Divide and Conquer Networks (DCN)

Code accompanying Divide and Conquer Networks

DCN Summary

Model

Weak supervision Loss

Gradients computation

Reproduce Experiments

Prerequisites

• Computer with Linux or OSX
• PyTorch
• For training, an NVIDIA GPU is needed. CPU not supported.

Comments

• For information about input arguments check main script.
• For changing datasets default configuration modify the corresponding data_generator init

Sorting

Baseline

python code/Sorting/main.py --path [experiment folder] --path_dataset [dataset folder]

Mergesort (Fix Split)

python code/Sorting/main.py --path [experiment folder] --path_dataset [dataset folder] --dynamic --mergesort_split

Quicksort (Fix Merge)

python code/Sorting/main.py --path [experiment folder] --path_dataset [dataset folder] --dynamic --quicksort_merge

Joint Training

python code/Sorting/main.py --path [experiment folder] --path_dataset [dataset folder] --dynamic

Convex Hull

Baseline

python code/ConvexHull2d/main.py --path [experiment folder] --path_dataset [dataset folder]

Without split computational regularization

python code/ConvexHull2d/main.py --path [experiment folder] --path_dataset [dataset folder] --dynamic

Add split computational regularization

python code/ConvexHull2d/main.py --path [experiment folder] --path_dataset [dataset folder] --dynamic --regularize_split

K-means

python code/K-means/main.py --path [experiment folder] --path_dataset [dataset folder] --num_clusters [number of clusters]