pytorch pso优化cnn-lstm 智慧海洋-渔船轨迹识别

1、摘要

本文主要讲解：pytorch pso优化cnn-lstm 智慧海洋-渔船轨迹识别
主要思路：

1. 根据经纬度和时间序列创建时序块
2. 数据集随机分成训练和测试
3. 定义PSO Parameters（粒子数量、 惯性权重）、LSTM Parameters、搜索维度和粒子的位置和速度等定义、所有粒子的位置和速度、个体经历的最佳位置和全局最佳位置、每个个体的历史最佳适应值
4. 开始搜索、初始粒子适应度计算、计算适应值、更新个体最优、更新全局最优
5. 训练模型 使用PSO找到的最好的神经元个数

2、数据介绍

赛题与数据
train_chusai.csv为我处理好的数据，如有需要请私聊

3、相关技术

1、基于滑动窗口和LSTM自动编码器的渔船作业类型识别
2、TOP1解决方案
3、基于滑动窗口和PSO-CNN-LSTM的渔船轨迹识别
PSO优化LSTM介绍

4、完整代码和步骤

代码输出如下：

主运行程序入口

import random
import time
import matplotlib.pyplot as plt
import numpy as np
import torch
from sklearn import preprocessing
from sklearn.model_selection import train_test_split
from torch import nn
from torch.utils.data import TensorDataset
from sklearn.metrics import classification_report, accuracy_score

torch.manual_seed(1)  # reproducible
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Hyper Parameters
EPOCH = 122  # train the training data n times, to save time, we just train 1 epoch
LR = 0.001  # learning rate
num_layers = 2
hidden_size = 256

import os
import pandas as pd

os.chdir(r'C:\Projects')
label_dict1 = {'拖网': 0, '围网': 1, '刺网': 2}
label_dict2 = {0: '拖网', 1: '围网', 2: '刺网'}
name_dict = {'渔船ID': 'id', '速度': 'v', '方向': 'dir', 'type': 'label', 'lat': 'x', 'lon': 'y'}
train = pd.read_csv('train_chusai.csv')
# train_chusai = pd.read_csv('train_chusai.csv')
train.rename(columns=name_dict, inplace=True)
# train_chusai.rename(columns=name_dict, inplace=True)

df = train
df['x'] = df['x'] * 100000 - 5630000
df['y'] = df['y'] * 110000 + 2530000
# df = pd.concat([train_chusai, df], axis=0, ignore_index=True)
df = df.drop(['time'], axis=1)
df['label'] = df['label'].map(label_dict1)
X_train = df.iloc[:, :-1]
y_train = df.iloc[:, -1]

INPUT_SIZE = X_train.shape[1]  # rnn input size
x_MinMax = preprocessing.MinMaxScaler()
x = x_MinMax.fit_transform(X_train.values)
x = pd.DataFrame(x, columns=['id', 'x', 'y', 'v', 'dir'])
df = pd.concat([x, y_train], axis=1, ignore_index=True)
df.columns = ['id', 'x', 'y', 'v', 'dir', 'label']
# 创建时序块
groups = df.groupby(['id'])
x_train3 = np.zeros(shape=(1, 300, 4))
y_train = []
for name, group in groups:
    group_nd = group.iloc[:300, 1:-1].values
    if group_nd.shape[0] == 300:
        group_nd = group_nd.reshape((1, 300, 4))
        x_train3 = np.append(x_train3, group_nd, axis=0)
        y_train.append(group.iloc[0, -1])
x_train3_nd = x_train3[1:]
y_train = np.array(y_train)
# 数据集随机分成训练和测试
X_train, X_test, y_train, y_test = train_test_split(x_train3_nd, y_train, test_size=0.2, shuffle=True)
train_data1 = torch.FloatTensor(X_train)
X_test = torch.FloatTensor(X_test)
y = np.eye(3)[y_train]
train_labels = torch.tensor(y).long()
y_test = np.eye(3)[y_test]
train_data, test_x, test_y = TensorDataset(train_data1, train_labels), X_test, y_test
test_x = test_x.to(device)
X_train = train_data1.to(device)


class CNN_LSTM(nn.Module):
    def __init__(self, hidden_size, num_layers):
        super(CNN_LSTM, self).__init__()
        lstm_units = hidden_size
        self.conv1d = nn.Conv1d(4, lstm_units, 1)
        self.act1 = nn.Sigmoid()
        self.maxPool = nn.MaxPool1d(kernel_size=4)
        self.drop = nn.Dropout(p=0.01)
        self.lstm = nn.LSTM(lstm_units, lstm_units, batch_first=True, num_layers=num_layers, bidirectional=True)
        self.act2 = nn.Tanh()
        self.cls = nn.Linear(lstm_units * 2, 3)

    def forward(self, x):
        x = x.transpose(-1, -2)  # tf和torch纬度有点不一样
        x = self.conv1d(x)  # in： bs, dim, window out: bs, lstm_units, window
        x = self.act1(x)
        x = self.maxPool(x)  # bs, lstm_units, 1
        x = self.drop(x)
        x = x.transpose(-1, -2)  # bs, 1, lstm_units
        x, (_, _) = self.lstm(x)  # bs, 1, 2*lstm_units
        x = self.act2(x)
        x = x.squeeze(dim=1)  # bs, 2*lstm_units
        x = self.cls(x)
        return x


def training(X):
    hidden_size = int(X[0])
    num_layers = int(X[1])
    LR = round(X[2], 6)
    batch_size = int(X[3])
    EPOCH = int(X[4])
    print([hidden_size, num_layers, LR, batch_size, EPOCH])
    lstm_model = CNN_LSTM(hidden_size, num_layers).to(device)
    optimizer = torch.optim.AdamW(lstm_model.parameters(), lr=LR)  # optimize all cnn parameters
    loss_func = nn.CrossEntropyLoss().to(device)  # the target label is not one-hotted
    # Data Loader for easy mini-batch return in training
    train_loader = torch.utils.data.DataLoader(dataset=train_data, batch_size=batch_size, shuffle=True, pin_memory=True,
                                               sampler=None)
    # training and testing
    for epoch in range(EPOCH):
        for step, (b_x, b_y) in enumerate(train_loader):  # gives batch data
            # b_x = b_x.view(-1, 1, INPUT_SIZE)  # reshape x to (batch, time_step, input_size)
            b_x = b_x.to(device)
            b_y = b_y.to(device)
            # b_x = b_x.unsqueeze(1)
            output = lstm_model(b_x)  # rnn output
            loss = loss_func(output, b_y)  # cross entropy loss
            optimizer.zero_grad()  # clear gradients for this training step
            loss.backward()  # backpropagation, compute gradients
            optimizer.step()  # apply gradients

        test_output = lstm_model(test_x)  # (samples, time_step, input_size)
        pred_y = torch.max(test_output, 1)[1].data.cpu().numpy()
        accuracy = float((pred_y == test_y).astype(int).sum()) / float(test_y.size)
    print('accuracy:', accuracy)
    return accuracy


# (1) PSO Parameters
MAX_EPISODES = 10
MAX_EP_STEPS = 10
c1 = 2
c2 = 2
pN = 10 # 粒子数量
w_start = 0.9  # 惯性权重
w_end = 0.1

t1 = time.time()
# (2) LSTM Parameters
'''
hidden_size = int(X[0])
num_layers = int(X[1])
LR = round(X[2], 6)
batch_size = int(X[3])
EPOCH = int(X[4])

'''
UP = [222, 4, 0.014, 18, 126]
DOWN = [50, 1, 0.0001, 6, 25]
# [234, 4, 0.001411, 39, 65]
# accuracy: 0.894124847001224
# (3) 搜索维度和粒子的位置和速度等定义
dim = len(UP)  # 搜索维度
X = np.zeros((pN, dim))  # 所有粒子的位置和速度
V = np.zeros((pN, dim))
pbest = np.zeros((pN, dim))  # 个体经历的最佳位置和全局最佳位置
gbest = np.zeros(dim)
p_fit = np.zeros(pN)  # 每个个体的历史最佳适应值
sum_epochs = MAX_EP_STEPS * MAX_EPISODES * pN + pN * MAX_EPISODES
print('总搜索参数组合数 =', sum_epochs)

# (4) 开始搜索
for i_episode in range(MAX_EPISODES):
    """初始化s"""
    random.seed(8)
    fit = 0.6  # 全局最佳适应值
    # 初始粒子适应度计算
    print("计算初始全局最优")
    for i in range(pN):
        for j in range(dim):
            V[i][j] = random.uniform(0, 1)
            if j in [2]:
                X[i][j] = random.uniform(DOWN[j], UP[j])
            else:
                X[i][j] = round(random.randint(DOWN[j], UP[j]))
        pbest[i] = X[i]
        # 计算适应值
        tmp = training(X[i])
        NN = 1
        p_fit[i] = tmp
        if tmp > fit:
            fit = tmp
            gbest = X[i]
    print("初始全局最优参数：{:}".format(gbest))

    fitness = []  # 适应度函数
    for j in range(MAX_EP_STEPS):

        for i in range(pN):
            temp = training(X[i])
            if temp > p_fit[i]:  # 更新个体最优
                p_fit[i] = temp
                pbest[i] = X[i]
                if p_fit[i] > fit:  # 更新全局最优
                    gbest = X[i]
                    fit = p_fit[i]

        for i in range(pN):
            w = w_start - (w_start - w_end) * (2 * (i / pN) - (i / pN) ** 2)  # 修改后的惯性权重表达式
            V[i] = w * V[i] + c1 * random.uniform(0, 1) * (pbest[i] - X[i]) + c2 * random.uniform(0, 1) * (
                    gbest - X[i])
            ww = 1
            for k in range(dim):
                if DOWN[k] < X[i][k] + V[i][k] < UP[k]:
                    continue
                else:
                    ww = 0
            X[i] = X[i] + V[i] * ww
        fitness.append(fit)
    print("全局最优参数：{:}".format(gbest))
    print("全局最佳适应值：{:}".format(fit))

print('Running time: ', time.time() - t1)

# 训练模型  使用PSO找到的最好的神经元个数
X = gbest
hidden_size = int(X[0])
num_layers = int(X[1])
LR = round(X[2], 6)
batch_size = int(X[3])
EPOCH = int(X[4])

# Data Loader for easy mini-batch return in training
train_loader = torch.utils.data.DataLoader(dataset=train_data, batch_size=batch_size, shuffle=True, pin_memory=True,
                                           sampler=None)

lstm_model = CNN_LSTM(hidden_size, num_layers).to(device)
print(lstm_model)

optimizer = torch.optim.AdamW(lstm_model.parameters(), lr=LR)  # optimize all cnn parameters
loss_func = nn.CrossEntropyLoss()  # the target label is not one-hotted
test_accs = []
train_accs = []
losss = []
# training and testing
for epoch in range(EPOCH):
    for step, (b_x, b_y) in enumerate(train_loader):  # gives batch data
        b_x = b_x.to(device)
        b_y = b_y.to(device)
        output = lstm_model(b_x)  # rnn output
        loss = loss_func(output, b_y)  # cross entropy loss
        optimizer.zero_grad()  # clear gradients for this training step
        loss.backward()  # backpropagation, compute gradients
        optimizer.step()  # apply gradients

    train_output = lstm_model(b_x)  # (samples, time_step, input_size)
    pred_train_y = torch.max(train_output, 1)[1].data.cpu().numpy()
    train_y_pred = np.argmax(pred_train_y, axis=1)

    train_y_real = torch.max(b_y, 1)[1].data.cpu().numpy()
    # train_y_real = np.argmax(train_y_real, axis=1)
    train_accuracy = accuracy_score(train_y_pred, train_y_real)
    test_output = lstm_model(test_x)  # (samples, time_step, input_size)
    pred_y = torch.max(test_output, 1)[1].data.cpu().numpy()
    test_accuracy = float((pred_y == test_y).astype(int).sum()) / float(test_y.size)
    print('Epoch: ', epoch, '| train loss: %.4f' % loss.data.cpu().numpy  (), '| test accuracy: %.3f' % test_accuracy)
    losss.append(loss.data.cpu().numpy())
    test_accs.append(test_accuracy)
    train_accs.append(train_accuracy)

photo_path = 'C:\Projects\photo\\'
plt.plot(losss)
plt.title('LSTM_loss')
plt.ylabel('loss')
plt.xlabel('Epoch')
plt.legend(['train'], loc='best')
plt.savefig(photo_path + "LSTM_PSO_train_loss.png", dpi=500, bbox_inches='tight')
plt.show()

plt.plot(train_accs)
plt.plot(test_accs)
plt.title('accuracy')
plt.ylabel('accuracy')
plt.xlabel('Epoch')
plt.legend(['train', 'test'], loc='best')
plt.savefig(photo_path + "CNN_LSTM_PSO_val_accuracy.png", dpi=500, bbox_inches='tight')
plt.show()
print('accuracy:', max(test_accs))

cls_report = classification_report(y_true=test_y, y_pred=pred_y)
print(cls_report)

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
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283