CAFT: Constraint-Guided Adaptive Fine-Tuning of Neural Networks

Project Overview

CAFT is a research codebase accompanying the paper:

Making Limited Data Count: Constraint-Guided Adaptive Fine-Tuning of Neural Networks for Cost-Aware Defect Classification
Kristina Dachtler, Alexander Schiendorfer
ETFA 2026 — Special Session: Addressing Data Scarcity

In manufacturing environments, labeled data is often scarce, imbalanced, and of uncertain quality. Under these conditions, optimizing for conventional aggregate metrics quickly reaches a plateau — and further improvements do not necessarily translate into better operational outcomes. Defining a suitable cost matrix for cost-sensitive learning is non-trivial, and static cost-sensitive training methods tend to collapse on small, imbalanced datasets.

CAFT addresses these challenges with a two-phase training strategy:

1. Phase 1 — Pre-Training: A neural network is trained with standard Cross-Entropy (or Weighted Cross-Entropy) loss to establish stable feature representations.
2. Phase 2 — Constraint-Aware Fine-Tuning: Starting from the pre-trained weights, misclassification penalties are incrementally increased until predefined operational decision constraints are satisfied. The method does not require a predefined cost matrix — instead, constraint-relevant cost entries are built up adaptively via Lagrangian multipliers.

Key Results

• CAFT achieves higher constraint satisfaction rates than static cost-sensitive methods across two datasets, without model collapse.
• CAFT maintains the highest F1-score among all methods that fulfill the decision constraint.
• Compared to post-hoc threshold tuning, CAFT shows more robust generalization from validation to test data.
• The approach scales naturally to multiple simultaneous constraints, a setting where threshold-based alternatives become impractical.

Datasets

• Paint Defects (real-world): Automotive paint defect classification with 2,681 samples, 5 features, and 3 classes (not included due to confidentiality reasons).
• Steel Plates Faults (public benchmark): Adapted from the UCI Steel Plates Faults dataset with 1,941 samples, 27 features, grouped into 3 classes.

Installation

1. Clone the repository:

   git clone https://github.com/AImotion-Bavaria/caft
   cd CAFT

2. Install dependencies (Python >=3.10):

   pip install -e .

   This uses the dependencies specified in pyproject.toml.

Directory Structure (excerpt)

• src/ — Core modules (models, training, evaluation, utilities)
• config/ — Central configuration (paths)
• artifacts/ — Data, models (model_weights from pre-training), and results (auto-generated)
• experiments/ — Experiment scripts for different datasets
• create_experiment_plots.py — Aggregates and plots experiment results
• create_experiment_summary.py — Summarizes results from text outputs

Running an Example Experiment

To run a neural network experiment on the steel plates dataset:

python experiments/steel_plates/01_torch_NN_training_SP.py

Aggregating Results

Results are aggregated and plotted after running experiments:

python create_experiment_plots.py
python create_experiment_summary.py

License

This project is licensed under the MIT License.

Author

Kristina Dachtler (kristina.dachtler@thi.de)