nystrom-ncut 0.1.2__tar.gz → 0.1.3__tar.gz

{nystrom_ncut-0.1.2/src/nystrom_ncut.egg-info → nystrom_ncut-0.1.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: nystrom_ncut
-Version: 0.1.2
+Version: 0.1.3
 Summary: Normalized Cut and Nyström Approximation
 Author-email: Huzheng Yang <huze.yann@gmail.com>, Wentinn Liao <wentinn.liao@gmail.com>
 Project-URL: Documentation, https://github.com/JophiArcana/Nystrom-NCUT/

{nystrom_ncut-0.1.2 → nystrom_ncut-0.1.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "nystrom_ncut"
-version = "0.1.2"
+version = "0.1.3"
 authors = [
     { name = "Huzheng Yang", email = "huze.yann@gmail.com" },
     { name = "Wentinn Liao", email = "wentinn.liao@gmail.com" },

nystrom_ncut-0.1.3/requirements.txt

@@ -0,0 +1,9 @@
+einops
+scikit-learn==1.6.1
+umap-learn
+pycolormap-2d
+tqdm
+torch==2.5.1
+pytorch3d==0.7.8
+torchdata==0.10.1
+torchvision==0.20.1

{nystrom_ncut-0.1.2 → nystrom_ncut-0.1.3}/src/nystrom_ncut/__init__.py

@@ -8,6 +8,7 @@ from .distance_utils import (
 )
 from .sampling_utils import (
     SampleConfig,
+    subsample_features,
 )
 from .visualize_utils import (
     extrapolate_knn,

{nystrom_ncut-0.1.2 → nystrom_ncut-0.1.3}/src/nystrom_ncut/distance_utils.py

@@ -41,7 +41,9 @@ def distance_from_features(
         D = torch.cdist(features, features_B, p=2) ** 2
 
         # Outlier-robust scale invariance using quantiles to estimate standard deviation
-        stds = torch.quantile(features, q=torch.tensor((0.158655, 0.841345), device=features.device), dim=0)
+        c = 2.0
+        p = torch.erf(torch.tensor((-c, c), device=features.device) * (2 ** -0.5))
+        stds = torch.quantile(features, q=(p + 1) / 2, dim=0)
         stds = (stds[1] - stds[0]) / 2
         D = D / (2 * torch.linalg.norm(stds) ** 2)
     else:

{nystrom_ncut-0.1.2 → nystrom_ncut-0.1.3}/src/nystrom_ncut/nystrom/nystrom_utils.py

@@ -12,7 +12,7 @@ from ..distance_utils import (
 )
 from ..sampling_utils import (
     SampleConfig,
-    run_subgraph_sampling,
+    subsample_features,
 )
 
 
@@ -180,7 +180,7 @@ class OnlineNystromSubsampleFit(OnlineNystrom):
         if precomputed_sampled_indices is not None:
             self.anchor_indices = precomputed_sampled_indices
         else:
-            self.anchor_indices = run_subgraph_sampling(
+            self.anchor_indices = subsample_features(
                 features=features,
                 disttype=self.distance,
                 config=self.sample_config,

{nystrom_ncut-0.1.2 → nystrom_ncut-0.1.3}/src/nystrom_ncut/sampling_utils.py

@@ -3,11 +3,10 @@ from dataclasses import dataclass
 from typing import Literal
 
 import torch
-from dgl.geometry import farthest_point_sampler
+from pytorch3d.ops import sample_farthest_points
 
 from .distance_utils import (
     DistanceOptions,
-    affinity_from_features,
     to_euclidean,
 )
 
@@ -25,7 +24,7 @@ class SampleConfig:
 
 
 @torch.no_grad()
-def run_subgraph_sampling(
+def subsample_features(
     features: torch.Tensor,
     disttype: DistanceOptions,
     config: SampleConfig,
@@ -57,25 +56,15 @@ def run_subgraph_sampling(
 
         elif config.method == "fps_recursive":
             features = to_euclidean(features, disttype)
-            sampled_indices = run_subgraph_sampling(
+            sampled_indices = subsample_features(
                 features=features,
                 disttype=disttype,
                 config=SampleConfig(method="fps", num_sample=config.num_sample, fps_dim=config.fps_dim)
             )
-
             nc = config._ncut_obj
-
-            A = affinity_from_features(features, affinity_focal_gamma=nc.kernel.affinity_focal_gamma, distance=nc.kernel.distance)
-            R = torch.diag(torch.sum(A, dim=-1) ** -0.5)
-            L = R @ A @ R
-
             for _ in range(config.n_iter):
                 fps_features, eigenvalues = nc.fit_transform(features, precomputed_sampled_indices=sampled_indices)
 
-                _L = fps_features @ torch.diag(eigenvalues) @ fps_features.mT
-                RE = torch.abs(_L / L - 1)
-
-                print(f"Iteration {_} --- max: {RE.max().item()}, mean: {RE.mean().item()}, min: {RE.min().item()}")
                 fps_features = to_euclidean(fps_features[:, :config.fps_dim], "cosine")
                 sampled_indices = torch.sort(fpsample(fps_features, config)).values
         else:
@@ -93,4 +82,4 @@ def fpsample(
         U, S, V = torch.pca_lowrank(features, q=config.fps_dim)
         features = U * S
 
-    return farthest_point_sampler(features[None], config.num_sample)[0]
+    return sample_farthest_points(features[None], K=config.num_sample)[1][0]

{nystrom_ncut-0.1.2 → nystrom_ncut-0.1.3}/src/nystrom_ncut/visualize_utils.py

@@ -19,7 +19,7 @@ from .distance_utils import (
 )
 from .sampling_utils import (
     SampleConfig,
-    run_subgraph_sampling,
+    subsample_features,
 )
 
 
@@ -120,7 +120,7 @@ def extrapolate_knn_with_subsampling(
     device = full_output.device if device is None else device
 
     # sample subgraph
-    anchor_indices = run_subgraph_sampling(
+    anchor_indices = subsample_features(
         features=full_features,
         disttype=distance,
         config=sample_config,
@@ -160,7 +160,7 @@ def _rgb_with_dimensionality_reduction(
 ) -> torch.Tensor:
 
     if True:
-        _subgraph_indices = run_subgraph_sampling(
+        _subgraph_indices = subsample_features(
             features=features,
             disttype=disttype,
             config=SampleConfig(method="fps"),
@@ -172,7 +172,7 @@ def _rgb_with_dimensionality_reduction(
             distance=disttype,
         )
 
-    subgraph_indices = run_subgraph_sampling(
+    subgraph_indices = subsample_features(
         features=features,
         disttype=disttype,
         config=SampleConfig(method="fps", num_sample=num_sample),

{nystrom_ncut-0.1.2 → nystrom_ncut-0.1.3/src/nystrom_ncut.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: nystrom_ncut
-Version: 0.1.2
+Version: 0.1.3
 Summary: Normalized Cut and Nyström Approximation
 Author-email: Huzheng Yang <huze.yann@gmail.com>, Wentinn Liao <wentinn.liao@gmail.com>
 Project-URL: Documentation, https://github.com/JophiArcana/Nystrom-NCUT/

{nystrom_ncut-0.1.2 → nystrom_ncut-0.1.3}/tests/test.py

@@ -1,8 +1,9 @@
 import numpy as np
 import torch
+import torch.nn as nn
 from matplotlib import pyplot as plt
 
-from src.nystrom_ncut import NCut, affinity_from_features, SampleConfig
+from src.nystrom_ncut import NCut, affinity_from_features, SampleConfig, subsample_features
 
 # from ncut_pytorch.src import rgb_from_umap_sphere
 # from ncut_pytorch.src.new_ncut_pytorch import NewNCUT
@@ -48,12 +49,13 @@ if __name__ == "__main__":
     torch.manual_seed(1212)
     np.random.seed(1212)
 
-    n, d = 1000, 2
-    num_sample = 900
+
+    n, d = 10000, 2
+    num_sample = 30
 
     M = torch.randn((n, d))
-    M[:int(0.9 * n)] += 2
-    M[int(0.9 * n):] -= 2
+    M[:int(0.9 * n)] += 3
+    M[int(0.9 * n):] -= 3
 
     distance = "rbf"
 
@@ -61,6 +63,58 @@ if __name__ == "__main__":
     R = torch.diag(torch.sum(A, dim=-1) ** -0.5)
     L = R @ A @ R
 
+    n_components = 30   # num_sample
+    eig_solver = "svd_lowrank"
+
+    nc = NCut(
+        n_components=n_components,
+        sample_config=SampleConfig(method="random", num_sample=num_sample),
+        distance=distance,
+        eig_solver=eig_solver,
+    )
+
+    torch.seed()
+    indices = subsample_features(M, disttype=distance, config=SampleConfig(method="fps", num_sample=num_sample))
+    samples = nn.Parameter(M[indices])
+
+    optimizer = torch.optim.AdamW((samples,), lr=1e-1)
+
+    output_dir = "../output/anchor_features_descent"
+    relative_losses, absolute_losses = [], []
+    for it in range(1000):
+        if it % 10 == 0:
+            plt.scatter(*M.mT, color="red")
+            plt.scatter(*samples.mT.detach(), color="black")
+            plt.title(f"Iteration {it}")
+
+            plt.savefig(f"{output_dir}/iteration{it}.png")
+            plt.show()
+
+        all_points = torch.cat((samples, M), dim=0)
+
+        X, eigs = nc.fit_transform(all_points, precomputed_sampled_indices=torch.arange(num_sample))
+        X = X[num_sample:]
+
+        _L = X @ torch.diag(eigs) @ X.mT
+
+        relative_loss = torch.linalg.norm(_L / L - 1) ** 2
+        with torch.no_grad():
+            absolute_loss = torch.linalg.norm(_L - L) ** 2
+
+        optimizer.zero_grad()
+        relative_loss.backward()
+        optimizer.step()
+
+        print(f"Relative: {relative_loss.item()}, Absolute: {absolute_loss.item()}")
+        relative_losses.append(relative_loss.item())
+        absolute_losses.append(absolute_loss.item())
+
+    torch.save(torch.tensor(relative_losses), f"{output_dir}/relative_losses.pt")
+    torch.save(torch.tensor(absolute_losses), f"{output_dir}/absolute_losses.pt")
+
+    raise Exception()
+
+
     # C = L[num_sample:, num_sample:]
     #
     # _A = L[:num_sample, :num_sample]
@@ -82,20 +136,20 @@ if __name__ == "__main__":
 
     max_rel = []
     for _ in range(1):
-        nc0 = NCut(
+        nc = NCut(
             n_components=n_components,
-            # sample_config=SampleConfig(method="random", n_iter=10),
-            sample_config=SampleConfig(method="fps_recursive", num_sample=num_sample, n_iter=101),
+            # sample_config=SampleConfig(method="random", num_sample=num_sample),
+            sample_config=SampleConfig(method="fps_recursive", num_sample=num_sample, n_iter=10),
             distance=distance,
             eig_solver=eig_solver,
         )
-        X0, eigs0 = nc0.fit_transform(M)
+        X, eigs = nc.fit_transform(M)
 
-        re0 = rel_error(X0, eigs0)
-        max_rel.append(re0.max().item())
+        re = rel_error(X, eigs)
+        max_rel.append(re.max().item())
 
         if _ % 100 == 0:
-            print_re(re0)
+            print_re(re)
         # print_re(re0)
 
     # plt.hist(max_rel, bins=30)

nystrom_ncut-0.1.2/requirements.txt

@@ -1,7 +0,0 @@
-einops
-scikit-learn
-umap-learn
-pycolormap-2d
-tqdm
-torch==2.1.0
-dgl==2.2.0