File size: 2,820 Bytes
a80391b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Fix: constraint_program — convert all numpy constants to torch tensors
# And CS-SDEdit gradient computation

import torch
import torch.nn.functional as F
import numpy as np

def constraint_program(x):
    """Compute constraint violation C(x).
    x: (batch, 25) tensor in log(1+x) space.
    Returns: (batch,) tensor of total violation (always differentiable).
    """
    b = x.shape[0]
    device = x.device
    
    # Property name to index map
    idx = {
        "compactness_2d": 12, "compactness_3d": 13, "cubeness": 19,
        "hemisphericality": 17, "area": 2, "volume": 5,
        "height_diff": 9, "num_vertices": 24, "bounding_box_width": 0,
        "bounding_box_length": 1, "aligned_bounding_box_height": 23,
        "num_floors": 10, "density": 14,
    }
    
    c2d  = x[:, idx["compactness_2d"]]
    c3d  = x[:, idx["compactness_3d"]]
    cub  = x[:, idx["cubeness"]]
    hemi = x[:, idx["hemisphericality"]]
    area_log = x[:, idx["area"]]
    volume_log = x[:, idx["volume"]]
    height = x[:, idx["height_diff"]]
    n_vert = x[:, idx["num_vertices"]]
    bb_w = x[:, idx["bounding_box_width"]]
    bb_l = x[:, idx["bounding_box_length"]]
    bb_h = x[:, idx["aligned_bounding_box_height"]]
    floors = x[:, idx["num_floors"]]
    density = x[:, idx["density"]]
    
    # All constants as torch tensors
    log2 = torch.tensor(np.log(2.0), device=device)
    log4 = torch.tensor(np.log(4.0), device=device)
    eps = torch.tensor(1e-6, device=device)
    tol = torch.tensor(2.0, device=device)
    
    v_list = []
    
    # H1-H4: Boundedness ∈ (0, 1] → in log space ∈ (0, log(2)]
    v_list.append(F.relu(-c2d + eps) + F.relu(c2d - log2))            # H1
    v_list.append(F.relu(-c3d + eps) + F.relu(c3d - log2))            # H2
    v_list.append(F.relu(-cub + eps) + F.relu(cub - log2))            # H3
    v_list.append(F.relu(-hemi + eps) + F.relu(hemi - log2))          # H4
    
    # H5-H9: Positivity
    v_list.append(F.relu(-area_log + eps))                             # H5
    v_list.append(F.relu(-volume_log + eps))                           # H6
    v_list.append(F.relu(-height + eps))                               # H7
    v_list.append(F.relu(-n_vert + log4))                              # H8: ≥4 vertices
    v_list.append(F.relu(-bb_w + eps))                                 # H9
    
    # S1-S4: Dimensional consistency (soft, tolerance=2.0)
    vol_est = area_log + height
    v_list.append(F.relu(torch.abs(volume_log - vol_est) - tol))      # S1
    
    v_list.append(F.relu(torch.abs(floors - height / log4) - tol))    # S2
    
    v_list.append(F.relu(density - log2))                              # S3
    
    v_list.append(F.relu(cub - log2))                                  # S4
    
    return sum(v_list)  # (batch,) scalar tensor per sample