OSDA-ST/benchmarktmp.py at master · yeedrag/OSDA-ST · GitHub

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
import numpy as np
import torch
import torch.nn.functional as F
from mmseg.core.evaluation.metrics import intersect_and_union
from typing import Tuple
import timeit

def get_optimal_thresholds(pred: torch.tensor, label: torch.tensor, num_classes: int) -> Tuple[float, dict]:
    '''
    Find the optimal threshold for pred that maximizes the H-score w.r.t label.
    '''
    device = pred.device
    thresholds = torch.full((num_classes,), 0.5, device=device)
    def getIoU(pred: torch.tensor, label: torch.tensor) -> torch.tensor:
        intersection = torch.zeros(num_classes)
        union = torch.zeros(num_classes)

        for cls in range(num_classes):
            pred_mask = (pred == cls)
            true_mask = (label == cls)
            intersection[cls] = torch.logical_and(pred_mask, true_mask).sum().float()
            union[cls] = torch.logical_or(pred_mask, true_mask).sum().float()

        iou = intersection / (union + 1e-10)  # adding epsilon to avoid division by zero
        return iou
    def get_H_score_fast(thresholds: torch.tensor) -> float:
        '''
        Calculate H score using the given thresholds.
        '''
        pseudo_prob, pseudo_label = torch.max(pred, dim=0)
        for i in range(num_classes):
            label_class_mask = (pseudo_label == i) & (pseudo_prob < thresholds[i])
            pseudo_label[label_class_mask] = num_classes  # Unknown label
        iou = getIoU(pseudo_label, label)
        miou_known = torch.mean(iou[:-1])  # Exclude unknown class
        H_score = (2 * miou_known * iou[-1]) / (miou_known + iou[-1])
        return H_score.item()

    best_thresholds = thresholds.clone()
    best_H_score = get_H_score_fast(thresholds)

    # Optimizing thresholds using binary search for each class
    for i in range(19):
        low, high = 0.0, 1.0
        while high - low > 1e-3:
            mid1 = low + (high - low) / 3
            mid2 = high - (high - low) / 3
            thresholds[i] = mid1
            H_score1 = get_H_score_fast(thresholds)
            thresholds[i] = mid2
            H_score2 = get_H_score_fast(thresholds)
            if H_score1 > H_score2:
                high = mid2
            else:
                low = mid1

        thresholds[i] = (low + high) / 2
    final_H_score = get_H_score_fast(thresholds)

    return final_H_score, thresholds.cpu().numpy()

# Example usage
torch.manual_seed(42)
A_512_512 = torch.rand(20, 512, 512)
B_512_512 = torch.randint(0, 20, (512, 512))
A_512_512 = A_512_512.cuda()
B_512_512 = B_512_512.cuda()
start_time = timeit.default_timer()
h, thres = get_optimal_thresholds(A_512_512, B_512_512, 20)
print(f"Time: {timeit.default_timer() - start_time}")
print(f"H-Score: {h}")


import torch
import torch.multiprocessing as mp
import timeit

def get_optimal_thresholds(pred: torch.tensor, label: torch.tensor, num_classes: int):
    device = pred.device
    thresholds = torch.full((num_classes,), 0.5, device=device)

    def getIoU(pred: torch.tensor, label: torch.tensor) -> torch.tensor:
        intersection = torch.zeros(num_classes, device=device)
        union = torch.zeros(num_classes, device=device)
        for cls in range(num_classes):
            pred_mask = (pred == cls)
            true_mask = (label == cls)
            intersection[cls] = torch.logical_and(pred_mask, true_mask).sum().float()
            union[cls] = torch.logical_or(pred_mask, true_mask).sum().float()
        iou = intersection / (union + 1e-10)
        return iou

    def get_H_score_fast(thresholds: torch.tensor) -> float:
        pseudo_prob, pseudo_label = torch.max(pred, dim=0)
        unknown_mask = pseudo_prob < thresholds[pseudo_label]
        pseudo_label[unknown_mask] = num_classes
        iou = getIoU(pseudo_label, label)
        miou_known = torch.mean(iou[:-1])
        H_score = (2 * miou_known * iou[-1]) / (miou_known + iou[-1])
        return H_score.item()

    best_thresholds = thresholds.clone()
    best_H_score = get_H_score_fast(thresholds)

    # Parallel optimization for all classes
    low = torch.zeros(num_classes - 1, device=device)
    high = torch.ones(num_classes - 1, device=device)

    while torch.max(high - low) > 1e-3:
        mid1 = low + (high - low) / 3
        mid2 = high - (high - low) / 3

        H_scores1 = torch.zeros(num_classes - 1, device=device)
        H_scores2 = torch.zeros(num_classes - 1, device=device)

        for i in range(num_classes - 1):
            thresholds[i] = mid1[i]
            H_scores1[i] = get_H_score_fast(thresholds)
            thresholds[i] = mid2[i]
            H_scores2[i] = get_H_score_fast(thresholds)
            thresholds[i] = best_thresholds[i]  # Reset to best known threshold

        mask = H_scores1 > H_scores2
        high[mask] = mid2[mask]
        low[~mask] = mid1[~mask]

    thresholds[:-1] = (low + high) / 2
    final_H_score = get_H_score_fast(thresholds)

    return final_H_score, thresholds.cpu().numpy()

# Example usage
torch.manual_seed(42)
A_512_512 = torch.rand(20, 512, 512)
B_512_512 = torch.randint(0, 20, (512, 512))
A_512_512 = A_512_512.cuda()
B_512_512 = B_512_512.cuda()

start_time = timeit.default_timer()
h, thres = get_optimal_thresholds(A_512_512, B_512_512, 20)
print(f"Time: {timeit.default_timer() - start_time}")
print(f"H-Score: {h}")
print(f"Threshold: {thres}")