Crate tsai 

tsai

Time series deep learning in Rust - a feature-parity port of Python tsai.

tsai-rs provides a comprehensive toolkit for time series analysis using deep learning:

• Data handling: Datasets, dataloaders, and preprocessing
• Transforms: Augmentations, label mixing, and imaging transforms
• Models: CNN, Transformer, ROCKET, RNN, and Tabular architectures
• Training: Learner, callbacks, metrics, and schedulers
• Analysis: Confusion matrix, top losses, importance
• Explainability: Attribution maps, activation capture

Quick Start

use tsai::prelude::*;

// Load data
let x = read_npy("data/X_train.npy")?;
let y = read_npy("data/y_train.npy")?;
let dataset = TSDataset::from_arrays(x, Some(y))?;

// Create dataloaders
let (train_ds, valid_ds) = train_test_split(&dataset, 0.2, Seed::new(42))?;
let dls = TSDataLoaders::builder(train_ds, valid_ds)
    .batch_size(64)
    .build()?;

// Create model
let config = InceptionTimePlusConfig::new(dls.n_vars(), dls.seq_len(), n_classes);
let model = config.init(&device);

// Train
let learner = Learner::new(model, dls, LearnerConfig::default(), &device);
learner.fit_one_cycle(25, 1e-3)?;

Feature Flags

• backend-ndarray (default): CPU backend using ndarray
• backend-wgpu: GPU backend using WGPU (Metal on macOS, Vulkan on Linux/Windows)
• backend-tch: PyTorch backend via tch-rs
• wandb: Weights & Biases integration

Re-exports

pub use tsai_analysis as analysis;

pub use tsai_core as core;

pub use tsai_data as data;

pub use tsai_explain as explain;

pub use tsai_models as models;

pub use tsai_train as train;

pub use tsai_transforms as transforms;

Modules

all

All module for importing everything.

compat

Compatibility module for sklearn-like API.

prelude

Prelude module for convenient imports.