pytorch使用Ray-tune对原有训练模型的代码改写，自动调参（一）

借助Ray-tune可以对pytorch自动调参，下面就一步步地改写，从原始的训练代码慢慢变为可以自动调参的代码的教程• 保姆级：

安装依赖

pip install -i https://mirrors.aliyun.com/pypi/simple/ 'ray[default]'

注意安装的内容是'ray[default]'，因为后来的各版本代码有调整，因此统一指定成了这种安装方式，否则代码运行中会有各种warnning。

pytorch模型改写至自动调参

为了增加学习效率，这里举一个没有加任何额外的buff的例子。

这里用随机数生成训练数据，然后随手撘一个三层感知机的模型，运行一下确保跑通即可。

步骤1：原有代码要跑通

假设原有的训练代码如下：

import torch
import torch.nn as nn
import numpy as np


class LinearRegressionModel(nn.Module):
    def __init__(self, input_shape, linear1, linear2, output_shape):
        super(LinearRegressionModel, self).__init__()
        self.linear1 = nn.Linear(input_shape, linear1)
        self.linear2 = nn.Linear(linear1, linear2)
        self.linear3 = nn.Linear(linear2, output_shape)

    def forward(self, x):
        l1 = self.linear1(x)
        l2 = self.linear2(l1)
        l3 = self.linear3(l2)
        return l3


def train_model(x_train, y_train, linear1, linear2):
    # 指定参数与损失函数
    model = LinearRegressionModel(x_train.shape[1], linear1, linear2, 1)
    epochs = 1000  # 迭代1000次
    learning_rate = 0.01  # 学习率
    optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)  # 优化函数
    criterion = nn.MSELoss()  # Loss使用MSE值，目标是使MSE最小

    loss_list = []
    for epoch in range(epochs):
        epoch += 1
        optimizer.zero_grad()  # 梯度清零
        outputs = model(x_train)  # 前向传播
        loss = criterion(outputs, y_train)  # 计算损失
        loss.backward()  # 返向传播
        loss_list.append(loss.detach().numpy())
        optimizer.step()  # 更新权重参数
    mean_loss = np.mean(loss_list)
    print("loss: ", mean_loss)


if __name__ == '__main__':
    x_train = torch.randn(100, 4)  # 生成100个4维的随机数，作为训练集的 X
    y_train = torch.randn(100, 1)  # 作为训练集的label
    train_model(x_train, y_train, 32, 8)

代码一定要跑通无报错哈！要不然会无限报错。

跑通模型后我们得到一个值，可能是loss: 0.6363947，我们之后的目的是希望找到一组参数linear1 与 linear2（模型中代表隐层数量），使这个loss值越小越好！

步骤2：修改原有代码

只需要有如下修改即可：

1. 导包from ray import tune
2. 执行依次模型后，返回一个或多个指标，这个指标只能是一个值，作为优化的目标
3. 所有的print()操作都要删掉，这个不会影响逻辑，但是会影响控制台的显示效果

一共有3处修改地点，修改后的代码如下：

import torch
import torch.nn as nn
import numpy as np
from ray import tune  # 修改地方1：导包


class LinearRegressionModel(nn.Module):
    def __init__(self, input_shape, linear1, linear2, output_shape):
        super(LinearRegressionModel, self).__init__()
        self.linear1 = nn.Linear(input_shape, linear1)
        self.linear2 = nn.Linear(linear1, linear2)
        self.linear3 = nn.Linear(linear2, output_shape)

    def forward(self, x):
        l1 = self.linear1(x)
        l2 = self.linear2(l1)
        l3 = self.linear3(l2)
        return l3


def train_model(x_train, y_train, linear1, linear2):
    # 指定参数与损失函数
    model = LinearRegressionModel(x_train.shape[1], linear1, linear2, 1)
    epochs = 1000  # 迭代1000次
    learning_rate = 0.01  # 学习率
    optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)  # 优化函数
    criterion = nn.MSELoss()  # Loss使用MSE值，目标是使MSE最小

    loss_list = []
    for epoch in range(epochs):
        epoch += 1
        optimizer.zero_grad()  # 梯度清零
        outputs = model(x_train)  # 前向传播
        loss = criterion(outputs, y_train)  # 计算损失
        loss.backward()  # 返向传播
        loss_list.append(loss.detach().numpy())
        optimizer.step()  # 更新权重参数
    mean_loss = np.mean(loss_list)
    # print("loss: ", mean_loss) # 修改地方2：不要有任何print()操作影响控制台的观察效果
    tune.report(my_loss=mean_loss)  # 修改地方3：加入tune.report(xxx=value)，这里xxx可以自定义


if __name__ == '__main__':
    x_train = torch.randn(100, 4)  # 生成100个4维的随机数，作为训练集的 X
    y_train = torch.randn(100, 1)  # 作为训练集的label
    train_model(x_train, y_train, 32, 8)

步骤3：使用ray的调参代码封装训练代码

这一步主要是指定调参的参数范围和一些调参规则

步骤如下：

1. 修改Model传参的方式，使用config替代所有的参数列表
2. 把main中直接训练模型的代码注释掉，咱们要修改成使用ray启动的方法
3. 添加ray的基本配置，进行封装

修改后的代码与注释如下：

import torch
import torch.nn as nn
import numpy as np
from ray import tune


class LinearRegressionModel(nn.Module):
    def __init__(self, input_shape, linear1, linear2, output_shape):
        super(LinearRegressionModel, self).__init__()
        self.linear1 = nn.Linear(input_shape, linear1)
        self.linear2 = nn.Linear(linear1, linear2)
        self.linear3 = nn.Linear(linear2, output_shape)

    def forward(self, x):
        l1 = self.linear1(x)
        l2 = self.linear2(l1)
        l3 = self.linear3(l2)
        return l3


def train_model(config):  # 修改1：修改参数，所有的参数都要借助config传递
    # 指定参数与损失函数
    model = LinearRegressionModel(x_train.shape[1], config['linear1'], config['linear2'], 1)
    epochs = 1000  # 迭代1000次
    learning_rate = 0.01  # 学习率
    optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)  # 优化函数
    criterion = nn.MSELoss()  # Loss使用MSE值，目标是使MSE最小

    loss_list = []
    for epoch in range(epochs):
        epoch += 1
        optimizer.zero_grad()  # 梯度清零
        outputs = model(x_train)  # 前向传播
        loss = criterion(outputs, y_train)  # 计算损失
        loss.backward()  # 返向传播
        loss_list.append(loss.detach().numpy())
        optimizer.step()  # 更新权重参数
    mean_loss = np.mean(loss_list)
    tune.report(my_loss=mean_loss)


if __name__ == '__main__':
    x_train = torch.randn(100, 4)  # 生成100个4维的随机数，作为训练集的 X
    y_train = torch.randn(100, 1)  # 作为训练集的label
    # train_model(x_train, y_train, 32, 8) # 修改2：就不需要这样启动了，注释掉这一行

    # 修改3：下面就是封装的方法
    config = {
        "linear1": tune.sample_from(lambda _: np.random.randint(2, 5)),  # 自定义采样
        "linear2": tune.choice([2, 4, 8, 16]),  # 从给定值中随机选择
    }
    result = tune.run(  # 执行训练过程，执行到这里就会根据config自动调参了
        train_model,  # 要训练的模型
        resources_per_trial={"cpu": 8, },  # 指定训练资源
        config=config,
        num_samples=20,  # 迭代的次数
    )
    # 得到最后的结果
    print("======================== Result =========================")
    print(result.results_df)

跑通后可以得到如下结果：

+-------------------------+------------+-------+-----------+-----------+--------+------------------+-----------+
| Trial name              | status     | loc   |   linear1 |   linear2 |   iter |   total time (s) |   my_loss |
|-------------------------+------------+-------+-----------+-----------+--------+------------------+-----------|
| train_model_559fb_00000 | TERMINATED |       |         2 |         8 |      1 |         0.325345 |   1.05185 |
| train_model_559fb_00001 | TERMINATED |       |         3 |        16 |      1 |         0.32275  |   1.06859 |
| train_model_559fb_00002 | TERMINATED |       |         4 |         8 |      1 |         0.292687 |   1.04742 |
| train_model_559fb_00003 | TERMINATED |       |         2 |        16 |      1 |         0.291877 |   1.06596 |
| train_model_559fb_00004 | TERMINATED |       |         3 |         2 |      1 |         0.298721 |   1.04942 |
| train_model_559fb_00005 | TERMINATED |       |         4 |         2 |      1 |         0.294107 |   1.06227 |
| train_model_559fb_00006 | TERMINATED |       |         2 |        16 |      1 |         0.310592 |   1.04826 |
| train_model_559fb_00007 | TERMINATED |       |         4 |         2 |      1 |         0.31578  |   1.0608  |
| train_model_559fb_00008 | TERMINATED |       |         3 |         2 |      1 |         0.286066 |   1.05879 |
| train_model_559fb_00009 | TERMINATED |       |         3 |         8 |      1 |         0.290412 |   1.05573 |
| train_model_559fb_00010 | TERMINATED |       |         2 |         2 |      1 |         0.282055 |   1.04826 |
| train_model_559fb_00011 | TERMINATED |       |         4 |         4 |      1 |         0.288696 |   1.05269 |
| train_model_559fb_00012 | TERMINATED |       |         2 |         2 |      1 |         0.311075 |   1.06867 |
| train_model_559fb_00013 | TERMINATED |       |         3 |         2 |      1 |         0.312876 |   1.06269 |
| train_model_559fb_00014 | TERMINATED |       |         3 |         2 |      1 |         0.273914 |   1.06514 |
| train_model_559fb_00015 | TERMINATED |       |         4 |         2 |      1 |         0.270813 |   1.05434 |
| train_model_559fb_00016 | TERMINATED |       |         4 |        16 |      1 |         0.292606 |   1.06485 |
| train_model_559fb_00017 | TERMINATED |       |         3 |        16 |      1 |         0.286667 |   1.05534 |
| train_model_559fb_00018 | TERMINATED |       |         4 |         8 |      1 |         0.272169 |   1.06625 |
| train_model_559fb_00019 | TERMINATED |       |         2 |         8 |      1 |         0.270342 |   1.0556  |
+-------------------------+------------+-------+-----------+-----------+--------+------------------+-----------+

最后的result.results_df记录了所有的参数与结果

在结尾处打个断点还能看到可视化的图表：View the Ray dashboard at http://127.0.0.1:8265，点进连接就可以看到运行状态了，可以通过这个连接拆看运行中的调参程序的运行状态。

程序跑通了就是信心十足，之后会解锁更多调参的内容与技巧