tscglue 0.1.1__tar.gz

tscglue-0.1.1/.github/workflows/publish.yml

@@ -0,0 +1,67 @@
+name: Build, Tag and Publish
+
+on:
+  push:
+    branches:
+      - main
+
+jobs:
+  build-tag-publish:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write   
+      id-token: write   
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v5
+        with:
+          enable-cache: true
+
+      - name: Get version from pyproject.toml
+        id: pyproject
+        run: |
+          VERSION=$(python3 -c "import pathlib, re; \
+            content = pathlib.Path('pyproject.toml').read_text(); \
+            match = re.search(r'^version\s*=\s*[\" \']([^\" \']+)[\" \']', content, re.MULTILINE); \
+            print(match.group(1)) if match else exit(1)")
+          echo "VERSION=$VERSION" >> $GITHUB_OUTPUT
+
+      - name: Check if Tag exists
+        id: tag_check
+        run: |
+          if git rev-parse "v${{ steps.pyproject.outputs.VERSION }}" >/dev/null 2>&1; then
+            echo "EXISTS=true" >> $GITHUB_OUTPUT
+          else
+            echo "EXISTS=false" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Build package
+        run: uv run --with build python -m build
+
+      - name: Update GitHub Release
+        uses: softprops/action-gh-release@v2
+        with:
+          tag_name: v${{ steps.pyproject.outputs.VERSION }}
+          name: Release v${{ steps.pyproject.outputs.VERSION }}
+          files: dist/*
+          # This forces GitHub to move the tag and overwrite the files
+          overwrite: true 
+          # Ensures the tag is updated to the current commit if it already exists
+          force_tag_update: true 
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Publish to PyPI
+        # CRITICAL: We skip this if the version already exists on PyPI
+        # because TestPyPI will reject the upload and fail the workflow.
+        if: steps.tag_check.outputs.EXISTS == 'false'
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          # repository-url: https://test.pypi.org/legacy/
+          print-hash: true

tscglue-0.1.1/.github/workflows/tests.yml

@@ -0,0 +1,24 @@
+name: Tests
+
+on:
+  push:
+    branches: [ main, develop ]
+  pull_request:
+    branches: [ main, develop ]
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+
+    steps:
+    - name: Checkout code
+      uses: actions/checkout@v4
+
+    - name: Install uv
+      run: |
+        curl -LsSf https://astral.sh/uv/install.sh | sh
+        echo "$HOME/.cargo/bin" >> $GITHUB_PATH
+
+    - name: Run tests
+      run: |
+        make tests

tscglue-0.1.1/.gitignore

@@ -0,0 +1,36 @@
+# Experiment files
+experiments/
+
+# Python cache
+__pycache__/
+*.py[cod]
+*$py.class
+
+# Virtual environments
+.venv/
+venv/
+env/
+ENV/
+
+# IDE
+.vscode/
+.idea/
+*.swp
+*.swo
+*~
+
+# Jupyter Notebook checkpoints
+.ipynb_checkpoints/
+
+# Distribution / packaging
+dist/
+build/
+*.egg-info/
+
+# uv lock file (library - not needed in git)
+uv.lock
+
+# Data
+data/
+figures/
+results/

tscglue-0.1.1/Makefile

@@ -0,0 +1,38 @@
+.PHONY: help install-uv setup list clean tests format download-ucr download-models
+.ONESHELL:
+
+help:   ## Show available commands
+	@grep -E '^[a-zA-Z_-]+:.*?## .*$$' $(MAKEFILE_LIST) | sort | awk 'BEGIN {FS = ":.*?## "}; {printf "\033[36m%-30s\033[0m %s\n", $$1, $$2}'
+
+install-uv:  ## Install uv package manager
+	curl -LsSf https://astral.sh/uv/install.sh | sh
+
+setup:  ## Sets up everything needed for a new deployment
+	uv sync --all-extras
+
+list:
+	@LC_ALL=C $(MAKE) -pRrq -f $(firstword $(MAKEFILE_LIST)) : 2>/dev/null | awk -v RS= -F: '/(^|\n)# Files(\n|$$)/,/(^|\n)# Finished Make data base/ {if ($$1 !~ "^[#.]") {print $$1}}' | sort | grep -E -v -e '^[^[:alnum:]]' -e '^$@$$'
+
+clean: ## Removes env, docs and caches
+	rm -rf build/docs
+	rm -rf ~/.exturion
+	rm -rf .venv
+	uv clean all
+	uv cache clean
+
+tests: ## Run the unit tests
+	uv run --extra dev pytest tests/ -vv -W ignore::DeprecationWarning --capture=no --durations=0 --cache-clear --maxfail=1
+
+format: ## Format the code with isort and ruff
+	uv run --extra dev isort . --profile black
+	uv run --extra dev ruff format .
+	uv run --extra dev ruff check . --fix
+
+download-models: ## Pre-download HF models (Mantis, Chronos-2) for offline/SLURM use
+	uv run --no-sync python -c "from tscglue.models_tsfm import download_models; download_models()"
+
+download-ucr: ## Download and unzip UCR archive (all folds) into data/
+	mkdir -p data
+	curl -L -o data/ucr.zip 'https://drive.usercontent.google.com/download?id=1V36LSZLAK6FIYRfPx6mmE5euzogcXS83&export=download&authuser=0&confirm=t&uuid=07e23200-74c3-4fd6-ba24-c5cde6e39a45&at=APcXIO39z41iEW4mVw4ltHUn9yYC%3A1769851071815'
+	unzip -o data/ucr.zip -d data/
+	rm data/ucr.zip

tscglue-0.1.1/PKG-INFO

@@ -0,0 +1,74 @@
+Metadata-Version: 2.4
+Name: tscglue
+Version: 0.1.1
+Summary: Automatic Time Series Classification
+Requires-Python: <3.14,>=3.12
+Requires-Dist: aeon>=1.3.0
+Requires-Dist: click
+Requires-Dist: huggingface-hub
+Requires-Dist: imblearn
+Requires-Dist: polars
+Requires-Dist: pyarrow
+Requires-Dist: pytorch-lightning
+Requires-Dist: scikit-learn
+Requires-Dist: seaborn
+Requires-Dist: statsmodels
+Requires-Dist: tabicl
+Requires-Dist: torch
+Requires-Dist: tqdm
+Requires-Dist: tsfresh
+Provides-Extra: dev
+Requires-Dist: awscli; extra == 'dev'
+Requires-Dist: boto3; extra == 'dev'
+Requires-Dist: isort; extra == 'dev'
+Requires-Dist: moto[s3]; extra == 'dev'
+Requires-Dist: pytest; extra == 'dev'
+Requires-Dist: ruff; extra == 'dev'
+Requires-Dist: s3fs; extra == 'dev'
+Provides-Extra: notebooks
+Requires-Dist: accelerate; extra == 'notebooks'
+Requires-Dist: catboost; extra == 'notebooks'
+Requires-Dist: chronos-forecasting; extra == 'notebooks'
+Requires-Dist: hvplot; extra == 'notebooks'
+Requires-Dist: ipykernel; extra == 'notebooks'
+Requires-Dist: jupyter; extra == 'notebooks'
+Requires-Dist: jupyterlab; extra == 'notebooks'
+Requires-Dist: lightgbm; extra == 'notebooks'
+Requires-Dist: mantis-tsfm; extra == 'notebooks'
+Requires-Dist: pandas; extra == 'notebooks'
+Requires-Dist: transformers; extra == 'notebooks'
+Description-Content-Type: text/markdown
+
+# TSCGlue
+
+Automatic Time Series Classification library built on top of aeon and scikit-learn.
+
+## Installation
+
+```bash
+pip install tscglue
+```
+
+## Quick Start
+
+```python
+from tscglue import utils
+from tscglue.models import TSCGlue
+from sklearn.metrics import accuracy_score
+
+# Load a time series classification dataset
+X_train, y_train, X_test, y_test = utils.load_dataset("ArrowHead")
+
+# Create and train the model
+model = TSCGlue(
+    random_state=270,
+    k_folds=10,
+    n_jobs=-1
+)
+model.fit(X_train, y_train)
+
+# Make predictions
+y_pred = model.predict(X_test)
+accuracy = accuracy_score(y_test, y_pred)
+print(f"Accuracy: {accuracy:.4f}")
+```

tscglue-0.1.1/README.md

@@ -0,0 +1,33 @@
+# TSCGlue
+
+Automatic Time Series Classification library built on top of aeon and scikit-learn.
+
+## Installation
+
+```bash
+pip install tscglue
+```
+
+## Quick Start
+
+```python
+from tscglue import utils
+from tscglue.models import TSCGlue
+from sklearn.metrics import accuracy_score
+
+# Load a time series classification dataset
+X_train, y_train, X_test, y_test = utils.load_dataset("ArrowHead")
+
+# Create and train the model
+model = TSCGlue(
+    random_state=270,
+    k_folds=10,
+    n_jobs=-1
+)
+model.fit(X_train, y_train)
+
+# Make predictions
+y_pred = model.predict(X_test)
+accuracy = accuracy_score(y_test, y_pred)
+print(f"Accuracy: {accuracy:.4f}")
+```

tscglue-0.1.1/experimental/README.md

@@ -0,0 +1,152 @@
+# AutoTSC
+
+- Does selection models based on val performance even work?
+- Which ensemble method should AutoML use as a function of dataset size? (stacking double stacking weights like now?)
+- Does the validation-test disconnect generalize across domains, or is it TSC-specific?
+- Does nested CV actually solve the small-sample problem?
+- Does downsampling work?
+- Does multifidelity work?
+- No AutoML due to small datasets?
+- How often is best val model best on train (shuffled/no shuffled)
+- Resampling should be done or not?
+
+
+✅ 1. First-order & local-transform views
+
+These change the shape or local structure of the series.
+
+✔ Differencing (Δx)
+
+Good for removing trends, enhancing sharp transitions.
+
+✔ Cumulative sum
+
+Smooths noise; emphasises long-term structure.
+
+✔ Moving average / smoothing (SG filter, EMA)
+
+Suppresses high-frequency noise → different model inductive bias.
+
+✔ Trend removal (detrending)
+
+Removes global shape and highlights wiggles.
+
+✅ 2. Frequency & phase transforms
+
+Often extremely useful because they create orthogonal representations.
+
+✔ FFT magnitude
+
+Spectral view of the series — deep models love it.
+
+✔ FFT phase
+
+Adds complementary structure to magnitude.
+
+✔ STFT / Sliding FFT
+
+Time-frequency representation (use 1D vector summary or window-level stats).
+
+✔ Wavelet transform (CWT, DWT)
+
+Great for multi-resolution patterns.
+
+✔ Hilbert transform
+
+Creates an analytic signal: amplitude envelope + instantaneous phase.
+
+✅ 3. Shape & geometric transforms
+
+Excellent for diversity because they are structurally different.
+
+✔ Time warping (random or deterministic)
+
+Warps the timeline → great for shape-based algorithms.
+
+✔ Curve length transform
+
+Turns a series into cumulative path length.
+
+✔ Polar coordinate transform
+
+Convert (x, diff(x)) into polar angle + magnitude.
+
+✔ Slope transform
+
+Use local slope or angle instead of raw values.
+
+✅ 4. Normalization-based transforms
+
+Don’t underestimate these — especially helpful with ROCKET/Hydra ensembles.
+
+✔ Z-normalization (per series)
+
+Baseline for most TSC but provides a different view from raw scale.
+
+✔ Min–max scaling
+
+Good for emphasising relative shape.
+
+✔ Unit energy / L2 normalization
+
+Highlights relative oscillations.
+
+✔ Robust scaling (median/IQR)
+
+When outliers distort structure.
+
+✅ 5. Feature extraction transforms
+
+These produce feature vectors that a classifier sees differently from the raw TS.
+
+✔ Catch22
+
+22 interpretable features — often complementary to ROCKET.
+
+✔ TSFresh / TSFEL feature subsets
+
+Huge diversity if you select subsets.
+
+✔ Autocorrelation / partial autocorrelation vectors
+
+Very different inductive bias.
+
+✔ Shapelet distances
+
+Distance to “prototype” shapes.
+
+✅ 6. Windowing & multi-resolution views
+
+Often extremely strong for ensembling.
+
+✔ Multi-scale segment averaging
+
+Compute downsampled versions at multiple resolutions.
+
+✔ Piecewise transforms
+
+PAA (Piecewise aggregate approximation)
+
+PLA (Piecewise linear approximation)
+
+SAX (Symbolic Aggregate Approximation)
+
+✔ Moving-window statistics
+
+Rolling min/max/std/skew/kurt.
+
+✅ 7. Noise & augmentation transforms (for diversity)
+
+Used often in Hydra/ROCKET ensembles to create diverse models.
+
+✔ Add small Gaussian noise
+
+Mild regularization; different learned filters.
+
+✔ Jittering / scaling
+
+Preserves topology but shifts amplitude.
+
+✔ Dropout segments
+
+Removes random subsequences → encourages robustness.