如何对自定义 Keras/Tensorflow 损失函数中的值进行排序?

How can I sort the values in a custom Keras / Tensorflow Loss Function?

简介

我想为 Keras 实现自定义损失函数。我想这样做,因为我对我的数据集的当前结果不满意。我认为这是因为目前内置的损失函数关注的是整个数据集。我只想关注数据集中的最高值。这就是为什么我想出了以下自定义损失函数的想法:

自定义损失函数思路

自定义损失函数应该取最高值的前 4 个预测并减去相应的真实值。然后从这个减法中取绝对值,乘以一些权重并将其添加到总损失总和中。

为了更好地理解这个自定义损失函数,我用列表输入对其进行了编程。希望这个例子更容易理解:

下面的例子计算loss = 4*abs(0.7-0.5)+3*abs(0.5-0.7)+2*abs(0.4-0.45) +1*abs(0.4-0.3) = 1.6 for我=0

然后它除以 div_top,在这个例子中是 10(对于 i=0,它将是 0.16),对所有其他 i 重复一切,最后取所有样本的平均值。

top = 4
div_top = 0.5*top*(top+1)


def own_loss(y_true, y_pred):
    loss_per_sample = [0]*len(y_pred)
    for i in range(len(y_pred)):
        sorted_pred, sorted_true = (list(t) for t in zip(*sorted(zip(y_pred[i], y_true[i]))))
        for k in range(top):
            loss_per_sample[i] += (top-k)*abs(sorted_pred[-1-k]-sorted_true[-1-k])
    loss_per_sample = [t/div_top for t in loss_per_sample]
    return sum(loss_per_sample)/len(loss_per_sample)


y_pred = [[0.1, 0.4, 0.7, 0.4, 0.4, 0.5, 0.3, 0.2],
          [0.3, 0.8, 0.5, 0.3, 0.1, 0.0, 0.1, 0.5],
          [0.5, 0.6, 0.6, 0.8, 0.3, 0.6, 0.7, 0.1]]

y_true = [[0.2, 0.45, 0.5, 0.3, 0.4, 0.7, 0.22, 0.1],
          [0.4, 0.9, 0.3, 0.0, 0.2, 0.1, 0.11, 0.8],
          [0.4, 0.7, 0.4, 0.3, 0.4, 0.7, 0.6, 0.05]]

print(own_loss(y_true, y_pred)) # Output is 0.196667

Keras 的实现

我想在 Keras 中使用这个函数作为自定义损失函数。这看起来像这样:

import numpy as np
from keras.datasets import boston_housing
from keras.layers import LSTM
from keras.models import Sequential
from keras.optimizers import RMSprop

(pre_x_train, pre_y_train), (x_test, y_test) = boston_housing.load_data()
"""
The following 8 lines are to format the dataset to a 3D numpy array
4*101*13. I do this so that it matches my real dataset with is formatted
to a 3D numpy array 47*731*179. It is not important to understand the following 
8 lines for the loss function itself.
"""
x_train = [[0]*101]*4
y_train = [[0]*101]*4
for i in range(4):
    for k in range(101):
        x_train[i][k] = pre_x_train[i*101+k]
        y_train[i][k] = pre_y_train[i*101+k]
train_x = np.array([np.array([np.array(k) for k in i]) for i in x_train])
train_y = np.array([np.array([np.array(k) for k in i]) for i in y_train])


top = 4
div_top = 0.5*top*(top+1)


def own_loss(y_true, y_pred):
    loss_per_sample = [0]*len(y_pred)
    for i in range(len(y_pred)):
        sorted_pred, sorted_true = (list(t) for t in zip(*sorted(zip(y_pred[i], y_true[i]))))
        for k in range(top):
            loss_per_sample[i] += (top-k)*abs(sorted_pred[-1-k]-sorted_true[-1-k])
    loss_per_sample = [t/div_top for t in loss_per_sample]
    return sum(loss_per_sample)/len(loss_per_sample)


model = Sequential()
model.add(LSTM(units=64, batch_input_shape=(None, 101, 13), return_sequences=True))
model.add(LSTM(units=101, return_sequences=False, activation='linear'))
# compile works with loss='mean_absolute_error' but not with loss=own_loss
model.compile(loss=own_loss, optimizer=RMSprop())

model.fit(train_x, train_y, epochs=16, verbose=2, batch_size=1, validation_split=None, shuffle=False)

显然,上面的 Keras 示例是行不通的。但我也不知道我怎么能这样工作。

解决问题的方法

我阅读了以下文章,试图解决问题:

How to use a custom objective function for a model?

我还阅读了 Keras 后端页面:

Keras Backends

和 Tensorflow Top_k 页面:

tf.nn.top_k

这对我来说似乎是最有前途的方法,但在实施了许多不同的方法之后它仍然行不通。使用 top_k 排序时,我可以获得正确的 pred_y 值,但无法获得相应的 true_y 值。

有人知道我如何实现自定义损失函数吗?

假设

  • 使用 tf.nn.top_k 对张量进行排序。这意味着 "If two elements are equal, the lower-index element appears first" 如 API document.
  • 中所述

建议的解决方案

top = 4
div_top = 0.5*top*(top+1)

def getitems_by_indices(values, indices):
    return tf.map_fn(
        lambda x: tf.gather(x[0], x[1]), (values, indices), dtype=values.dtype
    )

def own_loss(y_true, y_pred):
    y_pred_top_k, y_pred_ind_k = tf.nn.top_k(y_pred, top)
    y_true_top_k = getitems_by_indices(y_true, y_pred_ind_k)
    loss_per_sample = tf.reduce_mean(
        tf.reduce_sum(
            tf.abs(y_pred_top_k - y_true_top_k) *
                tf.range(top, 0, delta=-1, dtype=y_pred.dtype),
            axis=-1
        ) / div_top
    )
    return loss_per_sample

model = Sequential()
model.add(LSTM(units=64, batch_input_shape=(None, 101, 13), return_sequences=True))
model.add(LSTM(units=101, return_sequences=False, activation='linear'))
# compile works with loss='mean_absolute_error' but not with loss=own_loss
model.compile(loss=own_loss, optimizer=RMSprop())

model.train_on_batch(train_x, train_y)

评论

  • getitems_by_indices()有没有更好的实现方式?
  • getitems_by_indices()的当前实现采用了 Sungwoon Kim 的想法。