Metrics¶
Metrics for evaluating the quality of generative models. These metrics help quantify the performance of the models.
Base Metric¶
The base class for all evaluation metrics. It defines the interface that all metric implementations must follow.
Main Methods:
- __call__(real, generated): Computes the metric value between real and generated samples.
Bits Per Dimension¶
Evaluates probabilistic generative models based on their log-likelihood. Lower values indicate better models.
Main Methods:
- __call__(real, _): Computes bits per dimension for the real data.
Fréchet Inception Distance¶
Measures the distance between feature representations of real and generated images using the Inception-v3 model. Lower values indicate better quality and diversity.
Main Methods:
- _get_inception(): Creates and prepares the Inception-v3 model for feature extraction.
- _get_activations(images): Extracts Inception features from input images.
- _calculate_fid(real, gen): Calculates the FID score from feature activations.
- __call__(real, generated): Computes the FID score between real and generated images.
Inception Score¶
Evaluates the quality and diversity of generated images using the Inception-v3 model. Higher values indicate better quality and diversity.
Main Methods:
- _get_inception(): Creates and prepares the Inception-v3 model for feature extraction.
- _get_predictions(images): Gets softmax predictions from the Inception model.
- _calculate_is(predictions): Calculates the Inception Score from softmax predictions.
- __call__(_, generated): Computes the Inception Score for generated images.