ncut-pytorch 1.3.0

No project description provided

🌐Documentation | 🤗HuggingFace Demo

NCUT: Nyström Normalized Cut

Normalized Cut, aka. spectral clustering, is a graphical method to analyze data grouping in the affinity eigenvector space. It has been widely used for unsupervised segmentation in the 2000s.

Nyström Normalized Cut, is a new approximation algorithm developed for large-scale graph cuts, a large-graph of million nodes can be processed in under 10s (cpu) or 2s (gpu).

---

Installation & Quick Start

PyPI install, our package is based on PyTorch, please install PyTorch first

pip install ncut-pytorch

Minimal example on how to run NCUT:

import torch
from ncut_pytorch import NCUT, rgb_from_tsne_3d

model_features = torch.rand(20, 64, 64, 768)  # (B, H, W, C)

inp = model_features.reshape(-1, 768)  # flatten
eigvectors, eigvalues = NCUT(num_eig=100, device='cuda:0').fit_transform(inp)
tsne_x3d, tsne_rgb = rgb_from_tsne_3d(eigvectors, device='cuda:0')

eigvectors = eigvectors.reshape(20, 64, 64, 100)  # (B, H, W, num_eig)
tsne_rgb = tsne_rgb.reshape(20, 64, 64, 3)  # (B, H, W, 3)

Load Feature Extractor Model

We have implemented some backbone models, they can be used as feature extractors, Here is a list of available models:

from ncut_pytorch.backbone import list_models
print(list_models())
[
  'SAM2(sam2_hiera_t)', 'SAM2(sam2_hiera_s)', 'SAM2(sam2_hiera_b+)', 'SAM2(sam2_hiera_l)', 
  'SAM(sam_vit_b)', 'SAM(sam_vit_l)', 'SAM(sam_vit_h)', 'MobileSAM(TinyViT)', 
  'DiNOv2reg(dinov2_vits14_reg)', 'DiNOv2reg(dinov2_vitb14_reg)', 'DiNOv2reg(dinov2_vitl14_reg)', 'DiNOv2reg(dinov2_vitg14_reg)', 
  'DiNOv2(dinov2_vits14)', 'DiNOv2(dinov2_vitb14)', 'DiNOv2(dinov2_vitl14)', 'DiNOv2(dinov2_vitg14)', 
  'DiNO(dino_vits8_896)', 'DiNO(dino_vitb8_896)', 'DiNO(dino_vits8_672)', 'DiNO(dino_vitb8_672)', 'DiNO(dino_vits8_448)', 'DiNO(dino_vitb8_448)', 'DiNO(dino_vits16_448)', 'DiNO(dino_vitb16_448)',
  'CLIP(ViT-B-16/openai)', 'CLIP(ViT-L-14/openai)', 'CLIP(ViT-H-14/openai)', 'CLIP(ViT-B-16/laion2b_s34b_b88k)', 
  'CLIP(convnext_base_w_320/laion_aesthetic_s13b_b82k)', 'CLIP(convnext_large_d_320/laion2b_s29b_b131k_ft_soup)', 'CLIP(convnext_xxlarge/laion2b_s34b_b82k_augreg_soup)', 
  'CLIP(eva02_base_patch14_448/mim_in22k_ft_in1k)', "CLIP(eva02_large_patch14_448/mim_m38m_ft_in22k_in1k)",
  'MAE(vit_base)', 'MAE(vit_large)', 'MAE(vit_huge)', 
  'ImageNet(vit_base)'
]

Image model example:

import torch
from ncut_pytorch import NCUT, rgb_from_tsne_3d
from ncut_pytorch.backbone import load_model, extract_features

model = load_model(model_name="SAM(sam_vit_b)")
images = torch.rand(20, 3, 1024, 1024)
model_features = extract_features(images, model, node_type='attn', layer=6)
# model_features = model(images)['attn'][6]  # this also works

inp = model_features.reshape(-1, 768)  # flatten
eigvectors, eigvalues = NCUT(num_eig=100, device='cuda:0').fit_transform(inp)
tsne_x3d, tsne_rgb = rgb_from_tsne_3d(eigvectors, device='cuda:0')

eigvectors = eigvectors.reshape(20, 64, 64, 100)  # (B, H, W, num_eig)
tsne_rgb = tsne_rgb.reshape(20, 64, 64, 3)  # (B, H, W, 3)

Text model example:

import os
from ncut_pytorch import NCUT, rgb_from_tsne_3d
from ncut_pytorch.backbone_text import load_model

os.environ['HF_ACCESS_TOKEN'] = "your_huggingface_token"
llama = load_model("meta-llama/Meta-Llama-3.1-8B").cuda()
output_dict = llama("The quick white fox jumps over the lazy cat.")

model_features = output_dict['block'][31].squeeze(0)  # 32nd block output
token_texts = output_dict['token_texts']
eigvectors, eigvalues = NCUT(num_eig=5, device='cuda:0').fit_transform(model_features)
tsne_x3d, tsne_rgb = rgb_from_tsne_3d(eigvectors, device='cuda:0')
# eigvectors.shape[0] == tsne_rgb.shape[0] == len(token_texts)

---

paper in prep, Yang 2024

AlignedCut: Visual Concepts Discovery on Brain-Guided Universal Feature Space, Huzheng Yang, James Gee*, Jianbo Shi*,2024

Normalized Cuts and Image Segmentation, Jianbo Shi and Jitendra Malik, 2000