sid-composite-data-augmentation/evaluate.py at main · efthimia145/sid-composite-data-augmentation · 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
import os
import torch
import torch.nn as nn
import numpy as np
from torchvision import models
from sklearn.metrics import average_precision_score, accuracy_score
from torch.utils.tensorboard import SummaryWriter

from data.data import create_dataset
from eval_config import *
from options.test_options import TestOptions


def evaluate_model(model_path, evaluation_dataloader):
    """
    Evaluate a trained deep fake detection model on a test dataset.

    Args:
        model_path (str): Path to the saved model checkpoint
        evaluation_dataloader: PyTorch DataLoader containing test data

    Returns:
        dict: Dictionary containing evaluation metrics:
            - accuracy: Overall model accuracy
            - avg_precision: Average precision score
            - r_accuracy: Accuracy on real images only
            - f_accuracy: Accuracy on fake images only
    """
    m = nn.Sigmoid()
    model = models.resnet50(num_classes=1)
    state_dict = torch.load(model_path, map_location='cpu')
    model.load_state_dict(state_dict['model_state_dict'])
    model.cuda()
    model.eval()

    with torch.no_grad():
        y_pred, y_real = [], []

        for image, label in evaluation_dataloader:
            input_tensor = image.cuda()
            prediction_tensor = model(input_tensor)
            y_pred.extend(m(prediction_tensor).flatten().tolist())
            y_real.extend(label.flatten().tolist())

    y_real, y_pred = np.array(y_real).reshape(-1, 1), np.array(y_pred).reshape(-1, 1)

    model_accuracy = accuracy_score(y_real, y_pred > 0.5)
    avg_precision = average_precision_score(y_real, y_pred)
    reals_accurary = accuracy_score(y_real[y_real == 0], y_pred[y_real == 0] > 0.5)
    fakes_accurary = accuracy_score(y_real[y_real == 1], y_pred[y_real == 1] > 0.5)

    evaluation_result_metrics = {'accuracy': model_accuracy, 'avg_precision': avg_precision, \
                                'r_accuracy': reals_accurary, 'f_accuracy': fakes_accurary}

    return evaluation_result_metrics


def find_evaluation_classes(root_path):
    """
    Discover class subdirectories in evaluation datasets.

    Args:
        root_path (str): Root path containing evaluation datasets

    Returns:
        list: List of class directories for each evaluation dataset
    """
    classes = []

    for eval_dataset, multiclass_dataset in evaluation_datasets_multiclass.items():
        dataset_path = f'{root_path}/{eval_dataset}'
        classes.append(os.listdir(dataset_path) if multiclass_dataset else '')

    return classes

if __name__=="__main__":

    opt = TestOptions().parse()

    test_writer = SummaryWriter(f'runs/test/{opt.name}')

    data_directory = "dataset/test"
    model_path = f'{opt.checkpoints_dir}/{opt.model_path}/best_model.pth'
    eval_results = {}
    # classes = find_evaluation_classes(f"{data_directory}/{dataset}")

    evaluation_datasets_multiclass = evaluation_datasets_multiclass_local if opt.runner == 'local' else evaluation_datasets_multiclass_hpc

    for eval_dataset, multiclass_dataset in evaluation_datasets_multiclass.items():
        dataset_path = f'{data_directory}/{eval_dataset}'
        classes = os.listdir(dataset_path) if multiclass_dataset else ['']

        dataloader, dataset_size = create_dataset(dataset_path, opt, classes)

        eval_results[eval_dataset] = evaluate_model(model_path, dataloader)


    if opt.runner == 'local':
        table = f"""
                    | Eval dataset   | Accuracy  | AP |
                    |----------------|-----------|-----------|
                    | ProGAN | {round(eval_results['progan']['accuracy'],2)} |  {round(eval_results['progan']['avg_precision'],2)} |
                    | BigGAN  | {round(eval_results['biggan']['accuracy'],2)} |  {round(eval_results['biggan']['avg_precision'],2)} |
                    | CRN  | {round(eval_results['crn']['accuracy'],2)} |  {round(eval_results['crn']['avg_precision'],2)} |
                """
    else:
        table = f"""
                    | Eval dataset   | Accuracy  | AP |
                    |----------------|-----------|-----------|
                    | ProGAN | {round(eval_results['progan']['accuracy'],2)} |  {round(eval_results['progan']['avg_precision'],2)} |
                    | StyleGAN  | {round(eval_results['stylegan']['accuracy'],2)} |  {round(eval_results['stylegan']['avg_precision'],2)} |
                    | BigGAN  | {round(eval_results['biggan']['accuracy'],2)} |  {round(eval_results['biggan']['avg_precision'],2)} |
                    | CycleGAN  | {round(eval_results['cyclegan']['accuracy'],2)} |  {round(eval_results['cyclegan']['avg_precision'],2)} |
                    | StarGAN  | {round(eval_results['stargan']['accuracy'],2)} |  {round(eval_results['stargan']['avg_precision'],2)} |
                    | GauGAN  | {round(eval_results['gaugan']['accuracy'],2)} |  {round(eval_results['gaugan']['avg_precision'],2)} |
                    | CRN  | {round(eval_results['crn']['accuracy'],2)} |  {round(eval_results['crn']['avg_precision'],2)} |
                    | IMLE  | {round(eval_results['imle']['accuracy'],2)} |  {round(eval_results['imle']['avg_precision'],2)} |
                    | SITD  | {round(eval_results['seeingdark']['accuracy'],2)} |  {round(eval_results['seeingdark']['avg_precision'],2)} |
                    | SAN  | {round(eval_results['san']['accuracy'],2)} |  {round(eval_results['san']['avg_precision'],2)} |
                    | DeepFake  | {round(eval_results['deepfake']['accuracy'],2)} |  {round(eval_results['deepfake']['avg_precision'],2)} |
                    | StyleGAN2  | {round(eval_results['stylegan2']['accuracy'],2)} |  {round(eval_results['stylegan2']['avg_precision'],2)} |
                    | WFIR  | {round(eval_results['whichfaceisreal']['accuracy'],2)} |  {round(eval_results['whichfaceisreal']['avg_precision'],2)} |
                """

    table = '\n'.join(l.strip() for l in table.splitlines())

    test_writer.add_text('test', table)