python - Keras:如何将自定义损失中的 K.sum(y_true) 与 0 进行比较？

Question

我想定义自定义损失，但似乎无法将 keras 张量 K.sum(y_true) 与单个值 0 进行比较。

def custom_loss_keras(y_true, y_pred):
    if(K.sum(y_true) > 0):
        loss = K.mean(K.binary_crossentropy(y_true, y_pred), axis=-1)
        return loss
    else:
        loss = 0.0
        return loss

我还在损失函数中尝试了 K.eval() 来获取 numpy 数组，但失败了。

def custom_loss_keras(y_true, y_pred):
    y_true_np = K.eval(y_true) 
    #if(K.sum(y_true) > 0):
    if(np.sum(y_true_np) > 0):
        loss = K.mean(K.binary_crossentropy(y_true, y_pred), axis=-1)
        return loss
    else:
        loss = 0.0
        return loss

更新:

def custom_loss_keras(y_true, y_pred):
    if(K.greater(K.sum(y_true), 0)):
        loss = K.mean(K.binary_crossentropy(y_true, y_pred), axis=-1)
        return loss
    else:
        loss = 0.0
        return loss

它产生错误:

TypeError: Using a `tf.Tensor` as a Python `bool` is not allowed. Use `if t is not None:` instead of `if t:` to test if a tensor is defined, and use TensorFlow ops such as tf.cond to execute subgraphs conditioned on the value of a tensor.

我还尝试将建议的 tf.cond 与 keras 函数结合起来:

def custom_loss_keras(y_true, y_pred):
    loss = tf.cond(K.greater(K.sum(y_true),0), K.mean(K.binary_crossentropy(y_true, y_pred), axis=-1), 0.0)

    return loss

它产生错误:

     22 def custom_loss_keras(y_true, y_pred):
---> 23     loss = tf.cond(K.greater(K.sum(y_true),0), K.mean(K.binary_crossentropy(y_true, y_pred), axis=-1), 0.0)
     24 
     25     return loss

~/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/control_flow_ops.py in cond(pred, fn1, fn2, name)
   1718   with ops.name_scope(name, "cond", [pred]) as name:
   1719     if not callable(fn1):
-> 1720       raise TypeError("fn1 must be callable.")
   1721     if not callable(fn2):
   1722       raise TypeError("fn2 must be callable.")

TypeError: fn1 must be callable.

看来我需要用纯tensorflow来写？

Answer 1

在损失函数中使用 if 和 else(或 K.eval)将不起作用，因为 custom_loss_keras 中的行 在模型编译期间执行，而不是在模型拟合期间执行。

您可以使用 K.switch 而不是调用 tf.cond:

def custom_loss_keras(y_true, y_pred):
    loss = K.mean(K.binary_crossentropy(y_true, y_pred), axis=-1)
    condition = K.greater(K.sum(y_true), 0)
    return K.switch(condition, loss, K.zeros_like(loss))