python - 如何在 tensorflow.keras 模型中正确使用自定义损失（例如骰子系数）？

Question

我正在尝试将自定义损失或度量函数输入到 tensorflow.keras 模型编译中，但出现错误。看起来输入没有正确传递。

我已经能够使用以下张量流损失运行模型编译/拟合：

loss = tf.keras.losses.MeanSquaredError() # Breaks if I remove ()
loss = tf.nn.sigmoid_cross_entropy_with_logits # Breaks if I add ()

我对使这些工作的语法差异感到困惑（可能是不正确的？），如果我使用自定义骰子损失，它不适用于任何一种语法。

当我在下面运行自定义骰子损失时，输入标签正确传递，batch_size*height*width但输入 logits 传递为None,None,None,None（似乎不正确？）并且骰子损失函数出错。我正在寻找批量优化，所以损失应该由model.fit.

def generalized_dice(labels, logits):
    smooth = 1e-17
    shape = tf.TensorShape(logits.shape).as_list()
    depth = int(shape[-1])
    labels = tf.one_hot(labels, depth, dtype=tf.float32)
    logits = tf.nn.softmax(logits)
    weights = 1.0 / (tf.reduce_sum(labels, axis=[0, 1, 2])**2)

    numerator = tf.reduce_sum(labels * logits, axis=[0, 1, 2])
    numerator = tf.reduce_sum(weights * numerator)

    denominator = tf.reduce_sum(labels + logits, axis=[0, 1, 2])
    denominator = tf.reduce_sum(weights * denominator)

    loss = 2.0*(numerator + smooth)/(denominator + smooth)
    return loss


def generalized_dice_loss(dice):
    return 1-dice 


model = tf.keras.Model(inputs=[input_x], outputs=[predictions])

loss = tf.keras.losses.MeanSquaredError() # Breaks if I remove (), with inputs being passed as None's 
# loss = tf.nn.sigmoid_cross_entropy_with_logits 
metric = generalized_dice # If no parantehses, labels passed as None's; 
# Both tf.nn.sigmoid_cross_entropy_with_logits and generalized_dice don't work if I add () saying inputs and labels must be provided; since I am training in batches, these need to be provided by the model during training, not when I am defining loss?  

model.compile(optimizer=tf.keras.optimizers.Adam(),  
              loss=loss, 
              metrics=[metric]) 

由于 dice loss 的输入作为 None 传递，

labels = tf.one_hot(labels, depth, dtype=tf.float32) 

结果是：

TypeError: Value passed to parameter 'indices' has DataType float32 not in list of allowed values: uint8, int32, int64

标签： pythontensorflowkerasloss