Upload 6 files

spectral/notebooks/experiment_2_manifold_structures.ipynb

@@ -1765,8 +1765,8 @@
     "        self.k = k\n",
     "        n_patches_h = input_size // patch_size\n",
     "        self.n_patches = n_patches_h * n_patches_h\n",
-    "        # k singular values + k*(k-1)/2 principal angles = features per patch\n",
-    "        self.features_per_patch = k + k * (k - 1) // 2 + k * 3\n",
+    "        # k singular values + k log-ratios + k*(k+1)/2 gram upper-tri\n",
+    "        self.features_per_patch = k + k + k * (k + 1) // 2\n",
     "        self.output_dim = self.n_patches * self.features_per_patch\n",
     "        print(f\"[GRASS] {self.n_patches} patches, k={k}, dim={self.output_dim}\")\n",
     "\n",
@@ -1781,13 +1781,14 @@
     "        U, S, Vh = torch.linalg.svd(patches_flat, full_matrices=False)\n",
     "        # Top-k singular values\n",
     "        sk = S[:, :self.k]  # (B*n_p, k)\n",
+    "        # Log singular value ratios (numerically informative)\n",
+    "        sv_ratios = torch.log(sk / (sk[:, -1:] + 1e-8) + 1e-8)  # (B*n_p, k)\n",
     "        # Principal angles via gram matrix of top-k vectors\n",
     "        Uk = U[:, :, :self.k]  # (B*n_p, ps*ps, k)\n",
     "        gram = torch.bmm(Uk.transpose(1, 2), Uk)  # (B*n_p, k, k)\n",
-    "        # Features: singular values + flattened gram upper triangle\n",
     "        triu_idx = torch.triu_indices(self.k, self.k)\n",
     "        gram_feat = gram[:, triu_idx[0], triu_idx[1]]  # (B*n_p, k*(k+1)/2)\n",
-    "        feats = torch.cat([sk, gram_feat], dim=-1)  # (B*n_p, feat_per_patch)\n",
+    "        feats = torch.cat([sk, sv_ratios, gram_feat], dim=-1)  # (B*n_p, feat_per_patch)\n",
     "        return feats.reshape(B, -1)  # (B, n_patches * feat_per_patch)\n",
     "\n",
     "front = GrassmannianFrontEnd(patch_size=8, k=3).to(device)\n",
@@ -1828,8 +1829,10 @@
     "        self.levels = levels\n",
     "        n_patches_h = input_size // patch_size\n",
     "        self.n_patches = n_patches_h * n_patches_h\n",
-    "        # For each level k: k singular values + projection norms\n",
-    "        self.features_per_patch = sum(k + k for k in levels)\n",
+    "        # For each level k: min(k, C) singular values + min(k, C) projection norms\n",
+    "        # C=3 for RGB, so levels > 3 are clamped\n",
+    "        max_sv = min(3, patch_size * patch_size)  # C vs ps*ps\n",
+    "        self.features_per_patch = sum(min(k, max_sv) * 2 for k in levels)\n",
     "        self.output_dim = self.n_patches * self.features_per_patch\n",
     "        print(f\"[FLAG] {self.n_patches} patches, levels={levels}, dim={self.output_dim}\")\n",
     "\n",