系列文章目录
【时间序列篇】基于LSTM的序列分类-Pytorch实现 part1 案例复现
【时间序列篇】基于LSTM的序列分类-Pytorch实现 part2 自有数据集构建
【时间序列篇】基于LSTM的序列分类-Pytorch实现 part3 化为己用
在一个人体姿态估计的任务中，需要用深度学习模型来进行序列分类。
化为己用，实现成功。
文章目录
- 系列文章目录
- 前言
- 一、模型训练
-
- [1 导入库和自用函数](#1 导入库和自用函数)
- [2 导入数据集](#2 导入数据集)
- [3 设备部署](#3 设备部署)
- [4 网络模型](#4 网络模型)
- [5 训练过程](#5 训练过程)
- [6 运行结果](#6 运行结果)
- [7 完整代码](#7 完整代码)
- SequenceClassifier
- 二、模型预测
-
- [1 完整代码](#1 完整代码)
- [2 运行结果](#2 运行结果)
- 三、模型评估
-
- [1 完整代码](#1 完整代码)
- [2 运行结果](#2 运行结果)
- 总结
前言
结合了part1 和 part2的文章，处理现有序列分类任务。
基于LSTM的序列分类-Pytorch实现 这个部分先告一段落。
part3 主要是优化后的代码实现，包括训练，预测，模型评估。
一、模型训练
这一部分就是对part1文章中代码的优化和运行结果。每一节都是一整个代码的一部分，最后放完整代码。
1 导入库和自用函数
python
import os
import copy
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
import torch.optim as optim
from tqdm import tqdm
def calculate_accuracy(y_pred, y_true):
_, predicted_labels = torch.max(y_pred, 1)
correct = (predicted_labels == y_true).float()
accuracy = correct.sum() / len(correct)
return accuracy2 导入数据集
python
'''
/****************************************************/
导入数据集
/****************************************************/
'''
# ----------------------------------------------------#
# 路径指定
# ----------------------------------------------------#
ROOT_PATH = "DATA/RT_Position_dataset"
dataset_path = os.path.join(ROOT_PATH, "dataset")
target_path = os.path.join(ROOT_PATH, "groups/Movement4_target.csv")
groups_path = os.path.join(ROOT_PATH, "groups/Movement4_DatasetGroup.csv")
checkpoint_pth = "best_model.pth" # 模型参数
seq_len = 16 # 序列长度
# ----------------------------------------------------#
# sequences列表读取所有的样本
# ----------------------------------------------------#
path = os.path.join(dataset_path, "Movement4_")
sequences = list()
for i in range(1, 3731): # 3731为样本数
file_path = path + str(i) + '.csv'
# print(file_path)
df = pd.read_csv(file_path, header=0)
values = df.values
sequences.append(values)
# print(len(sequences))
# len_sequences = []
# for one_seq in sequences:
# len_sequences.append(len(one_seq))
# ----------------------------------------------------#
# 数据集标签
# ----------------------------------------------------#
targets = pd.read_csv(target_path)
targets = targets.values[:, 1]
# ----------------------------------------------------#
# 数据集划分
# ----------------------------------------------------#
groups = pd.read_csv(groups_path, header=0)
groups = groups.values[:, 1]
# ----------------------------------------------------#
# Padding the sequence with the values in last row to max length
# ----------------------------------------------------#
# 函数用于填充和截断序列
def pad_truncate_sequences(sequences, max_len, dim=4, truncating='post', padding='post'):
# 初始化一个空的numpy数组，用于存储填充后的序列
padded_sequences = np.zeros((len(sequences), max_len, dim))
for i, one_seq in enumerate(sequences):
if len(one_seq) > max_len: # 截断
if truncating == 'pre':
padded_sequences[i] = one_seq[-max_len:]
else:
padded_sequences[i] = one_seq[:max_len]
else: # 填充
padding_len = max_len - len(one_seq)
to_concat = np.repeat(one_seq[-1], padding_len).reshape(dim, padding_len).transpose()
if padding == 'pre':
padded_sequences[i] = np.concatenate([to_concat, one_seq])
else:
padded_sequences[i] = np.concatenate([one_seq, to_concat])
return padded_sequences
# ----------------------------------------------------#
# 设置序列长度
# ----------------------------------------------------#
# 使用自定义函数进行填充和截断
final_seq = pad_truncate_sequences(sequences, max_len=seq_len, dim=4, truncating='post', padding='post')
# 设置标签从 1~6 换为 0~5
targets = np.array(targets)
final_targets = targets - 1
# ----------------------------------------------------#
# 数据集划分
# ----------------------------------------------------#
# 将numpy数组转换为PyTorch张量
final_seq = torch.tensor(final_seq, dtype=torch.float)
# 划分样本为 训练集，验证集
train = [final_seq[i] for i in range(len(groups)) if groups[i] == 1]
validation = [final_seq[i] for i in range(len(groups)) if groups[i] == 2]
# 标签同理
train_target = [final_targets[i] for i in range(len(groups)) if groups[i] == 1]
validation_target = [final_targets[i] for i in range(len(groups)) if groups[i] == 2]
# 转换为PyTorch张量
train = torch.stack(train)
train_target = torch.tensor(train_target).long()
validation = torch.stack(validation)
validation_target = torch.tensor(validation_target).long()3 设备部署
python
'''
/****************************************************/
device
/****************************************************/
'''
def device_on():
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using {device} device")
return device
# ----------------------------------------------------#
# device
# ----------------------------------------------------#
device = device_on()4 网络模型
python
'''
/****************************************************/
网络模型
/****************************************************/
'''
# ----------------------------------------------------#
# 创建模型
# ----------------------------------------------------#
class TimeSeriesClassifier(nn.Module):
def __init__(self, n_features, hidden_dim=256, output_size=1):
super().__init__()
self.lstm = nn.LSTM(input_size=n_features, hidden_size=hidden_dim, batch_first=True)
self.fc = nn.Linear(hidden_dim, output_size) # output_size classes
def forward(self, x):
x, _ = self.lstm(x) # LSTM层
x = x[:, -1, :] # 只取LSTM输出中的最后一个时间步
x = self.fc(x) # 通过一个全连接层
return x
# ----------------------------------------------------#
# 模型实例化
# ----------------------------------------------------#
seq_len = 16 # 根据你的序列长度进行调整
n_features = 4 # 根据你的特征数量进行调整
output_size = 6
model = TimeSeriesClassifier(n_features=n_features, output_size=output_size)
# # 打印模型结构
print(model)
# ----------------------------------------------------#
# 模型部署
# ----------------------------------------------------#
model.to(device)5 训练过程
python
'''
/****************************************************/
训练过程
/****************************************************/
'''
# 设置训练参数
epochs = 100 # 训练轮数，根据需要进行调整
batch_size = 4 # 批大小，根据你的硬件调整
# DataLoader 加载数据集
train_dataset = torch.utils.data.TensorDataset(train, train_target)
train_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True)
validation_dataset = torch.utils.data.TensorDataset(validation, validation_target)
validation_loader = torch.utils.data.DataLoader(dataset=validation_dataset, batch_size=batch_size, shuffle=True)
# 定义损失函数和优化器
criterion = nn.CrossEntropyLoss()
criterion = criterion.to(device)
# 学习率和优化策略
learning_rate = 1e-3
optimizer = optim.Adam(params=model.parameters(), lr=learning_rate, weight_decay=5e-4)
lr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=2, gamma=0.5) # 设置学习率下降策略
# ----------------------------------------------------#
# 训练
# ----------------------------------------------------#
best_acc = 0.0
for epoch in range(epochs):
model.train() # 将模型设置为训练模式
train_epoch_loss = []
train_epoch_accuracy = []
pbar = tqdm(train_loader, total=len(train_loader))
for index, (inputs, labels) in enumerate(pbar, start=1):
# 获取输入数据和目标，并将它们转移到GPU（如果可用）
inputs = inputs.to(device)
labels = labels.to(device)
# 清零梯度
optimizer.zero_grad()
# 前向传播
outputs = model(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
train_epoch_loss.append(loss.item())
accuracy = calculate_accuracy(outputs, labels)
train_epoch_accuracy.append(accuracy.item())
pbar.set_description(f'Epoch [{epoch + 1}/{epochs}]')
pbar.set_postfix(**{'loss': loss.item(),
'accuracy': accuracy.item(),
})
# Validation accuracy
model.eval()
valid_epoch_loss = []
valid_epoch_accuracy = []
pbar = tqdm(validation_loader, total=len(validation_loader))
for index, (inputs, labels) in enumerate(pbar, start=1):
inputs = inputs.to(device)
labels = labels.to(device)
outputs = model(inputs)
loss = criterion(outputs, labels)
valid_epoch_loss.append(loss.item())
accuracy = calculate_accuracy(outputs, labels)
valid_epoch_accuracy.append(accuracy.item())
pbar.set_description('valid')
pbar.set_postfix(**{'total_loss': loss.item(),
'accuracy': accuracy.item(),
})
# 计算平均精度
print("--------------------------------------------")
train_epoch_loss = np.average(train_epoch_loss)
train_epoch_accuracy = np.average(train_epoch_accuracy)
print(f'Epoch {epoch + 1}, train Accuracy: {train_epoch_accuracy:.4f}')
valid_epoch_loss = np.average(valid_epoch_loss)
valid_epoch_accuracy = np.average(valid_epoch_accuracy)
print(f'Epoch {epoch + 1}, Validation Accuracy: {valid_epoch_accuracy:.4f}')
print("--------------------------------------------")
if valid_epoch_accuracy > best_acc:
best_acc = valid_epoch_accuracy
best_model_wts = copy.deepcopy(model.state_dict())
state = {
'state_dict': model.state_dict(),
'best_acc': best_acc,
'optimizer': optimizer.state_dict(),
}
torch.save(state, checkpoint_pth)
print('Finished Training')
print('Best val Acc: {:4f}'.format(best_acc))6 运行结果
可以看到相比起part1的实验，分类准确率更高。毕竟part1总共314个样本，自有数据集样本有3730个。
7 完整代码
SequenceClassifier
二、模型预测
输入样本，运行模型，输出预测
1 完整代码
这里直接贴完整代码。
python
"""
@file name:predict.py
@desc: 用于序列分类预测
"""
import os
import pandas as pd
import torch
import torch.nn as nn
'''
/****************************************************/
输入一个数据样本
/****************************************************/
'''
# 读取CSV样本文件
csv_file = "DATA/RT_Position_dataset/dataset/Movement4_7.csv"
data = pd.read_csv(csv_file)
# 将Pandas DataFrame转换为NumPy数组
data_array = data.values
# 将NumPy数组转换为PyTorch张量
input = torch.tensor(data_array, dtype=torch.float).unsqueeze(0)
'''
/****************************************************/
device
/****************************************************/
'''
def device_on():
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using {device} device")
return device
# ----------------------------------------------------#
# device
# ----------------------------------------------------#
device = device_on()
'''
/****************************************************/
网络模型
/****************************************************/
'''
# ----------------------------------------------------#
# 创建模型
# ----------------------------------------------------#
class TimeSeriesClassifier(nn.Module):
def __init__(self, n_features, hidden_dim=256, output_size=1):
super().__init__()
self.lstm = nn.LSTM(input_size=n_features, hidden_size=hidden_dim, batch_first=True)
self.fc = nn.Linear(hidden_dim, output_size) # output_size classes
def forward(self, x):
x, _ = self.lstm(x) # LSTM层
x = x[:, -1, :] # 只取LSTM输出中的最后一个时间步
x = self.fc(x) # 通过一个全连接层
return x
# ----------------------------------------------------#
# 模型实例化
# ----------------------------------------------------#
seq_len = 16 # 根据你的序列长度进行调整
n_features = 4 # 根据你的特征数量进行调整
output_size = 6
model = TimeSeriesClassifier(n_features=n_features, output_size=output_size)
# ----------------------------------------------------#
# 模型部署
# ----------------------------------------------------#
model.to(device)
# ----------------------------------------------------#
# 导入模型参数
# ----------------------------------------------------#
save_pth = "best_model.pth"
checkpoint = torch.load(save_pth)
best_acc = checkpoint['best_acc']
model.load_state_dict(checkpoint['state_dict'])
'''
/****************************************************/
预测输出
/****************************************************/
'''
model.eval() # 设置为评估模式
# 假设你有一个预处理好的序列数据，shape为(batch_size, seq_len, n_features)
# 例如：一个序列，长度为16，每个时间步有4个特征
# input = torch.randn(1, 16, 4) # 替换为你的数据
with torch.no_grad():
output = model(input)
# 将输出转换为概率分布
prediction = torch.softmax(output, dim=1)
# 取得最高概率的类别作为预测结果
predicted_class = torch.argmax(prediction, dim=1)
print(f"Predicted class: {predicted_class}")2 运行结果
三、模型评估
1 完整代码
python
"""
@file name:evaluate.py
@desc: 用于模型评估
"""
import os
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
from sklearn.metrics import accuracy_score, confusion_matrix, precision_score, recall_score, f1_score
def calculate_accuracy(y_pred, y_true):
_, predicted_labels = torch.max(y_pred, 1)
correct = (predicted_labels == y_true).float()
accuracy = correct.sum() / len(correct)
return accuracy
'''
/****************************************************/
导入数据集
/****************************************************/
'''
# ----------------------------------------------------#
# 路径指定
# ----------------------------------------------------#
ROOT_PATH = "DATA/RT_Position_dataset"
dataset_path = os.path.join(ROOT_PATH, "dataset")
target_path = os.path.join(ROOT_PATH, "groups/Movement4_target.csv")
groups_path = os.path.join(ROOT_PATH, "groups/Movement4_DatasetGroup.csv")
checkpoint_pth = "best_model.pth" # 模型参数
seq_len = 16 # 序列长度
# ----------------------------------------------------#
# sequences列表读取所有的样本
# ----------------------------------------------------#
path = os.path.join(dataset_path, "Movement4_")
sequences = list()
for i in range(1, 3731): # 3731为样本数
file_path = path + str(i) + '.csv'
# print(file_path)
df = pd.read_csv(file_path, header=0)
values = df.values
sequences.append(values)
# print(len(sequences))
# len_sequences = []
# for one_seq in sequences:
# len_sequences.append(len(one_seq))
# ----------------------------------------------------#
# 数据集标签
# ----------------------------------------------------#
targets = pd.read_csv(target_path)
targets = targets.values[:, 1]
# ----------------------------------------------------#
# 数据集划分
# ----------------------------------------------------#
groups = pd.read_csv(groups_path, header=0)
groups = groups.values[:, 1]
# ----------------------------------------------------#
# Padding the sequence with the values in last row to max length
# ----------------------------------------------------#
# 函数用于填充和截断序列
def pad_truncate_sequences(sequences, max_len, dim=4, truncating='post', padding='post'):
# 初始化一个空的numpy数组，用于存储填充后的序列
padded_sequences = np.zeros((len(sequences), max_len, dim))
for i, one_seq in enumerate(sequences):
if len(one_seq) > max_len: # 截断
if truncating == 'pre':
padded_sequences[i] = one_seq[-max_len:]
else:
padded_sequences[i] = one_seq[:max_len]
else: # 填充
padding_len = max_len - len(one_seq)
to_concat = np.repeat(one_seq[-1], padding_len).reshape(dim, padding_len).transpose()
if padding == 'pre':
padded_sequences[i] = np.concatenate([to_concat, one_seq])
else:
padded_sequences[i] = np.concatenate([one_seq, to_concat])
return padded_sequences
# ----------------------------------------------------#
# 设置序列长度
# ----------------------------------------------------#
# 使用自定义函数进行填充和截断
final_seq = pad_truncate_sequences(sequences, max_len=seq_len, dim=4, truncating='post', padding='post')
# 设置标签从 1~6 换为 0~5
targets = np.array(targets)
final_targets = targets - 1
# ----------------------------------------------------#
# 数据集划分
# ----------------------------------------------------#
# 将numpy数组转换为PyTorch张量
final_seq = torch.tensor(final_seq, dtype=torch.float)
# 划分样本为 训练集，验证集
train = [final_seq[i] for i in range(len(groups)) if groups[i] == 1]
validation = [final_seq[i] for i in range(len(groups)) if groups[i] == 2]
# 标签同理
train_target = [final_targets[i] for i in range(len(groups)) if groups[i] == 1]
validation_target = [final_targets[i] for i in range(len(groups)) if groups[i] == 2]
# 转换为PyTorch张量
train = torch.stack(train)
train_target = torch.tensor(train_target).long()
validation = torch.stack(validation)
validation_target = torch.tensor(validation_target).long()
'''
/****************************************************/
device
/****************************************************/
'''
def device_on():
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using {device} device")
return device
# ----------------------------------------------------#
# device
# ----------------------------------------------------#
device = device_on()
'''
/****************************************************/
网络模型
/****************************************************/
'''
# ----------------------------------------------------#
# 创建模型
# ----------------------------------------------------#
class TimeSeriesClassifier(nn.Module):
def __init__(self, n_features, hidden_dim=256, output_size=1):
super().__init__()
self.lstm = nn.LSTM(input_size=n_features, hidden_size=hidden_dim, batch_first=True)
self.fc = nn.Linear(hidden_dim, output_size) # output_size classes
def forward(self, x):
x, _ = self.lstm(x) # LSTM层
x = x[:, -1, :] # 只取LSTM输出中的最后一个时间步
x = self.fc(x) # 通过一个全连接层
return x
# ----------------------------------------------------#
# 模型实例化
# ----------------------------------------------------#
seq_len = 16 # 根据你的序列长度进行调整
n_features = 4 # 根据你的特征数量进行调整
output_size = 6
model = TimeSeriesClassifier(n_features=n_features, output_size=output_size)
# # 打印模型结构
print(model)
# ----------------------------------------------------#
# 模型部署
# ----------------------------------------------------#
model.to(device)
# ----------------------------------------------------#
# 导入模型参数
# ----------------------------------------------------#
checkpoint_pth = "best_model.pth"
checkpoint = torch.load(checkpoint_pth)
best_acc = checkpoint['best_acc']
model.load_state_dict(checkpoint['state_dict'])
batch_size = 4
# DataLoader 加载数据集
train_dataset = torch.utils.data.TensorDataset(train, train_target)
train_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True)
validation_dataset = torch.utils.data.TensorDataset(validation, validation_target)
validation_loader = torch.utils.data.DataLoader(dataset=validation_dataset, batch_size=batch_size, shuffle=True)
'''
/****************************************************/
模型评估
/****************************************************/
'''
# ----------------------------------------------------#
# 评估
# ----------------------------------------------------#
# Validation
model.eval()
# 存储所有预测和真实标签
all_preds = []
all_labels = []
# 不计算梯度，减少计算和内存消耗
with torch.no_grad():
for data, labels in validation_loader:
# data和labels的预处理（如：转移到GPU、标准化等）...
# 生成预测并获取最可能的类别
outputs = model(data)
_, predicted = torch.max(outputs, 1)
# 收集预测和真实标签
all_preds.extend(predicted.cpu().numpy())
all_labels.extend(labels.cpu().numpy())
# 计算性能指标
accuracy = accuracy_score(all_labels, all_preds)
conf_matrix = confusion_matrix(all_labels, all_preds)
precision = precision_score(all_labels, all_preds, average='macro')
recall = recall_score(all_labels, all_preds, average='macro')
f1 = f1_score(all_labels, all_preds, average='macro')
print(f"Accuracy: {accuracy}")
print(f"Confusion Matrix:\n{conf_matrix}")
print(f"Precision: {precision}")
print(f"Recall: {recall}")
print(f"F1 Score: {f1}")2 运行结果
总结
到此，基于LSTM的序列分类任务就结束啦。
相比于CNN图像处理，序列分类任务训练是非常快的。
正如李沐老师说的：
"文本是廉价的数据。"
完整项目
SequenceClassifier