Question

假设您有一个内部带有acos的Keras损失函数，并且想要保护acos免受无效值（[-1,1]之外的值）的影响。即，将<-1设置为-1，将1设置为1。
你该怎么做？

我已经考虑过使用tf.cond，但不确定如何正确应用它：

def loss(y_true, y_pred):
    # typically in [-1,1], but maybe slightly smaller/larger than that due to num precision
    unprotected_acos_input = ... # shape: [batchsize]
    shape = unprotected_acos_input.shape[0]

    cond_low = tf.constant(-1, shape=[shape], dtype="float32")
    cond_high = tf.constant(1, shape=[shape], dtype="float32")

    protected_acos_input = tf.cond(tf.greater_equal(unprotected_acos_input, cond_low), lambda: unprotected_acos_input, lambda: tf.constant(-1, shape=[shape], dtype="float32"))
    # same for cond_high

    loss = tf.acos(protected_acos_input)
    return loss

如果批次中只有一个值在[-1,1]之外，这会将所有 acos输入设置为1或-1。如何只更改acos_input范围之外的项目，而不是所有项目？

Answer 1

您可以使用@objc func keyboardWillHide(notification: NSNotification) { if let keyboardSize = (notification.userInfo?[UIKeyboardFrameBeginUserInfoKey] as? NSValue)?.cgRectValue { self.collectionView?.contentInset = UIEdgeInsets(top: 8, left: 0, bottom: 52, right: 0) UIView.animate(withDuration: 0.5, delay: 0, usingSpringWithDamping: 1, initialSpringVelocity: 1, options: .curveEaseOut, animations: { self.view.layoutIfNeeded() }, completion: { (completed:Bool) in }) } }。

K.clip(unprotected_acos_input, -1, 1)

在Keras损失中保护Tensorflow Acos功能免受Nans的影响

1 个答案: