ictonyx 0.4.6

Ictonyx: Iteration Comparison Testing Over N-runs: Yield eXamination

Ictonyx

Iteration Comparison Testing Over N-runs: Yield eXamination

A Python framework for studying machine learning model variability and performing rigorous statistical comparisons.

---

The problem

Training a machine learning model involves stochastic factors: random weight initialisation, data shuffling, dropout. Train the same architecture on the same dataset twice and you will get different weights, different predictions, and different evaluation metrics.

This means a model's performance is a random variable, not a constant — and treating it as a constant, as most practitioners do, leads to conclusions that cannot be replicated or trusted. Reporting a single accuracy from a single training run is not a measurement; it is a sample of size one.

Ictonyx trains a model N times under independent random seeds, collects the full distribution of outcomes, and provides the statistical machinery to reason about that distribution rigorously.

---

Installation

pip install ictonyx

Optional extras

Extra	What it includes	Install
sklearn	scikit-learn, joblib	pip install ictonyx[sklearn]
tensorflow	TensorFlow, Keras	pip install ictonyx[tensorflow]
torch	PyTorch	pip install ictonyx[torch]
mlflow	MLflow tracking	pip install ictonyx[mlflow]
explain	SHAP explainability	pip install ictonyx[explain]
tuning	Optuna hyperparameter tuning	pip install ictonyx[tuning]
isolation	Process isolation for GPU runs	pip install ictonyx[isolation]
progress	tqdm progress bars	pip install ictonyx[progress]
all	Everything above	pip install ictonyx[all]

Extras can be combined:

pip install "ictonyx[tensorflow,isolation]"

Requires Python 3.10+. Current release: 0.4.6 — changelog · PyPI

---

What variability does Ictonyx measure?

Ictonyx characterizes training-stochastic variance under a fixed data split. The reported uncertainty is the variance of your metric across different seeds — initialization, batch order, augmentation, dropout — holding the train/val/test split constant. This answers the question: "How stable is this training procedure?"

Ictonyx does not currently measure sampling variance — the uncertainty that comes from which examples landed in train vs. validation vs. test. If you change the split, you get a different answer about model quality, and that difference is its own source of uncertainty, distinct from seed stochasticity.

For sampling-variance analysis today, pair Ictonyx with an outer k-fold loop. In a later release, Ictonyx will ship ResamplingPolicy for nested (data × seed) designs with corrected comparison tests (Nadeau-Bengio, Bouckaert, Dietterich 5×2).

Reference: Bouthillier et al. (2021), Accounting for Variance in Machine Learning Benchmarks, MLSys.

---

Quick start

Train a small feed-forward network on the wine data from sklearn twenty times and observe the distribution of outcomes. Here, we'll use a simple Tensorflow model.

import tensorflow as tf
from sklearn.datasets import load_wine
from sklearn.preprocessing import StandardScaler
import ictonyx as ix

data = load_wine()
X = StandardScaler().fit_transform(data.data)
y = data.target


def build_model(config):
    model = tf.keras.Sequential([
        tf.keras.layers.Dense(16, activation='relu'),
        tf.keras.layers.Dense(16, activation='relu'),
        tf.keras.layers.Dense(3, activation='softmax')
    ])
    model.compile(optimizer='adam',
                  loss='sparse_categorical_crossentropy',
                  metrics=['accuracy'])
    return ix.KerasModelWrapper(model)

results = ix.variability_study(
    model=build_model,
    data=(X, y),
    runs=20,
    epochs=20,
    seed=42,
    verbose=False
)

print(results.summarize())

Variability Study Results
==============================
Successful runs: 20
Seed: 42

Test Set Metrics:
--------------------
accuracy:
  Mean:             0.9097
  SD (sample, N-1): 0.0467
  Min:              0.8333
  Max:              0.9722
loss:
  Mean:             0.5285
  SD (sample, N-1): 0.1085
  Min:              0.3791
  Max:              0.7748

Validation Metrics:
--------------------
train_accuracy:
  Mean:             0.8851
  SD (sample, N-1): 0.0436
  Min:              0.8065
  Max:              0.9355
train_loss:
  Mean:             0.5507
  SD (sample, N-1): 0.0987
  Min:              0.4288
  Max:              0.7760
val_accuracy:
  Mean:             0.9194
  SD (sample, N-1): 0.0709
  Min:              0.7222
  Max:              1.0000
val_loss:
  Mean:             0.4981
  SD (sample, N-1): 0.1105
  Min:              0.3793
  Max:              0.7476

On a 178-sample dataset, the same architecture achieves anywhere from 72% to 100% validation accuracy depending solely on the random seed — a 28-point range that a single-run evaluation would never reveal. Ictonyx also allows plotting of training histories:

ix.plot_variability_summary(results=results, metric='accuracy')

---

Comparing two models

A routine question in applied ML: does one model actually outperform another, or did it just draw a better random seed? A single run of each tells you nothing — the difference could be real, or it could be noise. Ictonyx answers this by running both models the same number of times and applying a statistical test to the resulting distributions.

Ictonyx works equally well with sklearn estimators — pass a class or configured instance directly, no wrapper required.

from sklearn.neural_network import MLPClassifier
from sklearn.ensemble import RandomForestClassifier

comparison = ix.compare_models(
    models=[
        MLPClassifier(hidden_layer_sizes=(64,), max_iter=200),
        RandomForestClassifier(n_estimators=40),
    ],
    data=(X, y),
    runs=20,
    metric='val_accuracy',
    seed=42,
    verbose=False,
)

print(comparison.get_summary())

ix.plot_comparison_boxplots(comparison)

Model Comparison Results (val_accuracy)
========================================
Models compared: 2
Omnibus test: Paired Wilcoxon Signed-Rank Test: 5.500, p=0.0005 ***, r (effect size)=0.815

Pairwise comparisons (none correction):
  MLPClassifier_vs_RandomForestClassifier: Paired Wilcoxon Signed-Rank Test: 5.500, p=0.0005 ***, r (effect size)=0.815 *

Significant pairs: MLPClassifier_vs_RandomForestClassifier

Both models received identical per-run seeds, so each MLP run is directly paired with the corresponding Random Forest run. The Wilcoxon signed-rank test exploits this pairing to remove noise common to both models.

The MLP outperformed the Random Forest with an effect size of r=0.815 (p=0.0005). That is not a marginal result that might reverse on a different seed.

The 'none correction' label indicates that with only two models there is a single pair to test, so no multiple-comparison correction is applied

---

Process isolation for GPU runs

Keras models accumulate GPU memory across training runs in a loop. For studies with many runs or large models, run each training session in a subprocess:

results = ix.variability_study(
    model=build_model,
    data=(X, y),
    runs=20,
    epochs=20,
    use_process_isolation=True,
    gpu_memory_limit=4096,
    seed=42,
)

Each run executes in a child process and exits cleanly, releasing all GPU memory before the next run begins.

---

PyTorch

import torch
import torch.nn as nn
import ictonyx as ix
from ictonyx import PyTorchModelWrapper, ArraysDataHandler

def build_net(config: ModelConfig) -> PyTorchModelWrapper:
    model = nn.Sequential(
        nn.Linear(30, 64), nn.ReLU(),
        nn.Linear(64, 32), nn.ReLU(),
        nn.Linear(32, 2),
    )
    return PyTorchModelWrapper(
        model,
        criterion=nn.CrossEntropyLoss(),
        optimizer_class=torch.optim.Adam,
        optimizer_params={'lr': config.get('learning_rate', 0.001)},
        task='classification',
    )

# Pass arrays directly — Ictonyx handles splitting
import numpy as np
from sklearn.datasets import load_breast_cancer
data = load_breast_cancer()
X = data.data.astype(np.float32)
y = data.target.astype(np.int64)

results = ix.variability_study(
    model=build_net,
    data=ArraysDataHandler(X, y, val_split=0.2, test_split=0.1),
    runs=20,
    seed=42,
)

ix.plot_variability_summary(results=results, metric='accuracy')

---

Working with results

# Full distribution of any metric across runs
results.get_metric_values('val_accuracy')       # List[float]

# Per-epoch statistics across all runs
results.get_epoch_statistics('val_accuracy')    # DataFrame: epoch, mean, sd, se, ci_lower, ci_upper

# All per-run, per-epoch DataFrames
results.all_runs_metrics                        # List[pd.DataFrame]

# Seed for exact reproducibility
results.seed

---

Examples

The examples/ directory contains Jupyter notebooks:

• quickstart.ipynb — wine dataset variability study and three-model comparison using Keras and sklearn. No GPU required for the comparison sections.
• 01_mnist_variability_study.ipynb — deep dive into Keras CNN variability on MNIST with full visualisation
• 02_mnist_model_comparison.ipynb — comparing two CNN architectures statistically
• 03_learning_rate_variability.ipynb — hyperparameter sweep across learning rates and batch sizes using run_grid_study()
• 04_pytorch_classification.ipynb — PyTorch classification variability study with epoch-level diagnostics
• 05_pytorch_regression.ipynb — PyTorch regression variability study with known ground truth
• 06_sklearn_models.ipynb — sklearn classification and regression: single-model variability and multi-model comparison

---

License

MIT. See LICENSE.

---

Citation

If you use Ictonyx in published work, please cite it using the metadata in CITATION.cff, or use the Cite this repository button on the GitHub repository page.

@software{kizlik_ictonyx,
  author  = {Kizlik, Stephen},
  title   = {Ictonyx: A Framework for Variability Analysis in Machine Learning Training},
  version = {0.4.6},
  url     = {https://github.com/skizlik/ictonyx},
  license = {MIT},
}