Python 网络训练中的奇异损失函数

我正在进行深度学习。我用NSL-KDD数据集创建了一个自动编码器网络。我想用真实的方式训练网络。但在训练过程中，我的损失函数显示了奇怪的数字，我无法理解。 这是我的源代码：

import os
import numpy as np
from keras.layers import Dense, Input
from keras.models import Model
def load_nsl_kdd_cup_dataset(filename):
    with open(filename, 'r')as f:
        list1 = []
        list2 = []
        for i in f:
            list1.append(i)

        for i in list1:
            i = str(i)
            labels = i.split(sep=',')
            list2.append(labels[:41])
        arraylist = np.array(list2)
        return arraylist

dirs = './NSL_KDD-master'
train_X = load_nsl_kdd_cup_dataset(os.path.join(dirs,'KDDTrain+.txt'))
test_X  = load_nsl_kdd_cup_dataset(os.path.join(dirs, 'KDDTest+.txt'))
print(train_X.shape)
print(test_X.shape)

预处理阶段1。 将标称值转换为数值

def convert_nominal_to_numerical(sets):
    list1 = []
    col1, col2, col3 = [], [], []
    for i in sets:
        # print(i[1])
        if i[1] == 'tcp':
            i[1] = 1
        elif i[1] == 'udp':
            i[1] = 2
        elif i[1] == 'icmp':
            i[1] = 3

        if i[2] == 'http':
            i[2] = 1
        elif i[2] == 'private':
            i[2] = 2
        elif i[2] == 'domain_u':
            i[2] = 3
        elif i[2] == 'smtp':
            i[2] = 4
        elif i[2] == 'ftp_data':
            i[2] = 5
        elif i[2] == 'eco_i':
            i[2] = 6
        elif i[2] == 'other':
            i[2] = 7
        elif i[2] == 'ecr_i':
            i[2] = 8
        elif i[2] == 'telnet':
            i[2] = 9
        elif i[2] == 'finger':
            i[2] = 10
        elif i[2] == 'ftp':
            i[2] = 11
        elif i[2] == 'auth':
            i[2] = 12
        elif i[2] == 'Z39_50':
            i[2] = 13
        elif i[2] == 'uucp':
            i[2] = 14
        elif i[2] == 'courier':
            i[2] = 15
        elif i[2] == 'bgp':
            i[2] = 16
        elif i[2] == 'whois':
            i[2] = 17
        elif i[2] == 'uucp_path':
            i[2] = 18
        elif i[2] == 'iso_tsap':
            i[2] = 19
        elif i[2] == 'time':
            i[2] = 20
        elif i[2] == 'imap4':
            i[2] = 21
        elif i[2] == 'nnsp':
            i[2] = 22
        elif i[2] == 'vmnet':
            i[2] = 23
        elif i[2] == 'urp_i':
            i[2] = 24
        elif i[2] == 'domain':
            i[2] = 25
        elif i[2] == 'ctf':
            i[2] = 26
        elif i[2] == 'csnet_ns':
            i[2] = 27
        elif i[2] == 'supdup':
            i[2] = 28
        elif i[2] == 'discard':
            i[2] = 29
        elif i[2] == 'http_443':
            i[2] = 30
        elif i[2] == 'daytime':
            i[2] = 31
        elif i[2] == 'gopher':
            i[2] = 32
        elif i[2] == 'efs':
            i[2] = 33
        elif i[2] == 'systat':
            i[2] = 34
        elif i[2] == 'link':
            i[2] = 35
        elif i[2] == 'exec':
            i[2] = 36
        elif i[2] == 'hostnames':
            i[2] = 37
        elif i[2] == 'name':
            i[2] = 38
        elif i[2] == 'mtp':
            i[2] = 39
        elif i[2] == 'echo':
            i[2] = 40
        elif i[2] == 'klogin':
            i[2] = 41
        elif i[2] == 'login':
            i[2] = 42
        elif i[2] == 'ldap':
            i[2] = 43
        elif i[2] == 'netbios_dgm':
            i[2] = 44
        elif i[2] == 'sunrpc':
            i[2] = 45
        elif i[2] == 'netbios_ssn':
            i[2] = 46
        elif i[2] == 'netstat':
            i[2] = 47
        elif i[2] == 'netbios_ns':
            i[2] = 48
        elif i[2] == 'ssh':
            i[2] = 49
        elif i[2] == 'kshell':
            i[2] = 50
        elif i[2] == 'nntp':
            i[2] = 51
        elif i[2] == 'pop_3':
            i[2] = 52
        elif i[2] == 'sql_net':
            i[2] = 53
        elif i[2] == 'IRC':
            i[2] = 54
        elif i[2] == 'ntp_u':
            i[2] = 55
        elif i[2] == 'rje':
            i[2] = 56
        elif i[2] == 'remote_job':
            i[2] = 57
        elif i[2] == 'pop_2':
            i[2] = 58
        elif i[2] == 'X11':
            i[2] = 59
        elif i[2] == 'printer':
            i[2] = 60
        elif i[2] == 'shell':
            i[2] = 61
        elif i[2] == 'urh_i':
            i[2] = 62
        elif i[2] == 'tim_i':
            i[2] = 63
        elif i[2] == 'red_i':
            i[2] = 64
        elif i[2] == 'pm_dump':
            i[2] = 65
        elif i[2] == 'tftp_u':
            i[2] = 66
        elif i[2] == 'http_8001':
            i[2] = 67
        elif i[2] == 'aol':
            i[2] = 68
        elif i[2] == 'harvest':
            i[2] = 69
        elif i[2] == 'http_2784':
            i[2] = 70

        if i[3] == 'SF':
            i[3] = 1
        elif i[3] == 'S0':
            i[3] = 2
        elif i[3] == 'REJ':
            i[3] = 3
        elif i[3] == 'RSTR':
            i[3] = 4
        elif i[3] == 'RSTO':
            i[3] = 5
        elif i[3] == 'S1':
            i[3] = 6
        elif i[3] == 'SH':
            i[3] = 7
        elif i[3] == 'S2':
            i[3] = 8
        elif i[3] == 'RSTOS0':
            i[3] = 9
        elif i[3] == 'S3':
            i[3] = 10
        elif i[3] == 'OTH':
            i[3] = 11
        list1.append(i)
        col1.append(i[1])
        col2.append(i[2])
        col3.append(i[3])
    return list1, col1, col2, col3

预处理阶段2。 在每列中查找最小值和最大值

def min_max_finder(col):
    minimum_col = 10
    maximum_col = 1
    for i in col:
        if float(i) < minimum_col:
            minimum_col = float(i)
        if float(i) > maximum_col:
            maximum_col = float(i)
    return minimum_col, maximum_col

预处理阶段4。 用原始数据替换标准化数据

def replace_normalized_with_original(normalized_col, which_col, data):
    for i in range(len(data)):
        data[i][which_col] = normalized_col[i]
    return data
def convert_list_to_array(list1):
    arrays = np.array(list1)
    return arrays

list1, col1, col2, col3 = convert_nominal_to_numerical(train_X)
min_col1, max_col1 = min_max_finder(col=col1)
min_col2, max_col2 = min_max_finder(col=col2)
min_col3, max_col3 = min_max_finder(col=col3)

normalized_col1 = data_normalizer(col=col1)
normalized_col2 = data_normalizer(col=col2)
normalized_col3 = data_normalizer(col=col3)

list1 = replace_normalized_with_original(normalized_col=normalized_col1,which_col=1, data=train_X)
list1 = replace_normalized_with_original(normalized_col=normalized_col2, which_col=2, data=train_X)
list1 = replace_normalized_with_original(normalized_col=normalized_col3, which_col=3, data=train_X)

X_train_normalized = convert_list_to_array(list1)
list_test, col1_test, col2_test, col3_test = convert_nominal_to_numerical(test_X)
min_col1_test, max_col1_test = min_max_finder(col=col1_test)
min_col2_test, max_col2_test = min_max_finder(col=col2_test)
min_col3_test, max_col3_test = min_max_finder(col=col3_test)

normalized_col1_test = data_normalizer(col=col1_test)
normalized_col2_test = data_normalizer(col=col2_test)
normalized_col3_test = data_normalizer(col=col3_test)

list_test = replace_normalized_with_original(normalized_col=normalized_col1_test, which_col=1, data=test_X)
list_test = replace_normalized_with_original(normalized_col=normalized_col2_test, which_col=2, data=test_X)
list_test = replace_normalized_with_original(normalized_col=normalized_col3_test, which_col=3, data=test_X)

X_test_normalized = convert_list_to_array(list_test)

X_train_normalized = X_train_normalized.astype('float32')
X_test_normalized = X_test_normalized.astype('float32')

training_X = load_nsl_kdd_cup_dataset(os.path.join(dirs, 'KDDTrain+.txt'))
testing_X  = load_nsl_kdd_cup_dataset(os.path.join(dirs, 'KDDTest+.txt'))

用密集层实现自动编码器网络

input_layer = Input(shape=(41,))
encoded = Dense(41, activation='relu')(input_layer)
encoded = Dense(20, activation='relu')(encoded)
encoded = Dense(10, activation='relu')(encoded)

decoded = Dense(10, activation='relu')(encoded)
decoded = Dense(20,activation='relu')(decoded)
decoded = Dense(41, activation='simoid')(decoded)

IDS_autoencoder = Model(inputs=input_layer, outputs=decoded)
IDS_autoencoder.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])
IDS_autoencoder.summary()
IDS_autoencoder.fit(X_train_normalized, X_train_normalized,
               epochs=5,
               batch_size=128,
               validation_data=(X_test_normalized,X_test_normalized))

这是我的输出：

对125973个样本进行训练，对22544个样本进行验证
纪元1/5
125973/125973[=========================================================3s 26us/阶跃-损耗：-25513.2260-附件：0.5736-val\u损耗：-5095.4520-val\u附件：0.6534
纪元2/5
125973/125973[==================================================================3s 23us/阶跃-损耗：-25675.6222-acc:0.6492-val_损耗：-5095.5203-val_acc:0.6512
纪元3/5
125973/125973[============================================================3s 23us/阶跃-损耗：-25675.5876-附件：0.6483-val\u损耗：-5095.5543-val\u附件：0.6520
纪元4/5
125973/125973[==============================================================================-3s 23us/阶跃-损耗：-25675.6049-附件：0.6493-val\u损耗：-5095.5665-val\u附件：0.6526
纪元5/5
125973/125973[===========================================================================-3s 23us/阶跃-损耗：-25675.6039-附件：0.6477-val\u损耗：-5095.5619-val\u附件：0.6536

您似乎有多类标签问题，因此您确定要使用二进制交叉熵损失和sigmoid激活函数吗？还有一个关于“sigmoid”的输入错误。我在中使用了很多损失函数，其中一些我得到了wierd损失函数，其中一些我得到了非常低的精度，我更改了损失和优化器，但仍然得到了wierd。你了解你使用的损失函数吗？有些损失函数适用于某些特定的东西，它们期望你的输出在一定的范围内，以便适用。不，我是这个领域的新手。你能帮我吗伙计？你想做什么分类？这是一个41类的分类，每个类都是独占的吗？也就是说，每一个物体都应该被划分为一个类别，并且只有一个类别？

input_layer = Input(shape=(41,))
encoded = Dense(41, activation='relu')(input_layer)
encoded = Dense(20, activation='relu')(encoded)
encoded = Dense(10, activation='relu')(encoded)

decoded = Dense(10, activation='relu')(encoded)
decoded = Dense(20,activation='relu')(decoded)
decoded = Dense(41, activation='simoid')(decoded)

IDS_autoencoder = Model(inputs=input_layer, outputs=decoded)
IDS_autoencoder.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])
IDS_autoencoder.summary()
IDS_autoencoder.fit(X_train_normalized, X_train_normalized,
               epochs=5,
               batch_size=128,
               validation_data=(X_test_normalized,X_test_normalized))