Code for Post-Training Neural Network Pruning using Graph Curvature

Neural Graph Definition (Default: w4)

• --metric w4

---

Example Command

(See run_vgg.sh for a full example)

python removal.py --image 1 --metric w4 --model_type cnn --model_name ori \
--model_path CNN/models/cnn/ \
--img_res_path res/w4/relu/cifarori_vgg9_new_curv0/ \
--dataset cifar10 \
--community 1 \
--edge 0 \
--activation relu \
--sample_num 1 \
--img_data_path res/w4/relu/cifarori_vgg9/ \
--alpha 0.9

---

Main Script

• removal.py – main pruning script

---

Arguments

Argument	Description
--image	Image classifier mode
--metric	Neural graph metric (default: w4)
--model_name	Training method: ori (CE), wd (Weight Decay), adv (Adversarial Training)
--model_path	Path to pretrained model
--img_res_path	Path to save pruning results
--dataset	Dataset name (e.g., cifar10)
--community	1: calculate graph curvature
--edge	1: count results and start pruning
--activation	Activation function (relu or tanh)
--sample_num	Number of calibration samples per class used for curvature calculation
--img_data_path	Path to curvature results used for pruning
--alpha	Parameter for curvature calculation (Definition 5 in the paper)

To Do

• dataset

• model pkl