Fast-and-Frugal-Transfer-Learning

This repository provides a lightweight and energy-aware transfer learning strategy designed for environments with limited hardware resources. Unlike traditional fine-tuning, the method decouples feature extraction from classifier training, avoiding unnecessary iterations through the backbone.

This work was published at the International Conference on Intelligent Data Engineering and Automated Learning (IDEAL 2025) and received the Best Paper Award🏆 (Springer Link).

Key Features

• Precomputed Features: Extract features once from the backbone to drastically reduce training time.
• Thresholded Batch Normalization Adaptation: Efficiently adapts batch normalization statistics to new domains.
• Redesigned Classifier Head: Improves generalization with negligible overhead.
• Margin-based Weighted Training: Focuses on ambiguous samples to improve accuracy.

Results on Brain Cancer MRI, BreakHis and PatchCamelyon datasets show up to 22× faster training and 25× lower CO₂ emissions, while maintaining or improving accuracy compared to full fine-tuning.

---

Setup

1. Clone the repository

   git clone https://github.com/DanielVilaCruz/Fast-and-Frugal-Transfer-Learning.git
   cd Fast-and-Frugal-Transfer-Learning

2. Install PyTorch

   This project requires PyTorch (with CUDA or CPU support).

   Example (CUDA 12.1):

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

   Example (CPU):

   pip install torch torchvision torchaudio

3. Install project dependencies

   pip install -r requirements.txt

4. Download datasets and place them under the /data folder.

• Brain Cancer MRI: https://data.mendeley.com/datasets/mk56jw9rns/1
• BreakHis: https://web.inf.ufpr.br/vri/databases/breast-cancer-histopathological-database-breakhis/
• PatchCamyleon:https://github.com/basveeling/pcam

Usage

Experiments are configured in experiments.yaml. Each entry corresponds to a single experiment, and multiple ones can be added to execute them sequentially.

experiments:
  - method: "ours"
    base_backbone: "resnet18"
    dataset: "brain_cancer"
    n_real_samples: 0
    num_runs: 3
    lr: 0.001
    amplify_factor: 10.0
    patience: 5
    unfreeze_layers: 2
    simple_head: false

  - method: "fine-tune"
    base_backbone: "densenet121"
    dataset: "breakhis"
    n_real_samples: 1000
    num_runs: 1

Parameters

• method: training mode
  • "ours": precomputed features + lightweight classifier
  • "fine-tune": standard fine-tuning
• base_backbone: model architecture
  • Options: "resnet18", "resnet50", "mobilenet_v3_large", "densenet121"
• dataset: dataset to use
  • Options: "brain_cancer", "breakhis", "pcam"
• n_real_samples: maximum number of samples to use
  • 0: use all available samples
• num_runs: number of times to repeat the experiment (averages results)
• lr: learning rate
• amplify_factor: weight assigned to ambiguous samples in margin-based training
• patience: early stopping patience
• unfreeze_layers (only for fine-tuning method): number of last layers to unfreeze in the backbone
• simple_head: whether to use a minimal classifier head (true) or the improved one (false)

Then launch experiments with:

python launch_exp.py

By default, results (metrics, timings, CO₂ emissions) are stored in results.json and exps/.

---

Results

Brain Cancer MRI Results

Architecture	Fine-tuning			Ours
	Acc. (%)	Time (s)	CO₂ (kg)	Acc. (%)	Time (s)	CO₂ (kg)
ResNet18	99.16	194.672	0.0038	98.42	32.486	0.0006
ResNet50	99.25	311.550	0.0064	97.75	57.407	0.0011
MobileNetV3	93.93	770.210	0.0151	99.71	34.780	0.0006
DenseNet121	99.58	625.140	0.0127	96.76	58.204	0.0011

BreakHis Results

Architecture	Fine-tuning			Ours
	Acc. (%)	Time (s)	CO₂ (kg)	Acc. (%)	Time (s)	CO₂ (kg)
ResNet18	99.33	367.299	0.0069	98.50	42.621	0.0008
ResNet50	97.66	513.383	0.0104	98.33	59.743	0.0012
MobileNetV3	99.00	430.983	0.0087	98.67	57.147	0.0010
DenseNet121	93.46	3970.622	0.0794	94.52	184.374	0.0038

PatchCamelyon (PCam) Results

Architecture	Fine-tuning			Ours
	Acc. (%)	Time (s)	CO₂ (kg)	Acc. (%)	Time (s)	CO₂ (kg)
ResNet18	79.78	5953.25	0.1202	79.03	841.32	0.0163
ResNet50	83.54	10191.44	0.2140	81.17	1114.45	0.0229
MobileNetV3	82.96	8009.09	0.1665	82.64	695.83	0.0134
DenseNet121	84.22	20530.34	0.4382	80.67	1287.82	0.0264

Citation

If you use this work, please cite:

@inproceedings{vila2025fast,
  title={Fast and Frugal Transfer Learning via Precomputed Features and Adaptive Normalization},
  author={Vila-Cruz, Daniel and Bol{\'o}n-Canedo, Ver{\'o}nica and Mor{\'a}n-Fern{\'a}ndez, Laura},
  booktitle={International Conference on Intelligent Data Engineering and Automated Learning},
  pages={143--149},
  year={2025},
  organization={Springer}
}