File size: 10,164 Bytes
dbbceb8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 | """
Morse Field Detection (MFD).
Project 768-dim patch tokens to K scalar potential fields via a linear projection.
Objects are peaks in the potential landscape. Bounding boxes from Hessian eigenvalues
via the Morse lemma. No box regression — boxes emerge from curvature.
Learned: W_psi (768 x K) projection + class prototypes
Fixed: Hessian (finite-difference kernels), peak detection, box extraction
"""
import json, os, sys, time, math
import torch
import torch.nn as nn
import torch.nn.functional as F
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, SCRIPT_DIR)
COCO_ROOT = os.environ.get("ARENA_COCO_ROOT")
VAL_CACHE = os.environ.get("ARENA_VAL_CACHE")
CACHE_DIR = os.environ.get("ARENA_CACHE_DIR")
DEVICE = "cuda"
RESOLUTION = 640
NUM_CLASSES = 80
STRIDE = 16
def cofiber_decompose(f, n_scales):
cofibers = []; residual = f
for _ in range(n_scales - 1):
omega = F.avg_pool2d(residual, 2)
sigma_omega = F.interpolate(omega, size=residual.shape[2:], mode="bilinear", align_corners=False)
cofibers.append(residual - sigma_omega); residual = omega
cofibers.append(residual); return cofibers
class MorseFieldDetector(nn.Module):
"""Morse theory detection head. Boxes from curvature, not regression."""
def __init__(self, feat_dim=768, n_fields=3, num_classes=80):
super().__init__()
# The only learned spatial component: project features to scalar fields
self.field_proj = nn.Linear(feat_dim, n_fields, bias=False)
# Class prototypes for classification at detected peaks
self.cls_prototypes = nn.Parameter(torch.randn(num_classes, feat_dim) * 0.01)
self.cls_bias = nn.Parameter(torch.zeros(num_classes))
# Fixed finite-difference Hessian kernels (Sobel-style)
# d²f/dx²
hxx = torch.tensor([[0, 0, 0], [1, -2, 1], [0, 0, 0]], dtype=torch.float32).unsqueeze(0).unsqueeze(0)
# d²f/dy²
hyy = torch.tensor([[0, 1, 0], [0, -2, 0], [0, 1, 0]], dtype=torch.float32).unsqueeze(0).unsqueeze(0)
# d²f/dxdy
hxy = torch.tensor([[1, 0, -1], [0, 0, 0], [-1, 0, 1]], dtype=torch.float32).unsqueeze(0).unsqueeze(0) * 0.25
self.register_buffer("hxx_kernel", hxx)
self.register_buffer("hyy_kernel", hyy)
self.register_buffer("hxy_kernel", hxy)
def compute_hessian(self, field):
"""Compute Hessian components of a scalar field. field: (B, 1, H, W)."""
fxx = F.conv2d(field, self.hxx_kernel, padding=1)
fyy = F.conv2d(field, self.hyy_kernel, padding=1)
fxy = F.conv2d(field, self.hxy_kernel, padding=1)
return fxx, fyy, fxy
def forward_detect(self, spatial, scale=1.0):
"""Run detection on one image. Returns boxes, scores, classes."""
B, C, H, W = spatial.shape
assert B == 1
f = F.layer_norm(spatial.permute(0, 2, 3, 1).reshape(-1, C), [C]) # (H*W, 768)
# Project to scalar potential fields
fields = self.field_proj(f).reshape(1, H, W, -1).permute(0, 3, 1, 2) # (1, K, H, W)
# Classification scores at every location
cls_scores = (f @ self.cls_prototypes.T + self.cls_bias).reshape(1, H, W, -1).permute(0, 3, 1, 2) # (1, 80, H, W)
all_boxes = []
all_scores = []
all_classes = []
n_fields = fields.shape[1]
for k in range(n_fields):
field = fields[:, k:k+1] # (1, 1, H, W)
# Compute Hessian
fxx, fyy, fxy = self.compute_hessian(field)
# Determinant and trace of Hessian
det_H = fxx * fyy - fxy * fxy # (1, 1, H, W)
tr_H = fxx + fyy
# Objectness: peak = det(H) > 0 AND tr(H) < 0 (local maximum)
objectness = torch.sigmoid(det_H * 10) * torch.sigmoid(-tr_H * 10)
objectness = objectness.squeeze(0).squeeze(0) # (H, W)
# Field values at each location
psi = field.squeeze(0).squeeze(0) # (H, W)
# Find peaks: local maxima of objectness
# Use max_pool to find locations that are local maxima
obj_padded = objectness.unsqueeze(0).unsqueeze(0)
local_max = F.max_pool2d(obj_padded, 3, stride=1, padding=1).squeeze()
is_peak = (objectness == local_max) & (objectness > 0.3)
peak_locs = is_peak.nonzero(as_tuple=False) # (M, 2) — row, col
for pi in range(min(len(peak_locs), 50)):
r, c = peak_locs[pi]
ri, ci = r.item(), c.item()
# Hessian eigenvalues at this peak
h11 = fxx[0, 0, ri, ci].item()
h22 = fyy[0, 0, ri, ci].item()
h12 = fxy[0, 0, ri, ci].item()
psi_val = max(psi[ri, ci].item(), 0.01)
# Eigenvalues of -H (should be positive at a maximum)
neg_tr = -(h11 + h22)
discriminant = (h11 - h22) ** 2 + 4 * h12 ** 2
sqrt_disc = math.sqrt(max(discriminant, 0))
lam1 = (neg_tr + sqrt_disc) / 2
lam2 = (neg_tr - sqrt_disc) / 2
if lam1 <= 0 or lam2 <= 0:
continue
# Morse lemma: box dimensions from curvature
# w = 2 * sqrt(psi / lam2), h = 2 * sqrt(psi / lam1)
box_w = 2 * math.sqrt(psi_val / lam2) * STRIDE
box_h = 2 * math.sqrt(psi_val / lam1) * STRIDE
# Box center in pixel coords
cx = (ci + 0.5) * STRIDE
cy = (ri + 0.5) * STRIDE
x1 = (cx - box_w / 2) / scale
y1 = (cy - box_h / 2) / scale
w = box_w / scale
h = box_h / scale
if w < 1 or h < 1:
continue
# Classification at this location
cls = cls_scores[0, :, ri, ci]
cls_score, cls_idx = cls.sigmoid().max(0)
score = objectness[ri, ci].item() * cls_score.item()
if score < 0.01:
continue
all_boxes.append([x1, y1, w, h])
all_scores.append(score)
all_classes.append(cls_idx.item())
return all_boxes, all_scores, all_classes
def main():
print("=" * 60)
print("Morse Field Detection (MFD)")
print("=" * 60, flush=True)
head = MorseFieldDetector().to(DEVICE)
n_params = sum(p.numel() for p in head.parameters())
print(f" {n_params:,} params ({head.field_proj.weight.numel()} projection + "
f"{head.cls_prototypes.numel() + head.cls_bias.numel()} classification)")
# Initialize class prototypes from analytical solution
analytical_path = os.path.join(SCRIPT_DIR, "heads", "cofiber_threshold",
"analytical_70k", "analytical_head_70k.pth")
if os.path.isfile(analytical_path):
ckpt = torch.load(analytical_path, map_location=DEVICE, weights_only=False)
head.cls_prototypes.data = ckpt["cls_weight"].to(DEVICE)
head.cls_bias.data = ckpt["cls_bias"].to(DEVICE)
print(" Loaded analytical class prototypes")
# Initialize field projection from PCA of features
# Use first 3 PCA components — the "smooth scalar fields" the paper shows
print(" Computing PCA for field projection init...", flush=True)
manifest = json.load(open(os.path.join(CACHE_DIR, "manifest.json")))
shard = torch.load(os.path.join(CACHE_DIR, "shard_0000.pt"), map_location="cpu", weights_only=False)
sample_features = []
for item in shard[:100]:
sp = item["spatial"].unsqueeze(0).float()
f = F.layer_norm(sp.permute(0, 2, 3, 1).reshape(-1, 768), [768])
sample_features.append(f)
sample_f = torch.cat(sample_features)
_, _, Vh = torch.linalg.svd(sample_f[:10000] - sample_f[:10000].mean(0), full_matrices=False)
head.field_proj.weight.data = Vh[:3].to(DEVICE)
print(f" PCA-initialized field projection")
del shard, sample_features, sample_f
# Eval (no training — pure derived detection)
val = torch.load(VAL_CACHE, map_location="cpu", weights_only=False)
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
coco_gt = COCO(os.path.join(COCO_ROOT, "annotations", "instances_val2017.json"))
cat_ids = sorted(coco_gt.getCatIds())
idx_to_cat = {i: c for i, c in enumerate(cat_ids)}
max_images = 1000
all_results = []
t0 = time.time()
head.eval()
with torch.no_grad():
for idx in range(min(max_images, len(val))):
item = val[idx]
spatial = item["spatial"].unsqueeze(0).float().to(DEVICE)
img_id = int(item["img_id"])
scale = item["scale"]
boxes, scores, classes = head.forward_detect(spatial, scale)
for b, s, c in zip(boxes, scores, classes):
all_results.append({
"image_id": img_id,
"category_id": idx_to_cat[c],
"bbox": b,
"score": s,
})
if (idx + 1) % 200 == 0:
elapsed = time.time() - t0
print(f" {idx+1}/{max_images} ({elapsed:.0f}s, {len(all_results)} dets)", flush=True)
print(f"\n{len(all_results)} total detections ({time.time()-t0:.0f}s)")
if all_results:
coco_dt = coco_gt.loadRes(all_results)
coco_eval = COCOeval(coco_gt, coco_dt, "bbox")
coco_eval.params.imgIds = sorted(coco_gt.getImgIds())[:min(max_images, len(val))]
coco_eval.evaluate(); coco_eval.accumulate(); coco_eval.summarize()
print(f"\nMFD (zero training): mAP@[.5:.95]={coco_eval.stats[0]:.4f} "
f"mAP@.50={coco_eval.stats[1]:.4f} mAP@.75={coco_eval.stats[2]:.4f}")
else:
print("No detections")
print(f"\n Field projection: {head.field_proj.weight.numel()} params (PCA-initialized)")
print(f" Classification: {head.cls_prototypes.numel()} params (analytical)")
print(f" Hessian + peak detection + Morse box extraction: 0 params")
if __name__ == "__main__":
main()
|