You can deploy this model and start earning money today!

AXAIIT/
AdaMCF

Views

No views yet

AdaMCF: Adaptive Multi-Criteria Fusion Pruning

This repository contains model artifacts and reference implementations for AdaMCF, an adaptive multi-criteria fusion framework for structured neural network pruning.

AdaMCF targets the two coupled decisions in structured pruning:

What to prune: dependency-group importance is estimated by fusing multiple criteria with loss-guided adaptive weights.
When to prune: a cosine-front sparsification schedule allocates more pruning budget to earlier and middle training stages, where networks are generally more plastic.

The method is designed for practical structured compression across CNNs, vision Transformers, object detectors, and pretrained language models.

🔗 Links: 💻 GitHub — Source Code

Highlights

Adaptive multi-criteria fusion: combines multiple importance criteria instead of relying on a fixed single heuristic.
Loss-guided online update: learns fusion weights through a soft-mask proxy and zeroth-order finite-difference estimation.
Dependency-group pruning: removes structurally coupled groups such as filters, channels, attention-related units, and model-specific dependency groups.
Cosine-front scheduling: progressively assigns pruning budget according to training-stage sensitivity.
Broad evaluation: includes CIFAR-10, CIFAR-100, ImageNet-1K, COCO2017, and SQuAD v1.1 experiments.

Repository Structure

text

test/
|-- Bert/          # BERT / SQuAD artifacts
|-- Deit-Small/    # DeiT-Small ImageNet artifacts
|-- DenseNet121/   # DenseNet-121 CIFAR-100 artifacts
|-- GoogLeNet/     # GoogLeNet CIFAR-100 artifacts
|-- ResNet18/      # ResNet-18 CIFAR-100 artifacts
|-- ResNet34/      # ResNet-34 CIFAR-100 artifacts
|-- ResNet50/      # ResNet-50 ImageNet artifacts
|-- ResNet56/      # ResNet-56 CIFAR-10 artifacts
|-- ResNet110/     # ResNet-110 CIFAR-10 artifacts
|-- VGG16/         # VGG-16 CIFAR-10 artifacts
`-- YOLOv5L/       # YOLOv5l / COCO2017 artifacts

Large model files are tracked with Git LFS.

Included Model Artifacts

The repository includes pruned or fine-tuned artifacts for:

Task	Dataset	Models
Image classification	CIFAR-10	VGG-16, ResNet-56, ResNet-110
Image classification	CIFAR-100	DenseNet-121, GoogLeNet, ResNet-18, ResNet-34
Image classification	ImageNet-1K	ResNet-50, DeiT-Small
Object detection	COCO2017	YOLOv5l
Question answering	SQuAD v1.1	BERT

Reported Results

Selected results from the accompanying paper are summarized below.

Model	Dataset / Task	Pruned Metric	FLOPs Reduction	Params Reduction
VGG-16	CIFAR-10	93.02% Top-1	70.58%	94.34%
ResNet-56-M	CIFAR-10	93.40% Top-1	53.53%	65.73%
ResNet-56-H	CIFAR-10	93.27% Top-1	66.60%	73.99%
ResNet-110-M	CIFAR-10	93.51% Top-1	66.87%	79.26%
ResNet-18	CIFAR-100	75.86% Top-1	62.61%	75.08%
ResNet-34	CIFAR-100	80.05% Top-1	78.22%	87.63%
DenseNet-121-H	CIFAR-100	79.00% Top-1	69.83%	57.46%
GoogLeNet-H	CIFAR-100	78.53% Top-1	57.57%	31.18%
ResNet-50-M	ImageNet-1K	75.60% Top-1	49.29%	63.89%
ResNet-50-H	ImageNet-1K	74.49% Top-1	79.85%	74.62%
DeiT-Small-M	ImageNet-1K	79.69% Top-1	34.25%	33.62%
DeiT-Small-H	ImageNet-1K	78.17% Top-1	58.21%	57.16%
YOLOv5l	COCO2017	64.50 mAP@0.5	34.49%	68.41%
BERT	SQuAD v1.1	88.11 F1 / 80.68 EM	33.71%	33.71%
BERT	SQuAD v1.1	87.27 F1 / 79.55 EM	53.93%	53.94%

Git LFS

Clone with Git LFS enabled to retrieve the model weights:

bash

git lfs install
git clone https://github.com/AXAIIT/AdaMCF.git
cd AdaMCF
git lfs pull

Without Git LFS, large files such as .pt, .pth, .bin, and .th will be checked out as pointer files rather than full model weights.

Method Overview

AdaMCF models structured pruning as a spatio-temporal optimization problem.

In the spatial dimension, each dependency group receives multiple importance signals. The method adaptively fuses these criteria with weights updated by task-loss feedback, allowing the pruning criterion to change as the model becomes sparser.

In the temporal dimension, AdaMCF uses a cosine-front sparsification schedule to avoid rigid uniform pruning. This schedule assigns a larger share of pruning to earlier and middle stages and reduces disruptive structural changes later in training.

Citation

If this repository is useful for your work, please cite:

bibtex

@article{adamcf2026,
  title   = {AdaMCF: Spatio-Temporal Collaborative Structured Pruning via Adaptive Multi-Criteria Fusion},
  author  = {AXAIIT},
  year    = {2026}
}

License

This project is released under the MIT License. See LICENSE for details.