Implement SAITS time series imputation model

[Copilot]

Implements a complete SAITS (Self-Attention-based Imputation for Time Series) project for multivariate time series data imputation using a dual-branch self-attention architecture.

Changes

• Model (saits/model.py): SAITS implementation with positional encoding, diagonal-masked self-attention, dual DMSA blocks, and combined ORT/MIT loss function
• Data utilities (saits/data.py): TimeSeriesDataset with artificial masking, normalization, sliding windows, train/test split
• Training (saits/train.py): SAITSTrainer with early stopping, LR scheduling, checkpointing
• Evaluation (saits/evaluate.py): MAE, MSE, RMSE, MRE metrics for imputation quality
• Tests: 36 tests covering model, data, and evaluation modules

Usage

from saits import SAITS, TimeSeriesDataset, SAITSTrainer

# Create model
model = SAITS(n_features=10, n_layers=2, d_model=128, n_heads=4)

# Impute missing values (mask: 1=observed, 0=missing)
x = torch.randn(32, 100, 10)
mask = torch.ones_like(x)
mask[:, 20:30, :5] = 0
imputed = model.impute(x, mask)

Original prompt

SAITS data imputation project

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.