• PyTorch之MLP

    一 、关于Graphviz 的问题

    首先手动下载对应的包并安装，添加环境变量，如果仍然不行，考虑如下方法

    graphviz.backend.execute.ExecutableNotFound: failed to execute WindowsPath(‘dot’), make sure the Graphviz executables are on your systems’ PATH

    解决方案

    在代码文件中导入路径

    import os
os.environ["PATH"] += os.pathsep + 'D:/IProgramming/python/Iintsall/Graphviz/bin'

    • 1
    • 2

    二、运行出现警告信息

    UserWarning: This figure includes Axes that are not compatible with tight_layout, so results might be incorrect.
    self.figure.tight_layout()

    解决方案

    File > Settings > Tools > Python Scientific > show plots in tool window 取消勾选的show plots in tool window 点击Apply再点击OK

    也就是将绘画的图不显示在pycharm的小窗口，直接外部显示

    示例代码，实现邮件的垃圾分类

    import numpy as np
import pandas as pd
import seaborn as sns
# read data and show the first lines
import torch
import torch.nn as nn

from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler,MinMaxScaler
import matplotlib.pyplot as plt
from torchviz import make_dot
import os
os.environ["PATH"] += os.pathsep + 'D:/IProgramming/python/Iintsall/Graphviz/bin'

spam = pd.read_csv("./data/spambase.csv")
'''显示基本数据显示'''
# spam.head()
# pd.value_counts(spam.label)
''' 训练集和测试集划分'''
X= spam.iloc[:,0:57].values
y = spam.label.values
X_train,X_test,y_trains,y_test = train_test_split(X,y,test_size=0.25,random_state=123)
scales = MinMaxScaler(feature_range= (0,1))
X_train_s = scales.fit_transform(X_train)
X_test_s = scales.transform(X_test)

'''绘画箱线图'''
# plt.figure(figsize = (20,14))
# colname = spam.columns.values[:-1]
# for ii in range(len(colname)):
#     plt.subplot(7,9,ii+1)
#     sns.boxplot(x = y_trains,y = X_train_s[:,ii])
#     plt.title(colname[ii])
# plt.subplots_adjust(hspace= 0.4)
# plt.show()

'''搭建网络'''
class MLPclassifica(nn.Module):
    def __init__(self):
        super(MLPclassifica, self).__init__()
        self.hidden1 = nn.Sequential(
            nn.Linear(
                in_features=57,
                out_features=30,
                bias= True
            ),
            nn.ReLU()
        )
        self.hidden2 = nn.Sequential(
            nn.Linear(30,10),
            nn.ReLU()
        )
        self.classifica = nn.Sequential(
            nn.Linear(10,2),
            nn.Sigmoid()
        )
    def forward(self,x):
        fc1 = self.hidden1(x)
        fc2 = self.hidden2(fc1)
        output = self.classifica(fc2)
        return fc1,fc2,output

'''网络可视化'''
mplc = MLPclassifica()
x = torch.randn(1,57).requires_grad_(True)
y = mplc(x)
Mymplcvis = make_dot(y,params=dict(list(mplc.named_parameters())+[('x',x)]))
Mymplcvis
Mymplcvis.view('model_structure.pdf',"./data")  #第一个参数是文件名 第二个是保存路径


    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    • 8
    • 9
    • 10
    • 11
    • 12
    • 13
    • 14
    • 15
    • 16
    • 17
    • 18
    • 19
    • 20
    • 21
    • 22
    • 23
    • 24
    • 25
    • 26
    • 27
    • 28
    • 29
    • 30
    • 31
    • 32
    • 33
    • 34
    • 35
    • 36
    • 37
    • 38
    • 39
    • 40
    • 41
    • 42
    • 43
    • 44
    • 45
    • 46
    • 47
    • 48
    • 49
    • 50
    • 51
    • 52
    • 53
    • 54
    • 55
    • 56
    • 57
    • 58
    • 59
    • 60
    • 61
    • 62
    • 63
    • 64
    • 65
    • 66
    • 67
    • 68
    • 69
    • 70

    Out[9]: ‘data\model_structure.pdf.pdf’

    三，邮件分类之一

    未初始化数据的训练结果如下：
    代码：

    import numpy as np
import pandas as pd
import seaborn as sns
# read data and show the first lines
import torch
import torch.nn as nn
import hiddenlayer as hl
from sklearn.metrics import accuracy_score,confusion_matrix,classification_report
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler, MinMaxScaler
import torch.utils.data as Data
import matplotlib.pyplot as plt
from torchviz import make_dot
from sklearn.manifold import TSNE

import os

os.environ["PATH"] += os.pathsep + 'D:/IProgramming/python/Iintsall/Graphviz/bin'

spam = pd.read_csv("./data/spambase.csv")
'''显示基本数据显示'''
# spam.head()
# pd.value_counts(spam.label)
''' 训练集和测试集划分'''
X = spam.iloc[:, 0:57].values
y = spam.label.values
X_train, X_test, y_trains, y_test = train_test_split(X, y, test_size=0.25, random_state=123)
# 数据归一化
scales = MinMaxScaler(feature_range=(0, 1))
X_train_s = scales.fit_transform(X_train)
X_test_s = scales.transform(X_test)

'''绘画箱线图'''
# plt.figure(figsize = (20,14))
# colname = spam.columns.values[:-1]
# for ii in range(len(colname)):
#     plt.subplot(7,9,ii+1)
#     sns.boxplot(x = y_trains,y = X_train_s[:,ii])
#     plt.title(colname[ii])
# plt.subplots_adjust(hspace= 0.4)
# plt.show()

'''搭建网络'''


class MLPclassifica(nn.Module):
    def __init__(self):
        super(MLPclassifica, self).__init__()
        self.hidden1 = nn.Sequential(
            nn.Linear(
                in_features=57,
                out_features=30,
                bias=True
            ),
            nn.ReLU()
        )
        self.hidden2 = nn.Sequential(
            nn.Linear(30, 10),
            nn.ReLU()
        )
        self.classifica = nn.Sequential(
            nn.Linear(10, 2),
            nn.Sigmoid()
        )

    def forward(self, x):
        fc1 = self.hidden1(x)
        fc2 = self.hidden2(fc1)
        output = self.classifica(fc2)
        return fc1, fc2, output


# '''网络可视化'''
mlpc = MLPclassifica()
# x = torch.randn(1,57).requires_grad_(True)
# y = mlpc(x)
# Mymlpcvis = make_dot(y,params=dict(list(mlpc.named_parameters())+[('x',x)]))
# Mymlpcvis

# 使用未归一化的数据进行训练
X_train_nots = torch.from_numpy(X_train.astype(np.float32))
y_train_t = torch.from_numpy(y_trains.astype(np.int64))
X_test_nots = torch.from_numpy(X_test.astype(np.float32))
y_test_t = torch.from_numpy(y_test.astype(np.int64))
train_data_nots = Data.TensorDataset(X_train_nots,y_train_t)
train_data_loader = Data.DataLoader(
    dataset=train_data_nots,
    batch_size=64,
    shuffle=True,
    # RuntimeError: DataLoader worker (pid(s) 8704) exited unexpectedly
    # 上述错误，必须使用单线程
    num_workers=0
)

'''定义优化器'''
optimizer = torch.optim.Adam(mlpc.parameters(),lr = 0.01)
loss_func = nn.CrossEntropyLoss()
history1 = hl.History()
canvas1 = hl.Canvas()
print_step = 25
for epoch in range(15):
    for step,(b_x,b_y) in enumerate(train_data_loader):
        _,_,output = mlpc(b_x)
        train_loss = loss_func(output,b_y)
        optimizer.zero_grad()
        train_loss.backward()
        optimizer.step()
        niter = epoch*len(train_data_loader)+step+1
        if niter%print_step == 0:
            _,_,output = mlpc(X_test_nots)
            _,pre_lab = torch.max(output,1)
            test_accuracy = accuracy_score(y_test_t,pre_lab)
            history1.log(niter,train_loss = train_loss,test_accuracy = test_accuracy)
            with canvas1:
                canvas1.draw_plot(history1['train_loss'])
                canvas1.draw_plot(history1['test_accuracy'])


    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    • 8
    • 9
    • 10
    • 11
    • 12
    • 13
    • 14
    • 15
    • 16
    • 17
    • 18
    • 19
    • 20
    • 21
    • 22
    • 23
    • 24
    • 25
    • 26
    • 27
    • 28
    • 29
    • 30
    • 31
    • 32
    • 33
    • 34
    • 35
    • 36
    • 37
    • 38
    • 39
    • 40
    • 41
    • 42
    • 43
    • 44
    • 45
    • 46
    • 47
    • 48
    • 49
    • 50
    • 51
    • 52
    • 53
    • 54
    • 55
    • 56
    • 57
    • 58
    • 59
    • 60
    • 61
    • 62
    • 63
    • 64
    • 65
    • 66
    • 67
    • 68
    • 69
    • 70
    • 71
    • 72
    • 73
    • 74
    • 75
    • 76
    • 77
    • 78
    • 79
    • 80
    • 81
    • 82
    • 83
    • 84
    • 85
    • 86
    • 87
    • 88
    • 89
    • 90
    • 91
    • 92
    • 93
    • 94
    • 95
    • 96
    • 97
    • 98
    • 99
    • 100
    • 101
    • 102
    • 103
    • 104
    • 105
    • 106
    • 107
    • 108
    • 109
    • 110
    • 111
    • 112
    • 113
    • 114
    • 115
    • 116
    • 117

    输出结果如下：
    在这里插入图片描述
    结果分析：
    准确率88%，损失函数一直在波动，MLP数据并未收敛
    分析原因如下：

    1. 数据没有标准化处理
    2. 训练的数据样本过少
    3. 网络的神经元太少，或者太多

    四、邮件分类正确分类之二

    对其进行归一化处理
    归一后代码如下：

    import numpy as np
import pandas as pd
import torch
import torch.nn as nn
import hiddenlayer as hl
from sklearn.metrics import accuracy_score
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import MinMaxScaler
import torch.utils.data as Data

import os
os.environ["PATH"] += os.pathsep + 'D:/IProgramming/python/Iintsall/Graphviz/bin'

spam = pd.read_csv("./data/spambase.csv")

''' 训练集和测试集划分 1：3'''
X = spam.iloc[:, 0:57].values
y = spam.label.values
X_train, X_test, y_trains, y_test = train_test_split(X, y, test_size=0.25, random_state=123)
# 数据归一化
scales = MinMaxScaler(feature_range=(0, 1))
X_train_s = scales.fit_transform(X_train)
X_test_s = scales.transform(X_test)

'''搭建网络'''
class MLPclassifica(nn.Module):
    def __init__(self):
        super(MLPclassifica, self).__init__()
        self.hidden1 = nn.Sequential(
            nn.Linear(
                in_features=57,
                out_features=30,
                bias=True
            ),
            nn.ReLU()
        )
        self.hidden2 = nn.Sequential(
            nn.Linear(30, 10),
            nn.ReLU()
        )
        self.classifica = nn.Sequential(
            nn.Linear(10, 2),
            nn.Sigmoid()
        )
    def forward(self, x):
        fc1 = self.hidden1(x)
        fc2 = self.hidden2(fc1)
        output = self.classifica(fc2)
        return fc1, fc2, output

mlpc = MLPclassifica()

# 使用未归一化的数据进行训练
X_train_t = torch.from_numpy(X_train_s.astype(np.float32))
y_train_t = torch.from_numpy(y_trains.astype(np.int64))
X_test_t = torch.from_numpy(X_test_s.astype(np.float32))
y_test_t = torch.from_numpy(y_test.astype(np.int64))
train_data_nots = Data.TensorDataset(X_train_t,y_train_t)
train_data_loader = Data.DataLoader(
    dataset=train_data_nots,
    batch_size=64,
    shuffle=True,
    # RuntimeError: DataLoader worker (pid(s) 8704) exited unexpectedly
    # 上述错误，必须使用单线程
    num_workers=0
)

'''定义优化器'''
optimizer = torch.optim.Adam(mlpc.parameters(),lr = 0.01)
loss_func = nn.CrossEntropyLoss()
history1 = hl.History()
canvas1 = hl.Canvas()
print_step = 25
for epoch in range(15):
    for step,(b_x,b_y) in enumerate(train_data_loader):
        _,_,output = mlpc(b_x)
        train_loss = loss_func(output,b_y)
        optimizer.zero_grad()
        train_loss.backward()
        optimizer.step()
        niter = epoch*len(train_data_loader)+step+1
        if niter%print_step == 0:
            _,_,output = mlpc(X_test_t)
            _,pre_lab = torch.max(output,1)
            test_accuracy = accuracy_score(y_test_t,pre_lab)
            history1.log(niter,train_loss = train_loss,test_accuracy = test_accuracy)
            with canvas1:
                canvas1.draw_plot(history1['train_loss'])
                canvas1.draw_plot(history1['test_accuracy'])

# 在测试集合上进行测试
_,_,output = mlpc(X_test_t)
_,pre_lab = torch.max(output,1)
test_accuracy = accuracy_score(y_test_t,pre_lab)
print("test_accuracy :",test_accuracy)


    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    • 8
    • 9
    • 10
    • 11
    • 12
    • 13
    • 14
    • 15
    • 16
    • 17
    • 18
    • 19
    • 20
    • 21
    • 22
    • 23
    • 24
    • 25
    • 26
    • 27
    • 28
    • 29
    • 30
    • 31
    • 32
    • 33
    • 34
    • 35
    • 36
    • 37
    • 38
    • 39
    • 40
    • 41
    • 42
    • 43
    • 44
    • 45
    • 46
    • 47
    • 48
    • 49
    • 50
    • 51
    • 52
    • 53
    • 54
    • 55
    • 56
    • 57
    • 58
    • 59
    • 60
    • 61
    • 62
    • 63
    • 64
    • 65
    • 66
    • 67
    • 68
    • 69
    • 70
    • 71
    • 72
    • 73
    • 74
    • 75
    • 76
    • 77
    • 78
    • 79
    • 80
    • 81
    • 82
    • 83
    • 84
    • 85
    • 86
    • 87
    • 88
    • 89
    • 90
    • 91
    • 92
    • 93
    • 94
    • 95
    • 96

    结果如下图所示
    在这里插入图片描述
    结果分析，可以看到准确率达到93%，损失函数也在不断下降。
    对训练的模型进行预测，结果如下，认为94.52%

    Backend TkAgg is interactive backend. Turning interactive mode on.
    test_accuracy : 0.945264986967854

    五、获取中间层的可视化工作

    ''中间层输出可视化'''
_, test_fc2, _ = mlpc(X_test_t)
# print('test_fc2.shape:', test_fc2.shape)

# 对输出进行降维并可视化
# test_fc2_tsne = TSNE(n_components = 2).fit_transform(test_fc2.data.numpy())
# # 特征可视化
# plt.figure(figsize=(8, 6))
# # 可视化前设置坐标系的取值范围
# plt.xlim([min(test_fc2_tsne[:, 0] - 1), max(test_fc2_tsne[:, 0] + 1)])
# plt.ylim([min(test_fc2_tsne[:, 1] - 1), max(test_fc2_tsne[:, 1] + 1)])
# plt.plot(test_fc2_tsne[y_test == 0, 0], test_fc2_tsne[y_test == 0, 1],
#          'bo', label = '0')
# plt.plot(test_fc2_tsne[y_test == 1, 0], test_fc2_tsne[y_test == 1, 1],
#          'rd', label = '1')
# plt.legend()
# plt.title('test_fc2_tsne')
# plt.show()

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    • 8
    • 9
    • 10
    • 11
    • 12
    • 13
    • 14
    • 15
    • 16
    • 17
    • 18

    对应结果如下
    在这里插入图片描述
    结果分析
    略
    使用hook输出的

    
'''使用钩子获取信息'''
# 定义一个辅助函数，获取指定层名称的特征
activation = {} # 保存不同层的输出

def get_activaion(name):
    def hook(model, input, output):
        activation[name] = output.detach()
    return hook
# 获取分类层输出
mlpc.classifica.register_forward_hook(get_activaion('classify'))
_, _, _ = mlpc(X_test_t)
classifica = activation['classify'].data.numpy()
print('classification.shape:', classifica.shape)

# 特征可视化
plt.figure(figsize=(8, 6))
plt.plot(classifica[y_test == 0, 0], classifica[y_test == 0, 1],
         'bo', label = '0')
plt.plot(classifica[y_test == 1, 0], classifica[y_test == 1, 1],
         'rd', label = '1')
plt.legend()
plt.title('classification')
plt.show()


    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    • 8
    • 9
    • 10
    • 11
    • 12
    • 13
    • 14
    • 15
    • 16
    • 17
    • 18
    • 19
    • 20
    • 21
    • 22
    • 23
    • 24
    • 25

    结果如下：

    classification.shape: (1151, 2)

    对应的输出图
    在这里插入图片描述
    分析如下：

    六、参数调整：

    lr学习率
    迭代次数
    bachsize的设定，epoch的设定等参数的理解何调参数对结果的影响

  • 相关阅读:
    功能比较：Redisson vs Jedis
    GenAI的“关键一跃”：推理与知识
    STC 32位8051单片机开发实例教程 二 I/O工作模式及其配置
    PHP接口自动化测试框架实现
    Kibana中的Dev Tools常用语法
    关系型数据库语言基础整理
    辅助生成：低延迟文本生成的新方向
    【Vue】第一个Vue程序
    VL813 VL817 VL820 是USB 3.0集线器芯片
    java Stream操作
  • 原文地址：https://blog.csdn.net/Android_WPF/article/details/125965904