python - Keras 损失函数取决于批量大小

Question

我正在尝试在 Keras 中构建一个损失函数，其中我对预测与一组给定值之间的最小距离进行惩罚。问题是我需要计算预测值和给定值之间的距离。

示例代码

def custom_loss(y_pred,y_test):


    #Given values
    centers=K.constant([[-2.5,-1],[-1.25,-2],[.5,-1],[1.5,.25]])
    num_centers=K.int_shape(centers)[0]


    #Begin constructing distance matrix
    height=K.int_shape(y_pred)[0]
    i=0
    current_center=K.reshape(K.repeat(K.reshape(centers[i,:],[1,-1]),height),[height,2])
    current_dist=K.sqrt(K.sum(K.square(y_pred-current_center),axis=1))


    #Values of distance matrix for first center
    Distance=K.reshape(current_dist,[height,1])


    for i in range(1,num_centers):
        current_center=K.reshape(K.repeat(K.reshape(centers[i,:],[1,-1]),height),[height,2])
        current_dist=K.sqrt(K.sum(K.square(y_pred-current_center),axis=-1))
        current_dist=K.reshape(current_dist,[height,1])


        #Iteratively concatenate distances of y_pred from remaining centers
        Distance=K.concatenate([Distance,current_dist],axis=-1)

    #Determine minimum distance from each predicted value to nearest center
    A=K.min(A,axis=1)


    #Return average minimum distance as loss
    return K.sum(A)/float(height)

但是，我无法消除函数对 y_pred 第一个维度的依赖，该维度是可变的。我使用数组广播来计算 y_pred 和每个给定值之间的差异，但我显式使用批量大小进行广播，因为我不知道如何在不使用 Keras 中的批量大小的情况下执行此操作。然而，这会产生错误，因为在构建计算图时并未明确知道批量大小。

如何避免显式广播？目前的方法非常笨拙，是否有更有效的计算距离矩阵的方法？

Answer 1

您的损失函数可以使用隐式广播来实现，如下所示:

import keras.backend as K


def custom_loss(y_true, y_pred):
    centers = K.constant([[-2.5, -1], [-1.25, -2], [.5, -1], [1.5, .25]])

    # Expand dimensions to enable implicit broadcasting
    y_pred_r = y_pred[:, None, :]  # Shape: (batch_size, 1, 2)
    centers_r = centers[None, :, :]  # Shape: (1, nb_centers, 2)

    # Compute minimum distance to centers for each element
    distances = K.sqrt(K.sum(K.square(y_pred_r - centers_r), axis=-1))  # Shape=(batch_size, nb_centers)
    min_distances = K.min(distances, axis=-1)  # Shape=(batch_size,)

    # Output average of minimum distances
    return K.mean(min_distances)

注意:未经测试。