plot_gradients.py 文件源码

def calc_mean_var_loss(epochsInds,loss_train):
    #Loss train is in dimension # epochs X #batchs
    num_of_epochs = loss_train.shape[0]
    #Average over the batchs
    loss_train_mean = np.mean(loss_train,1)
    #The diff divided by the sampled indexes
    d_mean_loss_to_dt = np.sqrt(np.abs(np.diff(loss_train_mean) / np.diff(epochsInds[:])))
    var_loss = []
    #Go over the epochs
    for epoch_index in range(num_of_epochs):
        #The loss for the specpic epoch
        current_loss = loss_train[epoch_index, :]
        #The derivative between the batchs
        current_loss_dt = np.diff(current_loss)
        #The mean of his derivative
        average_loss = np.mean(current_loss_dt)
        current_loss_minus_mean = current_loss_dt- average_loss
        #The covarince between the batchs
        cov_mat = np.dot(current_loss_minus_mean[:, None], current_loss_minus_mean[None, :])
        # The trace of the cov matrix
        trac_cov = np.trace(cov_mat)
        var_loss.append(trac_cov)
    return np.array(var_loss), d_mean_loss_to_dt