Reproducible classical and deep machine-learning utilities for scientific research.
Full Documentation · uv pip install scitex-ml[all]
| # | Problem | Solution |
|---|---|---|
| 1 | Boilerplate around scikit-learn — every paper re-implements Classifier factories, train/eval loops, classification reports, ROC/PR plots. |
Classifier + ClassificationReporter — thin factory over scikit-learn estimators that snaps directly into a reporter for cross-validation aware metrics, confusion matrices, and figure export. |
| 2 | Time-series CV done by hand — researchers re-derive blocking / sliding-window / calendar splitters per project, often with off-by-one bugs. | Time-series CV splitters — TimeSeriesStratifiedSplit, TimeSeriesBlockingSplit, TimeSeriesSlidingWindowSplit, TimeSeriesCalendarSplit ship with consistent APIs and tested edge cases. |
| 3 | Training-loop ergonomics — EarlyStopping, learning-curve logging, optimiser shortcuts, multi-task losses are all glue code that drifts between repos. |
First-class training utilities — EarlyStopping, LearningCurveLogger, MultiTaskLoss, get_optimizer / set_optimizer, vendored Ranger. |
| 4 | Heavy ML deps mixed with LLM SDKs — installing one pulls all of scikit-learn, torch, openai, anthropic. |
Split package — generative-AI lives in scitex-genai; scitex-ml keeps the classical / deep-ML stack and nothing else. |
pip install scitex-ml # core
pip install scitex-ml[heavy] # + torch / catboost / optuna / pytorch_pretrained_vit
pip install scitex-ml[mcp] # + fastmcp
pip install scitex-ml[all] # everythingThrough the umbrella: pip install scitex[ml]. Requires Python ≥ 3.10.
import scitex_ml
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
X, y = load_iris(return_X_y=True)
X_tr, X_te, y_tr, y_te = train_test_split(X, y, random_state=0, stratify=y)
# Classifier — factory over scikit-learn estimators.
clf = scitex_ml.Classifier("LogisticRegression")
clf.fit(X_tr, y_tr)
print(f"test accuracy: {clf.score(X_te, y_te):.3f}")
# ClassificationReporter — metric tracking + figure export.
reporter = scitex_ml.ClassificationReporter(save_dir="./results")
reporter.calc_metrics(y_te, clf.predict(X_te), clf.predict_proba(X_te))
reporter.summarize()
reporter.save()For a runnable walk-through see examples/01_classification.ipynb.
A complete classification + reporting walk-through (Iris, train/test
split, Classifier("LogisticRegression"), ClassificationReporter
metric persistence) lives in
examples/01_classification.ipynb.
flowchart LR
Data[load_iris<br/>train_test_split] --> Clf[scitex_ml.Classifier"LogisticRegression"]
Clf -->|fit / score| Pred[y_pred / y_proba]
Pred --> Reporter[scitex_ml.ClassificationReporter]
Reporter -->|calc_metrics| Metrics[bacc · ROC-AUC · Conf-Mat]
Reporter -->|save| Artefacts[results/<br/>metrics.csv · roc.png · pr.png]
A second examples/example_classifier.py runs the same flow as a script
so it can be wired into tests/examples/test_example_classifier.py for
CI smoke coverage.
scitex-ml sits in the middle layer of the SciTeX ecosystem:
scitex-python (umbrella)
└── scitex.ml ── thin sys.modules-aliasing shim
└── scitex_ml (this package)
├── classification/ Classifier, ClassificationReporter,
│ time-series CV splitters
├── training/ EarlyStopping, LearningCurveLogger
├── loss/ MultiTaskLoss + regularisers
├── optim/ get_optimizer / set_optimizer + Ranger
├── metrics/ calc_bacc, calc_conf_mat, calc_roc_auc
├── clustering/ PCA + UMAP wrappers
├── feature_extraction/ ViT embeddings
├── feature_selection/ univariate / multivariate
├── plt/ ROC / PR / learning-curve / conf-mat plots
├── sampling/ undersampling helpers
├── sklearn/ scikit-learn integration helpers
└── sk/ sktime compatibility
Cross-package dependencies are minimal: scitex-logging, scitex-io,
scitex-plt, scitex-repro, scitex-types. Heavy deps (torch,
catboost, optuna, pytorch_pretrained_vit) live behind the [heavy]
extra and are gated by _AVAILABLE flags in the source so a pip install scitex-ml install without [heavy] still imports cleanly.
Python API ⭐⭐⭐ (primary)
import scitex_ml
# Time-series cross-validation
from scitex_ml.classification import TimeSeriesStratifiedSplit
splitter = TimeSeriesStratifiedSplit(n_splits=5)
# Training utilities
stopper = scitex_ml.EarlyStopping(patience=10, direction="minimize")
logger = scitex_ml.LearningCurveLogger()
# Multi-task loss + optimiser
mtl = scitex_ml.MultiTaskLoss(are_regression=[False, False])
optimizer = scitex_ml.set_optimizer(model, "adam", lr=1e-3)
# Metrics
result = scitex_ml.metrics.calc_bacc(y_true, y_pred)
cm = scitex_ml.metrics.calc_conf_mat(y_true, y_pred)CLI ⭐ — none
scitex-ml ships no dedicated CLI. ML workflows are composed in Python and run via the umbrella scitex CLI / @scitex.session decorator.
MCP ⭐ — none
No MCP server in this package. The umbrella scitex CLI surfaces ML-adjacent MCP tools (e.g. scitex stats, scitex plt).
Skills ⭐⭐
Skill index for AI agents lives at src/scitex_ml/_skills/scitex-ml/SKILL.md. Sub-skills cover classification, training, loss, optim, clustering, metrics, sampling, feature-selection.
scitex-ml is part of SciTeX. Install via the umbrella with pip install scitex[ml] to use as scitex.ml (Python).
import scitex
scitex.ml.Classifier # same object as scitex_ml.Classifier
scitex.ml.classification.TimeSeriesStratifiedSplit # deep paths resolve via the umbrella shimscitex.ml delegates to scitex_ml — they share the same API.
The SciTeX system follows the Four Freedoms for Research below, inspired by the Free Software Definition:
Four Freedoms for Research
- The freedom to run your research anywhere — your machine, your terms.
- The freedom to study how every step works — from raw data to final manuscript.
- The freedom to redistribute your workflows, not just your papers.
- The freedom to modify any module and share improvements with the community.
AGPL-3.0 — because we believe research infrastructure deserves the same freedoms as the software it runs on.
