nan

[weixiaolian21]

the loss will be nan sometime due to tf.log().
L = -labels*(1-alpha)*((1-y_pred)**gamma)*tf.log(tf.clip_by_value(y_pred, 1e-8, 1.0)) -\ (1-labels)*alpha*(y_pred**gamma)*tf.log(tf.clip_by_value(1.0 - y_pred, 1e-8, 1.0))