EffectiveDimension.get_fisher_information

EffectiveDimension.get_fisher_information(gradients, model_outputs)

This method computes the average Jacobian for every set of gradients and model output as shown in Abbas et al.

Parametri:

• gradients (ndarray) – A numpy array, result of the neural network’s backward pass, of shape (num_input_samples * num_weight_samples, output_size, num_weights).

• model_outputs (ndarray) – A numpy array, result of the neural networks” forward pass, of shape (num_input_samples * num_weight_samples, output_size).

Ritorna:

A numpy array of shape

(num_input_samples * num_weight_samples, num_weights, num_weights) with the average Jacobian for every set of gradients and model output given.

Tipo di ritorno:

fisher