ncut-pytorch 2.1.2

Normalized Cut and Spectral Embedding

🌐Documentation (old version) | 🤗HuggingFace Demo

Nyström Normalized Cut

Normalized Cut and spectral embedding, 100x faster than sklean implementation. $O(n)$ time complexity, $O(1)$ space complexity.

https://github.com/user-attachments/assets/f0d40b1f-b8a5-4077-ab5f-e405f3ffb70f

Video: Ncut spectral embedding eigenvectors, on SAM features.

---

Installation

pip install -U ncut-pytorch

Quick Start: plain Ncut

import torch
from ncut_pytorch import Ncut

features = torch.rand(1960, 768)
eigvecs = Ncut(n_eig=20).fit_tranform(features)
  # (1960, 20)

from ncut_pytorch import kway_ncut
kway_eigvecs = kway_ncut(eigvecs)
cluster_assignment = kway_eigvecs.argmax(1)
cluster_centroids = kway_eigvecs.argmax(0)

Quick Start: Ncut DINOv3 Predictor

from ncut_pytorch.predictor import NcutDinov3Predictor
from PIL import Image

predictor = NcutDinov3Predictor(model_cfg="dinov3_vitl16")
predictor = predictor.to('cuda')

images = [Image.open(f"images/view_{i}.jpg") for i in range(4)]
predictor.set_images(images)

image = predictor.summary(n_segments=[10, 25, 50, 100], draw_border=True)
display(image)

More examples and detailed usage can be found in the examples directory.

Performance

• ncut_pytorch.Ncut is $O(n)$ time complexity

• sklearn.SpectralEmbedding is $O(n^2)$ time complexity.

Setup:

CPU: Intel(R) Core(TM) i9-13900K CPU

RAM: 128 GiB

GPU: RTX 4090 24 GiB

SYSTEM: Ubuntu 22.04.3 LTS

Run benchmark:

pytest unit_tests/bench_speed.py --benchmark-columns=mean,stddev --benchmark-sort=mean

Results:

------------- benchmark 'ncut-pytorch (CPU) vs sklearn': 8 tests ------------
Name (time in ms)                        Mean                StdDev          
-----------------------------------------------------------------------------
test_ncut_cpu_100_data_10_eig          2.5536 (1.0)          0.2782 (1.0)    
test_sklearn_100_data_10_eig           4.0913 (1.60)         1.6749 (6.02)   
test_ncut_cpu_300_data_10_eig          4.9034 (1.92)         1.6575 (5.96)   
test_sklearn_300_data_10_eig          10.1861 (3.99)         3.8870 (13.97)  
test_ncut_cpu_1000_data_10_eig        11.1968 (4.38)         1.7070 (6.13)   
test_ncut_cpu_3000_data_10_eig        38.6101 (15.12)        1.6379 (5.89)   
test_sklearn_1000_data_10_eig        193.5934 (75.81)        8.1933 (29.45)  
test_sklearn_3000_data_10_eig      1,246.4295 (488.11)   1,047.0191 (>1000.0)
-----------------------------------------------------------------------------

------------- benchmark 'ncut-pytorch (GPU) n_data': 5 tests -------------
Name (time in ms)                         Mean            StdDev          
--------------------------------------------------------------------------
test_ncut_gpu_100_data_10_eig           2.9564 (1.0)      0.1816 (1.0)    
test_ncut_gpu_1000_data_10_eig          4.6938 (1.59)     0.3933 (2.17)   
test_ncut_gpu_10000_data_10_eig        67.9607 (22.98)    4.0902 (22.52)  
test_ncut_gpu_100000_data_10_eig      396.9994 (134.29)   3.6202 (19.93)  
test_ncut_gpu_1000000_data_10_eig     798.4598 (270.08)   1.5704 (8.65)   
--------------------------------------------------------------------------

------------- benchmark 'ncut-pytorch (GPU) n_eig': 3 tests --------------
Name (time in ms)                         Mean            StdDev          
--------------------------------------------------------------------------
test_ncut_gpu_10000_data_10_eig        67.9607 (1.0)      4.0902 (10.76)  
test_ncut_gpu_10000_data_100_eig       74.0033 (1.09)     0.7856 (2.07)   
test_ncut_gpu_10000_data_1000_eig     179.8690 (2.65)     0.3801 (1.0)    
--------------------------------------------------------------------------

Run benchmark:

python unit_tests/bench_memory.py

Results:

ncut-pytorch.Ncut is $O(1)$ space complexity

+---------------+------------------------+
| Data Points   |   Peak GPU Memory (MB) |
+===============+========================+
| 1,000         |                   8.14 |
+---------------+------------------------+
| 10,000        |                   0.1  |
+---------------+------------------------+
| 100,000       |                   0.39 |
+---------------+------------------------+
| 1,000,000     |                   0.39 |
+---------------+------------------------+

Citation

@misc{yang2024alignedcutvisualconceptsdiscovery,
      title={AlignedCut: Visual Concepts Discovery on Brain-Guided Universal Feature Space}, 
      author={Huzheng Yang and James Gee and Jianbo Shi},
      year={2024},
      eprint={2406.18344},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2406.18344}, 
}