Basic Usage¶

Quick Start Example¶

Here’s a simple example to get you started with DSIPTS:

from dsipts.data_structure.d1_layers import MultiSourceTSDataSet
from dsipts.data_structure.d2_layers import EncoderDecoder
from dsipts.models import ITransformer
import pytorch_lightning as pl

# 1. Load your time series data (D1 Layer)
d1_dataset = MultiSourceTSDataSet(
    file_paths=['data/weather.csv'],
    target_cols=['temperature'],
    time_col='timestamp',
    enrich_cat=['hour', 'dow']  # Add temporal features
)

# 2. Create encoder-decoder structure (D2 Layer)
d2_dataset = EncoderDecoder(
    d1_dataset=d1_dataset,
    past_len=96,      # 96 timesteps for input
    future_len=24,    # 24 timesteps to predict
    scaling_method='standard'
)

# 3. Setup data splits
d2_dataset.setup(stage='fit')

# 4. Initialize model
model = ITransformer(
    out_channels=1,
    past_steps=96,
    future_steps=24,
    d_model=512,
    n_heads=8
)

# 5. Train with PyTorch Lightning
trainer = pl.Trainer(max_epochs=10)
trainer.fit(model, d2_dataset)

Data Layer Architecture¶

DSIPTS uses a two-layer data architecture:

D1 Layer (Data Loading): Handles raw data loading, preprocessing, and temporal enrichment.
D2 Layer (Windowing & Batching): Creates sliding windows, handles train/val/test splits, and manages scaling.

This separation provides flexibility and prevents data leakage.

Model Training¶

All models in DSIPTS are PyTorch Lightning modules, which means you get:

Automatic GPU/TPU support
Distributed training
Checkpointing
Logging integration
And more!

See the Advanced Features section for more details.