Keras聚合目标函数 [英] Keras aggregated objective function

问题描述

如何向keras模型添加汇总错误?
有桌子:

How to add aggregated error to keras model?
Having table:

   g  x  y 
0  1  1  1   
1  1  2  2   
2  1  3  3   
3  2  1  2   
4  2  2  1   

我希望能够将sum((y - y_pred) ** 2)错误以及 每组sum((sum(y) - sum(y_pred)) ** 2).
可以有更大的单个样本错误，但对我而言，拥有正确的总数至关重要.

I would like to be able to minimize sum((y - y_pred) ** 2) error along with sum((sum(y) - sum(y_pred)) ** 2) per group.
I'm fine to have bigger individual sample errors, but it is crucial for me to have right totals.

SciPy示例:

import pandas as pd
from scipy.optimize import differential_evolution
df = pd.DataFrame({'g': [1, 1, 1, 2, 2], 'x': [1, 2, 3, 1, 2], 'y': [1, 2, 3, 2, 1]})
g = df.groupby('g')
def linear(pars, fit=False):
    a, b = pars
    df['y_pred'] = a + b * df['x']
    if fit:
        sample_errors = sum((df['y'] - df['y_pred']) ** 2)
        group_errors = sum((g['y'].sum() - g['y_pred'].sum()) ** 2)
        total_error = sum(df['y'] - df['y_pred']) ** 2
        return sample_errors + group_errors + total_error
    else:
        return df['y_pred']

pars = differential_evolution(linear, [[0, 10]] * 2, args=[('fit', True)])['x']
print('SAMPLES:\n', df, '\nGROUPS:\n', g.sum(), '\nTOTALS:\n', df.sum())

输出:

SAMPLES:
   g  x  y  y_pred
0  1  1  1   1.232
1  1  2  2   1.947
2  1  3  3   2.662
3  2  1  2   1.232
4  2  2  1   1.947 
GROUPS:
   x  y  y_pred
g              
1  6  6   5.841
2  3  3   3.179 
TOTALS:
g        7.000
x        9.000
y        9.000
y_pred   9.020

推荐答案

对于分组，只要在整个训练过程中保持相同的组，损失函数就不会出现不可微分的问题.

For grouping, as long as you keep the same groups throughout training, your loss function will not have problems about being not differentiable.

作为一种简单的分组方式，您可以简单地将批次分开.

As a naive form of grouping, you can simply separate the batches.

我为此建议一个生成器.

I suggest a generator for that.

#suppose you have these three numpy arrays:
gTrain
xTrain
yTrain

#create this generator
def grouper(g,x,y):
    while True:
        for gr in range(1,g.max()+1):
            indices = g == gr
            yield (x[indices],y[indices])

对于损失功能，您可以自己制作:

For the loss function, you can make your own:

import keras.backend as K

def customLoss(yTrue,yPred):
    return K.sum(K.square(yTrue-yPred)) + K.sum(K.sum(yTrue) - K.sum(yPred))

model.compile(loss=customLoss, ....)

如果您使用负值，请小心第二项.

Just be careful with the second term if you have negative values.

现在，您使用方法fit_generator进行训练:

Now you train using the method fit_generator:

model.fit_generator(grouper(gTrain,xTrain, yTrain), steps_per_epoch=gTrain.max(), epochs=...)

这篇关于Keras聚合目标函数的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！