huber_loss

1.

tf.losses.huber_loss(
    labels,
    predictions,
    weights=1.0,
    delta=1.0,
    scope=None,
    loss_collection=tf.GraphKeys.LOSSES,
    reduction=Reduction.SUM_BY_NONZERO_WEIGHTS
)

2.

3.

import numpy as np
import matplotlib.pyplot as plt

def huber_loss(x, d):
    x = np.abs(x)
    return ( x<=d ) * x**2 /2 + ( x >d ) * d *( x - d /2.0)

if __name__ == '__main__':
    plt.figure(figsize=(6, 4.5), facecolor='w', edgecolor='k')
    x = np.linspace(-20, 20, 10000)
    plt.plot(x, x** 2 / 2, label='squared loss')
    for d in (10, 5, 3, 1):
        plt.plot(x, huber_loss(x, d), label=r'huber loss: $\delta$={}'.format(d))

    plt.legend(loc='best', frameon=False)
    plt.xlabel('residual')
    plt.ylabel('loss')
    plt.show()

4