# Classification Time Series Library > Use THUML Time-Series-Library/TSLib for deep learning time-series classification after ts-classification-data-prep, including UEA .ts TRAIN/TEST files, UEAloader, variable-length padding masks, multivariate channels, task_name=classification, CrossEntropyLoss training, softmax probabilities, accuracy evaluation, official classification scripts, supported model names such as TimesNet, Transformer, Autoformer, PatchTST, iTransformer, MambaSingleLayer/MambaSL, TimeMixer, and strict anti-leakage preprocessing. - Skill: `hibra999/classification-time-series-library` (Agent Skill, multi-file: 6 files) - Install (CLI): `npx skillmds@latest add hibra999/classification-time-series-library` - Raw SKILL.md: https://api.skillmd.com/api/skills/hibra999/classification-time-series-library/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: Hibra999 (https://skillmd.com/u/hibra999) - Updated: 2026-09-21 - Page: https://skillmd.com/skills/hibra999/classification-time-series-library --- # Time-Series-Library Classification Use this skill after `ts-classification-data-prep`. THUML Time-Series-Library (TSLib) is best when the user wants to run deep learning classification experiments with the repository's unified `run.py`, UEA `.ts` datasets, PyTorch models, official bash recipes, checkpoints, and reproducible benchmark-style runs. Do not present TSLib as a general sklearn-style classifier library. It is a research codebase: training is driven by CLI scripts, native output is accuracy, and many conveniences such as sklearn `fit/predict` objects, built-in stratified CV, and calibration reports are not documented. ## Minimum Install ```bash git clone https://github.com/thuml/Time-Series-Library.git cd Time-Series-Library conda create -n tslib python=3.11 conda activate tslib pip install torch==2.5.1 --index-url https://download.pytorch.org/whl/cu121 pip install -r requirements.txt ``` Install optional model dependencies only when needed, especially `mamba_ssm` for Mamba/MambaSL on Linux/CUDA. ## Data Contract - Require a completed `ts-classification-data-prep` contract: dataset name, TRAIN/TEST split, labels, class balance, channels/features, maximum sequence length, padding/truncation policy, missing-value policy, and leakage notes. - Use official UEA layout: `--data UEA`, `--root_path ./dataset//`, `--model_id `, with `_TRAIN.ts` and `_TEST.ts`. - The loader uses `sktime.datasets.load_from_tsfile_to_dataframe(..., replace_missing_vals_with='NaN')`, converts labels to integer class codes, interpolates missing values, normalizes features, and pads/clips batches through `collate_fn`. - Runtime tensors are `(batch_size, padded_length, feature_dim)` plus `padding_mask`; this is not sklearn/sktime `numpy3D`. - For unequal length, TSLib sets `seq_len` to the max train/test length and uses padding masks; if dimensions inside one sample have different length, the loader subsamples long series. Read `references/tslib-classification-data.md` before adapting data, using variable length, changing normalization, or adding custom splits. ## Model Selection Use `--task_name classification` and an official model name. Source-supported classification models include: `Autoformer`, `Crossformer`, `DLinear`, `ETSformer`, `FEDformer`, `FiLM`, `Informer`, `LightTS`, `MICN`, `MSGNet`, `MambaSingleLayer`, `Nonstationary_Transformer`, `PatchTST`, `Pyraformer`, `Reformer`, `SegRNN`, `TimeFilter`, `TimeMixer`, `TimesNet`, `Transformer`, and `iTransformer`. Official bash scripts exist for a smaller set: `Autoformer`, `Crossformer`, `DLinear`, `ETSformer`, `FEDformer`, `FiLM`, `Informer`, `LightTS`, `MICN`, `MambaSL` (`--model MambaSingleLayer`), `PatchTST`, `Pyraformer`, `Reformer`, `TimesNet`, `Transformer`, and `iTransformer`. Read `references/tslib-classification-models.md` before claiming support for models without classification scripts or models that raise `NotImplementedError`. ## Training And Prediction Pattern ```bash python -u run.py \ --task_name classification \ --is_training 1 \ --root_path ./dataset/Heartbeat/ \ --model_id Heartbeat \ --model TimesNet \ --data UEA \ --e_layers 3 \ --batch_size 16 \ --d_model 16 \ --d_ff 32 \ --top_k 1 \ --des Exp \ --itr 1 \ --learning_rate 0.001 \ --train_epochs 30 \ --patience 10 ``` `Exp_Classification` trains with `RAdam`, `CrossEntropyLoss`, gradient clipping, early stopping on validation accuracy, and saves `checkpoint.pth`. Test mode uses `--is_training 0` and loads the matching checkpoint. TSLib does not expose a documented sklearn-like `predict_proba`. Internally it computes `softmax(logits)` then `argmax`; use custom code around `Exp_Classification.test` if calibrated probabilities, ROC-AUC, or exported predictions are required. ## Evaluation - Native metric: accuracy via `cal_accuracy`. - For balanced classes, report accuracy plus optional external macro/weighted F1 from saved logits/predictions. - For imbalanced classes, add external `f1_macro`, balanced accuracy, per-class recall/precision, confusion matrix, ROC-AUC/PR-AUC where probabilities are exported. - TSLib does not document stratified cross-validation. To use stratified CV, create fold-specific UEA TRAIN/TEST directories outside the repo workflow, keep every preprocessing step inside each fold, and run one script per fold. ## Anti-Leakage Rules - Split TRAIN/TEST before normalization, interpolation choices, augmentation, padding/truncation tuning, and any external feature engineering. - Fit scalers/normalizers only on TRAIN for each fold. The stock `UEAloader` normalizes each loaded split independently; if strict train-fitted scaling is required, modify the loader deliberately and document it. - Use augmentation only on TRAIN; the official loader checks `flag == "TRAIN"` before applying augmentation. - Do not compute max sequence length, class mapping, or missing-value policies from a held-out final test set unless that metadata is part of a fixed public benchmark contract. - Use stratified folds when you create custom CV for imbalanced classes; preserve subject/device groups if present. ## Common Errors - Passing arrays directly instead of UEA `.ts` files and `--data UEA`. - Mismatching `--root_path`, `--model_id`, and filenames; the loader expects `_TRAIN.ts` or `_TEST.ts`. - Treating `--seq_len` like a user-chosen fixed window; classification resets it from max train/test sequence length during model build. - Expecting native `fit`, `predict`, or `predict_proba` methods. - Using models present in `models/` that explicitly do not support classification, such as `KANAD`, `MultiPatchFormer`, `TiDE`, `WPMixer`, or forecast-only Mamba variants. - Reading validation/test accuracy printed during training as a leak-free benchmark without checking the experiment split design. ## References - Read `references/tslib-classification-models.md` for supported models, script coverage, and model limitations. - Read `references/tslib-classification-data.md` for UEA `.ts` format, tensor shapes, padding, normalization, CV, and metrics. - Read `references/official-sources.md` for official sources consulted. - Use `scripts/validate_uea_ts_split.py` to check required TRAIN/TEST `.ts` files before running TSLib. ## Ready Checklist - `ts-classification-data-prep` contract is complete and leakage risks are documented. - Dataset directory has `_TRAIN.ts` and `_TEST.ts`. - Labels/classes, channel count, sequence lengths, missing values, and imbalance are checked. - Model name is source-supported for `task_name=classification` and required dependencies are installed. - TRAIN-only preprocessing, augmentation, and fold generation are enforced. - Accuracy plus imbalance-aware external metrics are reported on held-out folds/test data.