Skip to content

class FID

Source code in aigve/metrics/video_quality_assessment/distribution_based/fid.py 
 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
class FIDScore:
    def __init__(self, model_name='inception_v3', input_shape=(299, 299, 3), pooling='avg'):
        """
        Initialize the FIDScore evaluator.

        Parameters:
        model_name (str): The model to use for feature extraction (default: 'inception_v3').
        input_shape (tuple): Input shape of the images (default: (299, 299, 3)).
        pooling (str): Pooling type to use ('avg' or 'max') (default: 'avg').
        """
        if model_name == 'inception_v3':
            self.model = InceptionV3(include_top=False, pooling=pooling, input_shape=input_shape)
        else:
            raise ValueError("Only 'inception_v3' is currently supported.")

    def preprocess_images(self, images):
        """
        Preprocess the images for the InceptionV3 model.

        Parameters:
        images (numpy array): Input images.

        Returns:
        numpy array: Preprocessed images.
        """
        return preprocess_input(images)

    def calculate_statistics(self, images):
        """
        Calculate the feature statistics (mean and covariance) for a set of images.

        Parameters:
        images (numpy array): Preprocessed images.

        Returns:
        tuple: Mean and covariance of the features.
        """
        features = self.model.predict(images)
        mu = features.mean(axis=0)
        sigma = np.cov(features, rowvar=False)
        return mu, sigma

    def calculate_fid(self, images1, images2):
        """
        Calculate the FID score between two sets of images.

        Parameters:
        images1 (numpy array): First set of images.
        images2 (numpy array): Second set of images.

        Returns:
        float: The FID score.
        """
        # Preprocess images
        images1 = self.preprocess_images(images1)
        images2 = self.preprocess_images(images2)

        # Calculate statistics
        mu1, sigma1 = self.calculate_statistics(images1)
        mu2, sigma2 = self.calculate_statistics(images2)

        # Compute FID score
        ssdiff = np.sum((mu1 - mu2) ** 2.0)
        covmean = sqrtm(sigma1.dot(sigma2))

        # Check and correct for imaginary numbers
        if np.iscomplexobj(covmean):
            covmean = covmean.real

        fid = ssdiff + np.trace(sigma1 + sigma2 - 2.0 * covmean)
        return fid

__init__(model_name='inception_v3', input_shape=(299, 299, 3), pooling='avg')

Initialize the FIDScore evaluator.

Parameters: model_name (str): The model to use for feature extraction (default: 'inception_v3'). input_shape (tuple): Input shape of the images (default: (299, 299, 3)). pooling (str): Pooling type to use ('avg' or 'max') (default: 'avg').

Source code in aigve/metrics/video_quality_assessment/distribution_based/fid.py 
 8
 9
10
11
12
13
14
15
16
17
18
19
20
def __init__(self, model_name='inception_v3', input_shape=(299, 299, 3), pooling='avg'):
    """
    Initialize the FIDScore evaluator.

    Parameters:
    model_name (str): The model to use for feature extraction (default: 'inception_v3').
    input_shape (tuple): Input shape of the images (default: (299, 299, 3)).
    pooling (str): Pooling type to use ('avg' or 'max') (default: 'avg').
    """
    if model_name == 'inception_v3':
        self.model = InceptionV3(include_top=False, pooling=pooling, input_shape=input_shape)
    else:
        raise ValueError("Only 'inception_v3' is currently supported.")

calculate_fid(images1, images2)

Calculate the FID score between two sets of images.

Parameters: images1 (numpy array): First set of images. images2 (numpy array): Second set of images.

Returns: float: The FID score.

Source code in aigve/metrics/video_quality_assessment/distribution_based/fid.py 
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
def calculate_fid(self, images1, images2):
    """
    Calculate the FID score between two sets of images.

    Parameters:
    images1 (numpy array): First set of images.
    images2 (numpy array): Second set of images.

    Returns:
    float: The FID score.
    """
    # Preprocess images
    images1 = self.preprocess_images(images1)
    images2 = self.preprocess_images(images2)

    # Calculate statistics
    mu1, sigma1 = self.calculate_statistics(images1)
    mu2, sigma2 = self.calculate_statistics(images2)

    # Compute FID score
    ssdiff = np.sum((mu1 - mu2) ** 2.0)
    covmean = sqrtm(sigma1.dot(sigma2))

    # Check and correct for imaginary numbers
    if np.iscomplexobj(covmean):
        covmean = covmean.real

    fid = ssdiff + np.trace(sigma1 + sigma2 - 2.0 * covmean)
    return fid

calculate_statistics(images)

Calculate the feature statistics (mean and covariance) for a set of images.

Parameters: images (numpy array): Preprocessed images.

Returns: tuple: Mean and covariance of the features.

Source code in aigve/metrics/video_quality_assessment/distribution_based/fid.py 
34
35
36
37
38
39
40
41
42
43
44
45
46
47
def calculate_statistics(self, images):
    """
    Calculate the feature statistics (mean and covariance) for a set of images.

    Parameters:
    images (numpy array): Preprocessed images.

    Returns:
    tuple: Mean and covariance of the features.
    """
    features = self.model.predict(images)
    mu = features.mean(axis=0)
    sigma = np.cov(features, rowvar=False)
    return mu, sigma

preprocess_images(images)

Preprocess the images for the InceptionV3 model.

Parameters: images (numpy array): Input images.

Returns: numpy array: Preprocessed images.

Source code in aigve/metrics/video_quality_assessment/distribution_based/fid.py 
22
23
24
25
26
27
28
29
30
31
32
def preprocess_images(self, images):
    """
    Preprocess the images for the InceptionV3 model.

    Parameters:
    images (numpy array): Input images.

    Returns:
    numpy array: Preprocessed images.
    """
    return preprocess_input(images)