|
|
|
"""Contains the functions to compute Frechet Inception Distance (FID). |
|
|
|
FID metric is introduced in paper |
|
|
|
GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash |
|
Equilibrium. Heusel et al. NeurIPS 2017. |
|
|
|
See details at https://arxiv.org/pdf/1706.08500.pdf |
|
""" |
|
|
|
import numpy as np |
|
import scipy.linalg |
|
|
|
__all__ = ['extract_feature', 'compute_fid'] |
|
|
|
|
|
def extract_feature(inception_model, images): |
|
"""Extracts feature from input images with given model. |
|
|
|
NOTE: The input images are assumed to be with pixel range [-1, 1]. |
|
|
|
Args: |
|
inception_model: The model used to extract features. |
|
images: The input image tensor to extract features from. |
|
|
|
Returns: |
|
A `numpy.ndarray`, containing the extracted features. |
|
""" |
|
features = inception_model(images, output_logits=False) |
|
features = features.detach().cpu().numpy() |
|
assert features.ndim == 2 and features.shape[1] == 2048 |
|
return features |
|
|
|
|
|
def compute_fid(fake_features, real_features): |
|
"""Computes FID based on the features extracted from fake and real data. |
|
|
|
Given the mean and covariance (m_f, C_f) of fake data and (m_r, C_r) of real |
|
data, the FID metric can be computed by |
|
|
|
d^2 = ||m_f - m_r||_2^2 + Tr(C_f + C_r - 2(C_f C_r)^0.5) |
|
|
|
Args: |
|
fake_features: The features extracted from fake data. |
|
real_features: The features extracted from real data. |
|
|
|
Returns: |
|
A real number, suggesting the FID value. |
|
""" |
|
|
|
m_f = np.mean(fake_features, axis=0) |
|
C_f = np.cov(fake_features, rowvar=False) |
|
m_r = np.mean(real_features, axis=0) |
|
C_r = np.cov(real_features, rowvar=False) |
|
|
|
fid = np.sum((m_f - m_r) ** 2) + np.trace( |
|
C_f + C_r - 2 * scipy.linalg.sqrtm(np.dot(C_f, C_r))) |
|
return np.real(fid) |
|
|
