【小笔记】用tsai库实现Rocket家族算法

2024.1.16
Rocket家族算法是用于时间序列分类的强baseline（性能比较参考【小笔记】时序数据分类算法最新小结），Rocket/MiniRocket/MultiRocket官方都有开源实现，相比较而言，用tsai来实现有三个好处：1是快速跑通模型；2是更简洁优雅；3是掌握一个框架能举一反三。

1.tsai简介

项目：https://github.com/timeseriesAI/tsai

简介：
用于处理时间序列的工具库，包含TCN、Rockert等众多时间序列处理算法

安装：

pip install tsai

2.Rocket：最优雅的实现

这个例子是基于UCR的Beef数据集，运行时，会自动下载数据集到项目的data路径下

from tsai.all import *
from sklearn.linear_model import RidgeClassifierCV
from dsets_build import get_my_dsets


device = 'cuda' if torch.cuda.is_available() else 'cpu'
print(device)

# 加载UCR数据集
X, y, splits = get_UCR_data('Beef', return_split=False, on_disk=True, verbose=True)
tfms  = [None, [Categorize()]]
batch_tfms = [TSStandardize(by_sample=True)]
dsets = TSDatasets(X, y, tfms=tfms, splits=splits)

# 标准示例
dls = TSDataLoaders.from_dsets(dsets.train, dsets.valid, bs=768, drop_last=False, shuffle_train=False,
                               device=device,
                               batch_tfms=[TSStandardize(by_sample=True)])
model = create_model(ROCKET, dls=dls)
# model = model.to(device)

print("构造特征...")
X_train, y_train = create_rocket_features(dls.train, model, verbose=False)
X_valid, y_valid = create_rocket_features(dls.valid, model, verbose=False)
print(X_train.shape, X_valid.shape)

print("基于特征开始训练...")
ridge = RidgeClassifierCV(alphas=np.logspace(-8, 8, 17))
ridge.fit(X_train, y_train)
print(f'alpha: {ridge.alpha_:.2E}  train: {ridge.score(X_train, y_train):.5f}  valid: {ridge.score(X_valid, y_valid):.5f}')

3.MiniRocket:（比Rocket更快）

待补充

4.MultiRocket:（比MiniRocket更强）

待补充

5.Hydra-MultiRocket:（Rocket家最强王者）

待补充

6.用自己的数据集训练模型

上面的例子都是用的UCR数据集，若要用自己的数据集进行训练怎么解决？
官方教程：
tsai-main\tutorial_nbs路径下的00c_Time_Series_data_preparation.ipynb

我总结了一下，基于单变量时间序列构建数据集就是下面这样：
dsets_build.py

from tsai.all import *
import numpy as np
import pandas as pd

def get_my_dsets():
    # 导入数据集
    train_data, valid_data, test_data = [[], []], [[], []], [[], []]
    radio_train, radio_valid, radio_test = 0.6, 0.2, 0.2
	
	# !这是我的读取读取例子,读者需要进行替换----------------------------------
    path = "train.csc"
	data = pd.read_csv(path)	

    train_data[0] = data['x'].tolist()
    train_data[1] = data['y'].tolist()
	# -----------------------------------------------------------------------
	
	# 将数据转换为np.array即可，剩下的都是通用了
    X_2d, y = np.array(train_data[0]), np.array(train_data[1])
    print(X_2d.shape, y.shape)    # (4000, 4096) (4000,)

    splits = get_splits(y, valid_size=0.2, stratify=True, random_state=23, shuffle=True, show_plot=False)
    print(splits)

    tfms = [None, [Categorize()]]
    dsets = TSDatasets(X_2d, y, tfms=tfms, splits=splits, inplace=True)
    print(dsets)
    return dsets

将数据集转换为tsai的dsets后，就可以直接用于训练模型了。

from tsai.all import *
from sklearn.linear_model import RidgeClassifierCV
from dsets_build import get_my_dsets


device = 'cuda' if torch.cuda.is_available() else 'cpu'
print(device)

# 加载UCR数据集
# X, y, splits = get_UCR_data('Beef', return_split=False, on_disk=True, verbose=True)
# tfms  = [None, [Categorize()]]
# batch_tfms = [TSStandardize(by_sample=True)]
# dsets = TSDatasets(X, y, tfms=tfms, splits=splits)

# 加载自定义的数据集
dsets = get_my_dsets()         # 和Rockert例子只有这里的区别

# 标准示例
dls = TSDataLoaders.from_dsets(dsets.train, dsets.valid, bs=768, drop_last=False, shuffle_train=False,
                               device=device,
                               batch_tfms=[TSStandardize(by_sample=True)])
model = create_model(ROCKET, dls=dls)
# model = model.to(device)

print("构造rocket特征...")
X_train, y_train = create_rocket_features(dls.train, model, verbose=False)
X_valid, y_valid = create_rocket_features(dls.valid, model, verbose=False)
print(X_train.shape, X_valid.shape)

print("基于特征开始训练...")
ridge = RidgeClassifierCV(alphas=np.logspace(-8, 8, 17))
ridge.fit(X_train, y_train)
print(f'alpha: {ridge.alpha_:.2E}  train: {ridge.score(X_train, y_train):.5f}  valid: {ridge.score(X_valid, y_valid):.5f}')