完整代码在个人主页简介链接pytorch路径下可找到
1 单车预测器1.0
1.1 人工神经元
对于sigmoid函数来说,w控制函数曲线的方向,b控制曲线水平方向位移,w'控制曲线在y方向的幅度
1.2 多个人工神经元
模型如下
数学上可证,有限神经元绘制的曲线可以逼近任意有限区间内的曲线(闭区间连续函数有界)
1.3 模型与代码
通过训练可得到逼近真实曲线的神经网络参数
通过梯度下降法寻找局部最优(如何寻找全局最优后面考虑)
思考 n个峰需在一个隐层要多少隐单元?材料说3个峰10个单元就够了,理论上算,最少需要5个,可能保险起见,加其他一些不平滑处,就弄了10个
初次代码如下
from os import path
import numpy as np
import pandas as pd
import torch
import torch.optim as optim
import matplotlib.pyplot as plotDATA_PATH = path.realpath('pytorch/jizhi/bike/data/hour.csv')class Bike():def exec(self):self.prepare_data_and_params()self.train()def prepare_data_and_params(self):self.data = pd.read_csv(DATA_PATH)counts = self.data['cnt'][:50]self.x = torch.FloatTensor(np.arange(len(counts)))self.y = torch.FloatTensor(np.array(counts, dtype=float))self.size = 10self.weights = torch.randn((1, self.size), requires_grad=True)self.biases = torch.randn((self.size), requires_grad=True)self.weights2 = torch.randn((self.size, 1), requires_grad=True)def train(self):rate = 0.001losses = []x, y = self.x.view(50, -1), self.y.view(50, -1) # reshapefor num in range(30000):hidden = x * self.weights + self.biaseshidden = torch.sigmoid(hidden)predictions = hidden.mm(self.weights2)loss = torch.mean((predictions - y) ** 2)losses.append(loss.data.numpy())if num % 3000 == 0:print(f'loss: {loss}')loss.backward()self.weights.data.add_(- rate * self.weights.grad.data)self.biases.data.add_(- rate * self.biases.grad.data)self.weights2.data.add_(- rate * self.weights2.grad.data)self.weights.grad.data.zero_()self.biases.grad.data.zero_()self.weights2.grad.data.zero_()# plot loss#plot.plot(losses)#plot.xlabel('epoch')#plot.ylabel('loss')#plot.show()# plot predictx_data = x.data.numpy()plot.figure(figsize=(10, 7))xplot, = plot.plot(x_data, y.data.numpy(), 'o')yplot, = plot.plot(x_data, predictions.data.numpy())plot.xlabel('x')plot.ylabel('y')plot.legend([xplot, yplot], ['Data', 'prediction with 30000 epoch'])plot.show()def main():Bike().exec()if __name__ == '__main__':main()
拟合有问题,原因是拟合次数不够,为啥不够?从sklearn学习了解到,神经网络对输入参数敏感,一般来说需要对数据做标准化处理。具体来说,第一个隐层输出范围变成-50-50,0.0001学习率情况下100000次也不够,可以对数据做预处理,减小x跨度,变为0-1,可加快训练速度,进行如下改动再次训练
self.x = torch.FloatTensor(np.arange(len(counts))) / len(counts)
正确了,再取50个点预测一下
def predict_and_plot(self):counts_predict = self.data['cnt'][50:100]x = torch.FloatTensor((np.arange(len(counts_predict), dtype=float) + 50) / 100)y = torch.FloatTensor(np.array(counts_predict, dtype=float))# num multiply replace matrix multiplyhidden = x.expand(self.size, len(x)).t() * self.weights.expand(len(x), self.size)hidden = torch.sigmoid(hidden)predictions = hidden.mm(self.weights2)loss = torch.mean((predictions - y) ** 2)print(f'loss: {loss}')x_data = x.data.numpy()plot.figure(figsize=(10, 7))xplot, = plot.plot(x_data, y.data.numpy(), 'o')yplot, = plot.plot(x_data, predictions.data.numpy())plot.xlabel('x')plot.ylabel('y')plot.legend([xplot, yplot], ['data', 'prediction'])plot.show()
预测失败,可能是过拟合
2 单车预测器2.0
2.1 数据预处理
通过上节学习和之前写的sklearn博客发现,神经网络训练前需要预处理数据,主要有1数值型变量需要范围标准化2数值型类型变量需处理为onehot。标准化可用sklearn的scaler,也可手动标准化,类型变量可用pd.get_dummies操作。直接开始操作
def prepare_data_and_params_2(self):# type columns to dummyself.data = pd.read_csv(DATA_PATH)dummy_fields = ['season', 'weathersit', 'mnth', 'hr', 'weekday']for each in dummy_fields:dummies = pd.get_dummies(self.data[each], prefix=each, drop_first=False)self.data = pd.concat([self.data], dummies)drop_fields = ['season', 'weathersit', 'mnth', 'hr', 'weekday', 'instant', 'dteday', 'workingday', 'atemp']self.data = self.data.drop(drop_fields, axis=1)# decimal columns to scalerquant_features = ['cnt', 'temp', 'hum', 'windspeed']scaled_features = {}for each in quant_features:mean, std = self.data[each].mean(), self.data[each].std()scaled_features[each] = [mean, std]self.data.loc[:, each] = (self.data[each] - mean) / stdself.tr, self.te = self.data[:-21 * 24], self.data[-21 * 24:]target_fields = ['cnt', 'casual', 'registered']self.xtr, self.ytr = self.tr.drop(self.tr.drop[target_fields], axis=1), self.tr[target_fields]self.xte, self.yte = self.te.drop(self.te.drop[target_fields], axis=1), self.te[target_fields]self.x = self.xtr.valuesy = self.ytr.values.astype(float)self.y = np.reshape(y, [len(y), 1]) self.loss = []
2.2 构造神经网络
def train_and_plot2(self):input_size = self.xtr.shape[1]hidden_size=10output_size=1batch_size=128neu = torch.nn.Sequential(torch.nn.Linear(input_size, hidden_size),torch.nn.Sigmoid(),torch.nn.Linear(hidden_size, output_size))cost = torch.nn.MSELoss()optimizer = torch.optim.SGD(neu.parameters(), lr=0.01)
2.3 数据批处理
为啥要批处理?如果数据太多,每个iter直接处理所有数据会比较慢
for i in range(1000):batch_loss = []for start in range(0, len(self.x), batch_size):end = start + batch_size if start + batch_size < len(self.x) else len(self.x)xx = torch.FloatTensor(self.x[start:end])yy = torch.FloatTensor(self.y[start:end])predictions = neu(xx)loss = cost(predictions, yy)optimizer.zero_grad()loss.backward()optimizer.step()batch_loss.append(loss.data.numpy())if i % 100 == 0:self.loss.append(np.mean(batch_loss))print(i, np.mean(batch_loss))plot.plot(np.arange(len(self.loss)) * 100, self.loss)plot.xlabel('epoch')plot.ylabel('MSE')plot.show()
2.4 测试神经网络
原始数据是从2011-2012两个完整年,按教材,取2012最后21天作测试集预测
def predict_and_plot2(self):targets = self.yte['cnt']targets = targets.values.reshape([len(targets), 1]).astype(float)x = torch.FloatTensor(self.xte.values.astype(float))y = torch.FloatTensor(targets)predict = self.neu(x)predict = predict.data.numpy()fig, ax = plot.subplots(figsize=(10, 7))mean, std = self.scaled_features['cnt']ax.plot(predict * std + mean, label='prediction')ax.plot(targets * std + mean, label='data')ax.legend()ax.set_xlabel('date-time')ax.set_ylabel('counts')dates = pd.to_datetime(self.rides.loc[self.te.index]['dteday'])dates = dates.apply(lambda d: d.strftime('%b %d'))ax.set_xticks(np.arange(len(dates))[12::24])ax.set_xticklabels(dates[12::24], rotation=45)plot.show()
发现2012最后21天前半段还行,后半段有差异,看日历发现临近圣诞节,可能不能用正常日程预测
2.5 改进与分析(重要)
这节有啥用?上节圣诞节预测不准,为啥?这节可以通过分析神经网络回答这个问题
怎么分析?本节主要通过分析神经网络参数来在底层寻找原因,帮助分析问题
在异常处将多个神经源绘制独自的曲线,绘制其图像,分析找原因,比如趋势相同,趋势相反这种曲线,重点分析对象。适用于神经元较少,可以一个一个神经元看,多了就不行了