameters())
loss_function = MSELoss()

# train
model.train()
for scaled_x, scaled_y, x, y, x_date_enc, y_date_enc in dataloader.train_loader:
    optimizer.zero_grad()
    
    scaled_x = scaled_x.float()
    scaled_y = scaled_y.float()
    scaled_pred_y = model(scaled_x) 
    
    loss = loss_function(scaled_pred_y, scaled_y)
    loss.backward()
    optimizer.step()
    print(loss)

1.4 val/test

# val
model.eval()
for scaled_x, scaled_y, x, y, x_date_enc, y_date_enc in dataloader.val_loader:
    ....your validation code here...

# test
model.eval()
for scaled_x, scaled_y, x, y, x_date_enc, y_date_enc in dataloader.test_loader:
    ....your test code here...

Dev Install

install requirements

Note:This library assumes that you've installed Pytorch according to it's official website, the basic dependencies of torch > > related libraries may not be listed in the requirements files: https://pytorch.org/get-started/locally/

The recommended python version is 3.8.1+.

1. fork this project

2. clone this project (latest version)

git clone https://github.com/wayne155/pytorch_timeseries

3. install requirements.

pip install -r ./requirements.txt

4. change some code and push to the forked repo

5. create a pull request to this repo