我有一些数据和代码，例如示例数据，以及下面的代码。在代码中，我有两个示例，其中一个是定义自定义损失函数并训练神经网络分类器。还有一个示例代码，其中我定义了一个自定义损失函数并针对同一数据训练了adaboost分类器。我真正想弄清楚的部分是，在两种情况下，我是否都使用自定义损失功能实现了相同的目标。
也就是说，“最小化损失函数以仅使用训练数据来选择最佳模型。”
如果我考虑正确，则应在反向传播过程中使用神经网络中的损失函数，以从前馈过程中训练的模型中获取初始输出，并根据一些训练数据值对预测进行基本分级，然后调整模型以获得较小的损失函数值。
在带有adaboost分类器的gridsearchcv步骤中，我使用相同的自定义损失函数来评估从给定的一组超参数创建的模型，并将其输出到训练数据中，然后选择使训练中的损失函数最小化的模型数据。
Adaboost分类器和神经网络并不完全相同，但是整个过程似乎是相同的。
我阅读了以下有关创建自定义损失函数的博客，以及他们谈论培训损失和验证损失的方式使我感到困惑：
https://towardsdatascience.com/custom-loss-functions-for-gradient-boosting-f79c1b40466d
如果我仅将gridsearchCV与训练数据一起使用，那将是“训练损失”吧？
这也与我发表的一篇早期文章有关，但这是一个完全不同的问题。
较早的帖子： make custom scorer with GridSearchCV
样本数据：
print(x_train_scld[:5])
[[ 0.37773519 2.0109691 0.49644224 0.21679945 0.538941 1.99144889
2.15011467 1.20312084 0.86114816 0.79507318 -0.45602028 0.07146743
-0.19524294 -0.33405545 -0.60264522 1.26724727 1.44991588 0.74630967
0.16529837 0.89613455 0.3253014 2.19166429 0.64865429 0.12894674
0.46995314 3.41479052 4.44308499 1.83182458 1.54348561 2.50155582]
[ 0.32029317 0.1214269 0.28824456 0.13510828 -0.0851059 -0.0057386
-0.31671716 0.0303454 0.32754165 -0.15354084 -0.36310852 -0.34419771
-0.28347519 -0.28927174 -0.39507256 -0.2039463 -0.49919802 0.12281647
-0.56756272 -0.30637335 0.10701249 0.21461633 0.17531634 -0.04414507
0.19574444 0.36354262 -1.23318869 0.59029124 0.28936372 0.19248437]
[ 0.25843254 0.29037034 0.21339798 0.12738073 0.28185716 -0.47995085
-0.13321816 0.14228058 -3.69915162 -0.10246162 0.26193423 0.12807553
0.18956053 0.12487671 -0.28174435 -0.71770499 -0.34455425 0.00729992
-0.70102685 -0.57022389 0.59171701 0.77319193 0.52065985 -1.37655715
0.59387438 -1.52826854 0.18054306 0.76212977 0.3639211 0.08726502]
[-0.70482588 -0.32963569 -0.74849491 -0.86505667 0.10026287 -0.87877366
-1.06584707 -1.19559926 0.34039964 0.10112554 -0.62427503 -0.3134676
-0.65996358 -0.52932857 0.11989554 -0.95345177 -0.67459484 -0.82130922
-0.52228025 -0.38191412 -0.75239269 -0.31180246 -0.7418967 -0.7432583
0.12191902 -0.97620932 -1.02049823 -1.20098216 -0.02333216 -0.24853266]
[-0.36680171 -0.14757043 -0.41413663 -0.56754624 -0.34512544 -0.76162172
-0.72684687 -0.61557149 0.31896966 -0.25351016 -0.6357623 0.12484078
-0.71632135 -0.51097128 0.26933611 -0.53549047 -0.54070413 -0.36472263
-0.24581883 -0.67901706 -0.44128802 0.16221265 -0.42239358 -0.52459003
0.34339528 -0.43064345 -1.23318869 -0.23310168 0.44404246 -0.40964978]]
print(x_test_scld[:5])
[[ 2.60641850e-01 -7.18369636e-01 3.27138629e-01 -1.76172773e+00
4.67645320e-01 1.53766591e+00 7.62837058e-01 4.07109050e-01
7.71142242e-01 9.80417766e-01 5.10262027e-01 5.66383900e-01
9.28678845e-01 2.06576727e-01 9.68389151e-01 1.48288576e+00
7.53349504e-01 7.04842193e-01 7.80186706e-01 6.43850055e-01
1.43107505e-01 -7.20312971e-01 2.96065817e-01 -4.51322867e-02
1.93107816e-01 7.41280492e-01 3.28514299e-01 4.47039330e-02
1.39136160e-01 4.94989991e-01]
[-7.51730115e-02 4.92568820e-02 -7.29146850e-02 -2.86318841e-01
1.00026599e+00 4.43886212e-01 4.80336890e-01 6.71683119e-01
8.61148159e-01 5.21434522e-01 -3.65135682e-01 -4.32021118e-01
-4.10049198e-01 -3.01778906e-01 -4.27568719e-02 -1.34413479e+00
-4.09570872e-02 1.64283954e-01 -3.04209384e-01 -7.10176931e-03
7.32148655e-03 -2.90459367e+00 2.31719950e-02 -1.37655715e+00
1.44286672e+00 1.07281572e+00 1.19548020e+00 1.44805187e+00
1.33316704e+00 1.55622575e+00]
[-1.23777794e-01 -3.83763205e-01 -1.65737513e-01 -3.43999436e-01
3.58604868e-01 -3.45623859e-01 -2.89602186e-01 -3.38277511e-01
8.23494778e-03 2.97415674e-01 -6.27653637e-01 -6.42441486e-01
-7.17707195e-01 -4.34516210e-01 6.01100047e-01 -2.64325075e-01
-2.31751338e-01 4.13624916e-02 7.46820672e-01 3.84336779e-01
-3.24408912e-01 -5.30945125e-01 -3.14685046e-01 -4.13363730e-01
6.43970206e-01 -2.37091815e-01 -1.45963962e-01 -2.97594271e-02
7.54512744e-01 6.49530907e-01]
[ 1.06041146e+00 3.61350612e-02 9.93240469e-01 1.11126264e+00
-2.54537983e-01 -2.50709092e-01 -3.56042668e-02 -1.19559926e+00
-2.25351836e-01 -4.65124054e-01 -4.64466800e-01 -1.10808348e+00
-4.47005113e-01 -2.07571731e-01 -1.11908130e+00 -8.49190558e-01
-5.40704133e-01 -6.40037086e-01 -1.10737748e+00 -9.30940117e-01
9.76730527e-01 2.34863210e-01 9.02228200e-01 9.43399666e-01
-1.25487123e-02 -1.70804996e-03 4.83277659e-01 7.07714236e-01
5.60886115e-01 -4.38009686e-01]
[ 3.57851416e-01 1.87811066e+00 2.77785646e-01 2.23975029e-01
-3.66933526e-01 -9.49100986e-01 -4.74866806e-01 -4.98802740e-01
2.69680706e-01 -5.60715159e-01 2.46392629e-01 7.53999293e-01
1.19344293e-01 1.24473258e-01 4.50284535e-02 -5.74844494e-01
-1.80203418e-01 -2.89340672e-01 1.37362545e+00 -6.91305992e-01
2.80612333e-01 1.49136056e+00 1.99466234e-01 1.55930637e-01
-2.39298218e-01 -9.12274848e-01 -4.82659170e-01 -6.00406523e-01
5.90931626e-01 -7.55722792e-01]]
print(y_train[:5])
562 1
291 0
16 1
546 0
293 0
Name: diagnosis, dtype: int64
print(y_test[:5])
421 0
47 1
292 0
186 1
414 1
Name: diagnosis, dtype: int64
代码：
# custom loss function
# importing libraries
import io
import os
import time
import pandas as pd
import numpy as np
import keras
from keras.models import Sequential
from keras.layers import Dense
import keras.backend as K
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import classification_report, roc_curve, roc_auc_score, precision_recall_fscore_support, accuracy_score
import matplotlib.pyplot as plt
from IPython.core.display import display, HTML
# from sklearn.grid_search import GridSearchCV
from sklearn.model_selection import GridSearchCV
from sklearn.metrics import fbeta_score, make_scorer
from sklearn.ensemble import AdaBoostClassifier, GradientBoostingClassifier
# neural network version
# custom loss function
def custom_loss_wrapper(fn_cost=1, fp_cost=1):
def custom_loss(y_true, y_pred, fn_cost=fn_cost, fp_cost=fp_cost):
h = K.ones_like(y_pred)
fn_value = fn_cost * h
fp_value = fp_cost * h
weighted_values = y_true * K.abs(1-y_pred)*fn_value + (1-y_true) * K.abs(y_pred)*fp_value
loss = K.mean(weighted_values)
return loss
return custom_loss
#The 'custom' model is trained on our custom loss function that weighs
#false negatives 5 time's more heavily than false positives.
model_five = Sequential()
model_five.add(Dense(units=num_classes, input_dim=input_dim, activation='sigmoid'))
# compile model
#Trained using custom loss function with false negatives weighted 5 times more heavily than false positives
custom_loss_five = custom_loss_wrapper(fn_cost=5, fp_cost=1)
model_five.compile(loss=custom_loss_five,
optimizer='sgd',
metrics=['accuracy'])
def custom_loss2(y_true, y_pred):
fn_cost, fp_cost = 5, 1
h = np.ones(len(y_pred))
fn_value = fn_cost * h
fp_value = fp_cost * h
weighted_values = y_true * np.abs(1-y_pred)*fn_value + (1-y_true) * np.abs(y_pred)*fp_value
loss = np.mean(weighted_values)
return loss
# TODO: Initialize the classifier
clf = AdaBoostClassifier(random_state=0)
# TODO: Create the parameters list you wish to tune
parameters = {'n_estimators':[100,200,300],'learning_rate':[1.0,2.0,4.0]}
# TODO: Make an fbeta_score scoring object
scorer2 = make_scorer(custom_loss2,greater_is_better=False)
# TODO: Perform grid search on the classifier using 'scorer' as the scoring method
grid_obj2 = GridSearchCV(clf,parameters,scoring=scorer2)
# TODO: Fit the grid search object to the training data and find the optimal parameters
grid_fit2 = grid_obj2.fit(x_train_scld,y_train)
# Get the estimator
best_clf2 = grid_fit2.best_estimator_
# Make predictions using the optimized and model
best_predictions = best_clf2.predict(x_test_scld)