以Titanic乘客生存预测任务为例，进一步熟悉Ray Tune调参工具。

titanic数据集的目标是根据乘客信息预测他们在Titanic号撞击冰山沉没后能否生存。

本示例的基础代码参考了下面两篇文章：

• 1-1,结构化数据建模流程范例（一个不错的PyTorch教程）

• How to use Tune with PyTorch

也可以看一下上一篇文章：PyTorch + Ray Tune 调参

教程中的原始代码如下：

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import torch
from torch import nn
from torch.utils.data import Dataset,DataLoader,TensorDataset
from sklearn.metrics import accuracy_score
import datetimedftrain_raw = pd.read_csv('train.csv')
dftest_raw = pd.read_csv('test.csv')
dftrain_raw.head(10)def preprocessing(dfdata):dfresult= pd.DataFrame()#PclassdfPclass = pd.get_dummies(dfdata['Pclass'])dfPclass.columns = ['Pclass_' +str(x) for x in dfPclass.columns ]dfresult = pd.concat([dfresult,dfPclass],axis = 1)#SexdfSex = pd.get_dummies(dfdata['Sex'])dfresult = pd.concat([dfresult,dfSex],axis = 1)#Agedfresult['Age'] = dfdata['Age'].fillna(0)dfresult['Age_null'] = pd.isna(dfdata['Age']).astype('int32')#SibSp,Parch,Faredfresult['SibSp'] = dfdata['SibSp']dfresult['Parch'] = dfdata['Parch']dfresult['Fare'] = dfdata['Fare']#Carbindfresult['Cabin_null'] =  pd.isna(dfdata['Cabin']).astype('int32')#EmbarkeddfEmbarked = pd.get_dummies(dfdata['Embarked'],dummy_na=True)dfEmbarked.columns = ['Embarked_' + str(x) for x in dfEmbarked.columns]dfresult = pd.concat([dfresult,dfEmbarked],axis = 1)return(dfresult)x_train = preprocessing(dftrain_raw).values
y_train = dftrain_raw[['Survived']].valuesx_test = preprocessing(dftest_raw).values
y_test = dftest_raw[['Survived']].valuesprint("x_train.shape =", x_train.shape )
print("x_test.shape =", x_test.shape )print("y_train.shape =", y_train.shape )
print("y_test.shape =", y_test.shape )dl_train = DataLoader(TensorDataset(torch.tensor(x_train).float(),torch.tensor(y_train).float()),shuffle = True, batch_size = 8)
dl_valid = DataLoader(TensorDataset(torch.tensor(x_test).float(),torch.tensor(y_test).float()),shuffle = False, batch_size = 8)def create_net():net = nn.Sequential()net.add_module("linear1", nn.Linear(15, 20))net.add_module("relu1", nn.ReLU())net.add_module("linear2", nn.Linear(20, 15))net.add_module("relu2", nn.ReLU())net.add_module("linear3", nn.Linear(15, 1))net.add_module("sigmoid", nn.Sigmoid())return netnet = create_net()
print(net)loss_func = nn.BCELoss()
optimizer = torch.optim.Adam(params=net.parameters(),lr = 0.01)
metric_func = lambda y_pred,y_true: accuracy_score(y_true.data.numpy(),y_pred.data.numpy()>0.5)
metric_name = "accuracy"epochs = 10
log_step_freq = 30dfhistory = pd.DataFrame(columns = ["epoch","loss",metric_name,"val_loss","val_"+metric_name])
print("Start Training...")
nowtime = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
print("=========="*8 + "%s"%nowtime)for epoch in range(1, epochs + 1):# 1，训练循环-------------------------------------------------net.train()loss_sum = 0.0metric_sum = 0.0step = 1for step, (features, labels) in enumerate(dl_train, 1):# 梯度清零optimizer.zero_grad()# 正向传播求损失predictions = net(features)loss = loss_func(predictions, labels)metric = metric_func(predictions, labels)# 反向传播求梯度loss.backward()optimizer.step()# 打印batch级别日志loss_sum += loss.item()metric_sum += metric.item()if step % log_step_freq == 0:print(("[step = %d] loss: %.3f, " + metric_name + ": %.3f") %(step, loss_sum / step, metric_sum / step))# 2，验证循环-------------------------------------------------net.eval()val_loss_sum = 0.0val_metric_sum = 0.0val_step = 1for val_step, (features, labels) in enumerate(dl_valid, 1):# 关闭梯度计算with torch.no_grad():predictions = net(features)val_loss = loss_func(predictions, labels)val_metric = metric_func(predictions, labels)val_loss_sum += val_loss.item()val_metric_sum += val_metric.item()# 3，记录日志-------------------------------------------------info = (epoch, loss_sum / step, metric_sum / step,val_loss_sum / val_step, val_metric_sum / val_step)dfhistory.loc[epoch - 1] = info# 打印epoch级别日志print(("\nEPOCH = %d, loss = %.3f," + metric_name + \"  = %.3f, val_loss = %.3f, " + "val_" + metric_name + " = %.3f")% info)nowtime = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')print("\n" + "==========" * 8 + "%s" % nowtime)print('Finished Training...')

运行效果如下：

x_train.shape = (712, 15)
x_test.shape = (179, 15)
y_train.shape = (712, 1)
y_test.shape = (179, 1)
Sequential((linear1): Linear(in_features=15, out_features=20, bias=True)(relu1): ReLU()(linear2): Linear(in_features=20, out_features=15, bias=True)(relu2): ReLU()(linear3): Linear(in_features=15, out_features=1, bias=True)(sigmoid): Sigmoid()
)
Start Training...
================================================================================2021-01-07 10:57:41
[step = 30] loss: 0.706, accuracy: 0.613
[step = 60] loss: 0.641, accuracy: 0.673EPOCH = 1, loss = 0.653,accuracy  = 0.660, val_loss = 0.595, val_accuracy = 0.745================================================================================2021-01-07 10:57:42
[step = 30] loss: 0.593, accuracy: 0.700
[step = 60] loss: 0.591, accuracy: 0.700EPOCH = 2, loss = 0.574,accuracy  = 0.719, val_loss = 0.472, val_accuracy = 0.772================================================================================2021-01-07 10:57:42
[step = 30] loss: 0.574, accuracy: 0.767
[step = 60] loss: 0.537, accuracy: 0.769EPOCH = 3, loss = 0.525,accuracy  = 0.772, val_loss = 0.454, val_accuracy = 0.804================================================================================2021-01-07 10:57:42
[step = 30] loss: 0.523, accuracy: 0.783
[step = 60] loss: 0.503, accuracy: 0.794EPOCH = 4, loss = 0.506,accuracy  = 0.792, val_loss = 0.473, val_accuracy = 0.804================================================================================2021-01-07 10:57:42
[step = 30] loss: 0.492, accuracy: 0.796
[step = 60] loss: 0.482, accuracy: 0.802EPOCH = 5, loss = 0.485,accuracy  = 0.796, val_loss = 0.429, val_accuracy = 0.799================================================================================2021-01-07 10:57:42
[step = 30] loss: 0.461, accuracy: 0.808
[step = 60] loss: 0.461, accuracy: 0.798EPOCH = 6, loss = 0.471,accuracy  = 0.792, val_loss = 0.454, val_accuracy = 0.799================================================================================2021-01-07 10:57:42
[step = 30] loss: 0.467, accuracy: 0.779
[step = 60] loss: 0.490, accuracy: 0.762EPOCH = 7, loss = 0.464,accuracy  = 0.787, val_loss = 0.401, val_accuracy = 0.810================================================================================2021-01-07 10:57:42
[step = 30] loss: 0.444, accuracy: 0.804
[step = 60] loss: 0.444, accuracy: 0.808EPOCH = 8, loss = 0.442,accuracy  = 0.808, val_loss = 0.446, val_accuracy = 0.777================================================================================2021-01-07 10:57:42
[step = 30] loss: 0.460, accuracy: 0.775
[step = 60] loss: 0.458, accuracy: 0.792EPOCH = 9, loss = 0.440,accuracy  = 0.802, val_loss = 0.429, val_accuracy = 0.810================================================================================2021-01-07 10:57:42
[step = 30] loss: 0.436, accuracy: 0.808
[step = 60] loss: 0.438, accuracy: 0.806EPOCH = 10, loss = 0.439,accuracy  = 0.803, val_loss = 0.418, val_accuracy = 0.810================================================================================2021-01-07 10:57:43
Finished Training...

回顾上一篇博客，要使用Ray Tune进行调参，我们要做以下几件事：

• 将数据加载和训练过程封装到函数中；
• 使一些网络参数可配置；
• 增加检查点（可选）；
• 定义模型参数搜索空间。

下面我们看一下按照上面的要求怎么修改代码：

import os
import pandas as pd
import torch
from torch import nn
from functools import partial
from torch.utils.data import random_split
from torch.utils.data import DataLoader, TensorDataset
from sklearn.metrics import accuracy_scorefrom ray import tune
from ray.tune import CLIReporter
from ray.tune.schedulers import ASHAScheduler# 执行数据预处理功能
def preprocessing(dfdata):dfresult = pd.DataFrame()# PclassdfPclass = pd.get_dummies(dfdata['Pclass'])dfPclass.columns = ['Pclass_' + str(x) for x in dfPclass.columns]dfresult = pd.concat([dfresult, dfPclass], axis=1)# SexdfSex = pd.get_dummies(dfdata['Sex'])dfresult = pd.concat([dfresult, dfSex], axis=1)# Agedfresult['Age'] = dfdata['Age'].fillna(0)dfresult['Age_null'] = pd.isna(dfdata['Age']).astype('int32')# SibSp,Parch,Faredfresult['SibSp'] = dfdata['SibSp']dfresult['Parch'] = dfdata['Parch']dfresult['Fare'] = dfdata['Fare']# Carbindfresult['Cabin_null'] = pd.isna(dfdata['Cabin']).astype('int32')# EmbarkeddfEmbarked = pd.get_dummies(dfdata['Embarked'], dummy_na=True)dfEmbarked.columns = ['Embarked_' + str(x) for x in dfEmbarked.columns]dfresult = pd.concat([dfresult, dfEmbarked], axis=1)return (dfresult)# 执行数据加载功能的函数
def load_data(train_dir='F:/Data_Analysis/Visualization/7_ray_tune/train.csv',test_dir='F:/Data_Analysis/Visualization/7_ray_tune/test.csv'):dftrain_raw = pd.read_csv(train_dir)dftest_raw = pd.read_csv(test_dir)x_train = preprocessing(dftrain_raw).valuesy_train = dftrain_raw[['Survived']].valuesx_test = preprocessing(dftest_raw).valuesy_test = dftest_raw[['Survived']].valuestrainset = TensorDataset(torch.tensor(x_train).float(), torch.tensor(y_train).float())testset = TensorDataset(torch.tensor(x_test).float(), torch.tensor(y_test).float())return trainset, testset# 定义神经网络模型
class Net(nn.Module):def __init__(self, l1=20, l2=15):super(Net, self).__init__()self.net = nn.Sequential(nn.Linear(15, l1),nn.ReLU(),nn.Linear(l1, l2),nn.ReLU(),nn.Linear(l2, 1),nn.Sigmoid())def forward(self, x):x = self.net(x)return xmetric_func = lambda y_pred,y_true: accuracy_score(y_true.data.numpy(),y_pred.data.numpy()>0.5)# 执行模型训练过程的函数
def train_titanic(config, checkpoint_dir=None, train_dir=None, test_dir=None):net = Net(config["l1"], config["l2"])       # 可配置参数device = "cpu"net.to(device)loss_func = nn.BCELoss()optimizer = torch.optim.Adam(params=net.parameters(), lr=config["lr"])      # 可配置参数if checkpoint_dir:# 模型的状态、优化器的状态model_state, optimizer_state = torch.load(os.path.join(checkpoint_dir, "checkpoint"))net.load_state_dict(model_state)optimizer.load_state_dict(optimizer_state)# numpy.ndarraytrainset, testset = load_data(train_dir, test_dir)train_abs = int(len(trainset) * 0.8)train_subset, val_subset = random_split(trainset, [train_abs, len(trainset) - train_abs])trainloader = torch.utils.data.DataLoader(train_subset,batch_size=int(config["batch_size"]),       # 可配置参数shuffle=True,num_workers=4)valloader = torch.utils.data.DataLoader(val_subset,batch_size=int(config["batch_size"]),       # 可配置参数shuffle=True,num_workers=4)for epoch in range(1, 11):net.train()loss_sum = 0.0metric_sum = 0.0step = 1for step, (features, labels) in enumerate(trainloader, 1):# 梯度清零optimizer.zero_grad()# 正向传播求损失predictions = net(features)loss = loss_func(predictions, labels)metric = metric_func(predictions, labels)# 反向传播求梯度loss.backward()optimizer.step()net.eval()val_loss_sum = 0.0val_metric_sum = 0.0val_step = 1for val_step, (features, labels) in enumerate(valloader, 1):# 关闭梯度计算with torch.no_grad():predictions = net(features)val_loss = loss_func(predictions, labels)val_metric = metric_func(predictions, labels)val_loss_sum += val_loss.item()val_metric_sum += val_metric.item()with tune.checkpoint_dir(epoch) as checkpoint_dir:path = os.path.join(checkpoint_dir, "checkpoint")torch.save((net.state_dict(), optimizer.state_dict()), path)# 打印平均损失和平均精度tune.report(loss=(val_loss_sum / val_step), accuracy=(val_metric_sum / val_step))print("Finished Training")# 执行测试功能的函数
def test_accuracy(net, device="cpu"):trainset, testset = load_data()testloader = DataLoader(testset, shuffle = False, batch_size = 4, num_workers=6)loss_func = nn.BCELoss()net.eval()test_loss_sum = 0.0test_metric_sum = 0.0test_step = 1for test_step, (features, labels) in enumerate(testloader, 1):# 关闭梯度计算with torch.no_grad():predictions = net(features)test_loss = loss_func(predictions, labels)test_metric = metric_func(predictions, labels)test_loss_sum += test_loss.item()test_metric_sum += test_metric.item()return test_loss_sum / test_step, test_metric_sum / test_step# 主程序
def main(num_samples=10, max_num_epochs=10):checkpoint_dir = 'F:/Data_Analysis/Visualization/7_ray_tune/checkpoints/'train_dir = 'F:/Data_Analysis/Visualization/7_ray_tune/train.csv'test_dir = 'F:/Data_Analysis/Visualization/7_ray_tune/test.csv'load_data(train_dir, test_dir)# 定义模型参数搜索空间config = {"l1": tune.choice([8, 16, 32, 64]),"l2": tune.choice([8, 16, 32, 64]),"lr": tune.choice([0.1, 0.01, 0.001, 0.0001]),"batch_size": tune.choice([2, 4, 8, 16])}# ASHAScheduler会根据指定标准提前中止坏实验scheduler = ASHAScheduler(metric="loss",mode="min",max_t=max_num_epochs,grace_period=1,reduction_factor=2)# 在命令行打印实验报告reporter = CLIReporter(# parameter_columns=["l1", "l2", "lr", "batch_size"],metric_columns=["loss", "accuracy", "training_iteration"])# 执行训练过程result = tune.run(partial(train_titanic, checkpoint_dir=checkpoint_dir, train_dir=train_dir, test_dir=test_dir),# 指定训练资源resources_per_trial={"cpu": 6},config=config,num_samples=num_samples,scheduler=scheduler,progress_reporter=reporter)# 找出最佳实验best_trial = result.get_best_trial("loss", "min", "last")# 打印最佳实验的参数配置print("Best trial config: {}".format(best_trial.config))print("Best trial final validation loss: {}".format(best_trial.last_result["loss"]))print("Best trial final validation accuracy: {}".format(best_trial.last_result["accuracy"]))# 打印最优超参数组合对应的模型在测试集上的性能best_trained_model = Net(best_trial.config["l1"], best_trial.config["l2"])device = "cpu"best_trained_model.to(device)best_checkpoint_dir = best_trial.checkpoint.valuemodel_state, optimizer_state = torch.load(os.path.join(best_checkpoint_dir, "checkpoint"))best_trained_model.load_state_dict(model_state)test_loss, test_acc = test_accuracy(best_trained_model, device)print("Best trial test set loss: {}".format(test_loss))print("Best trial test set accuracy: {}".format(test_acc))if __name__ == "__main__":# You can change the number of GPUs per trial here:main(num_samples=10, max_num_epochs=10)

效果如下：