教你用pytorch训练五子棋ai示例代码

 
class game:
    def __init__(self, h, w):
        # 行数
        self.h = h
        # 列数
        self.w = w
        # 棋盘
        self.l = [['-' for _ in range(w)] for _ in range(h)]
        # 当前玩家 - 表示空 x先下 然后是o
        self.cur = 'x'
        # 游戏胜利者
        self.win_user = none
 
    # 检查下完这步后有没有赢 y是行 x是列 返回true表示赢
    def check_win(self, y, x):
        directions = [
            # 水平、垂直、两个对角线方向
            (1, 0), (0, 1), (1, 1), (1, -1)
        ]
        player = self.l[y][x]
        for dy, dx in directions:
            count = 0
            # 检查四个方向上的连续相同棋子
            for i in range(-4, 5):  # 检查-4到4的范围，因为五子连珠需要5个棋子
                ny, nx = y + i * dy, x + i * dx
                if 0 <= ny < self.h and 0 <= nx < self.w and self.l[ny][nx] == player:
                    count += 1
                    if count == 5:
                        return true
                else:
                    count = 0
        return false
 
    # 检查能不能下这里 y行 x列 返回true表示能下
    def check(self, y, x):
        return self.l[y][x] == '-' and self.win_user is none
 
    # 打印棋盘 可视化用得到
    def __str__(self):
        # 确定行号和列号的宽度
        row_width = len(str(self.h - 1))
        col_width = len(str(self.w - 1))
        
        # 生成带有行号和列号的棋盘字符串表示
        result = []
        # 添加列号标题
        result.append(' ' * (row_width + 1) + ' '.join(f'{i:>{col_width}}' for i in range(self.w)))
        # 添加分隔线（可选）
        result.append(' ' * (row_width + 1) + '-' * (col_width * self.w))
        # 添加棋盘行
        for y, row in enumerate(self.l):
            # 添加行号
            result.append(f'{y:>{row_width}} ' + ' '.join(f'{cell:>{col_width}}' for cell in row))
        return '\n'.join(result)
 
    # 一步棋
    def set(self, y, x):
        if self.win_user or not self.check(y, x):
            return false
        self.l[y][x] = self.cur
        if self.check_win(y, x):
            self.win_user = self.cur
            return true
        self.cur = 'x' if self.cur == 'o' else 'o'
        return true
    #和棋
    def heqi(self):
        for y in range(self.h):
            for x in range(self.w):
                if self.l[y][x]=='-':
                    return false
        return true
    

#玩家自己下
def run_game01():
    g = game(15, 15)
    while not g.win_user:
        # 打印当前棋盘状态
        while 1:
            print(g)
            try:
                y,x=input(g.cur+':').split(',')
                x=int(x)
                y=int(y)
                if g.set(y,x):
                    break
            except exception as e:
                print(e)
    print(g)
    print('胜利者',g.win_user)
 
 
 

import torch
import torch.nn as nn
import torch.optim as optim
from game import game

class mymod(nn.module):
    def __init__(self, input_channels=1, output_size=15*15):
        super(mymod, self).__init__()
        
        # 定义卷积层，用于提取特征
        self.conv1 = nn.conv2d(input_channels, 32, kernel_size=3, padding=1)  # 输出 32 x 15 x 15
        self.conv2 = nn.conv2d(32, 64, kernel_size=3, padding=1)  # 输出 64 x 15 x 15
        self.conv3 = nn.conv2d(64, 128, kernel_size=3, padding=1)  # 输出 128 x 15 x 15
        
        # 定义全连接层，用于最后的得分预测
        self.fc1 = nn.linear(128 * 15 * 15, 1024)  # 展平后传入全连接层
        self.fc2 = nn.linear(1024, output_size)  # 输出 15*15 的得分预测

    def forward(self, x):
        # 卷积层 -> 激活函数 -> 最大池化
        x = torch.relu(self.conv1(x))
        x = torch.relu(self.conv2(x))
        x = torch.relu(self.conv3(x))
        
        # 将卷积层输出展平为一维
        x = x.view(x.size(0), -1)
        
        # 全连接层
        x = torch.relu(self.fc1(x))
        x = self.fc2(x)
        return x
    # 保存模型权重
    def save(self, path):
        torch.save(self.state_dict(), path)

    # 加载模型权重
    def load(self, path):
        self.load_state_dict(torch.load(path))

#改进一下  output 把有棋子的地方的概率=0避免下这些地方
# 输入game对象和mymod对象，用于得到概率最大的落棋点 (行y, 列x)
def input_qi(g: game, m: mymod):
    # 获取当前棋盘状态
    board_state = g.l  # 使用 game.l 获取当前棋盘的状态 (15x15的二维列表)
    
    # 将棋盘状态转换为pytorch的tensor并增加一个维度（batch_size = 1）
    board_tensor = torch.tensor([[1 if cell == 'x' else -1 if cell == 'o' else 0 for cell in row] for row in board_state], 
                                 dtype=torch.float32).unsqueeze(0).unsqueeze(0)  # 形状变为 (1, 1, 15, 15)
    
    # 传入模型获取每个位置的得分
    output = m(board_tensor)
    
    # 将输出转为概率值（可以使用softmax来归一化）
    probabilities = torch.softmax(output, dim=-1).view(g.h, g.w).detach().numpy()  # 变为 (15, 15) 大小
    
    # 将已有棋子的位置的概率设置为 -inf，避免选择这些位置
    for y in range(g.h):
        for x in range(g.w):
            if board_state[y][x] != '-':
                probabilities[y, x] = -float('inf')  # 设置已经有棋子的地方的概率为 -inf
    
    # 找到概率最大的落子点
    max_prob_pos = divmod(probabilities.argmax(), g.w)  # 得到最大概率的行列坐标
    
    # 确保返回的是合法的位置
    y, x = max_prob_pos
    
    return (y, x), output  # 返回坐标和模型输出

import os
import torch
import torch.optim as optim
import torch.nn.functional as f
from mod import mymod, input_qi, game

# 两个权重文件，分别代表 x 棋和 o 棋
mx = 'mx'
mo = 'mo'

# 加载模型，若文件不存在则初始化
def load_model(model, path):
    if os.path.exists(path):
        model.load(path)
        print(f"loaded model from {path}")
    else:
        print(f"{path} not found, initializing new model.")
        # 这里可以加一些初始化模型的代码，例如：
        # model.apply(init_weights) 如果需要初始化权重

# 初始化模型
modx = mymod()
load_model(modx, mx)

modo = mymod()
load_model(modo, mo)

# 定义优化器
lr=0.001
optimizer_x = optim.adam(modx.parameters(), lr=lr)
optimizer_o = optim.adam(modo.parameters(), lr=lr)

# 损失函数：根据游戏结果调整损失
def compute_loss(winner: int, player: str, model_output):
    # 将目标值转换为相应的张量
    if player == "x":
        if winner == 1:  # x 胜
            target = torch.tensor(1.0, dtype=torch.float32)
        elif winner == 0:  # 平局
            target = torch.tensor(0.5, dtype=torch.float32)
        else:  # x 输
            target = torch.tensor(0.0, dtype=torch.float32)
    else:
        if winner == -1:  # o 胜
            target = torch.tensor(1.0, dtype=torch.float32)
        elif winner == 0:  # 平局
            target = torch.tensor(0.5, dtype=torch.float32)
        else:  # o 输
            target = torch.tensor(0.0, dtype=torch.float32)

    # 确保目标值的形状和 model_output 一致，假设 model_output 是单一的值
    target = target.unsqueeze(0).unsqueeze(0)  # 形状变为 (1, 1)

    # 使用均方误差损失计算
    return f.mse_loss(model_output, target)



# 训练模型的过程
def train_game():
    modx.train()
    modo.train()

    # 创建新的游戏实例
    game = game(15, 15)  # 默认是 15x15 棋盘
    # 反向传播和优化
    optimizer_x.zero_grad()
    optimizer_o.zero_grad()

    while not game.win_user:  # 游戏未结束
        # x 方落子
        x_move, x_output = input_qi(game, modx)  # 获取落子位置和模型输出（x_output 是模型的输出）
        game.set(x_move[0], x_move[1])  # x 下棋
        
        if game.win_user:
            break
        
        # o 方落子
        
        o_move, o_output = input_qi(game, modo)  # 获取落子位置和模型输出（o_output 是模型的输出）
        #print(o_move,game)
        game.set(o_move[0], o_move[1])  # o 下棋
       

    # 获取比赛结果
    winner = 0 if game.heqi() else (1 if game.win_user == 'x' else -1)  # 1为x胜，-1为o胜，0为平局

    # 计算损失
    loss_x = compute_loss(winner, "x", x_output)  # 传递模型输出给计算损失函数
    loss_o = compute_loss(winner, "o", o_output)  # 传递模型输出给计算损失函数

    # 计算损失并进行反向传播
    loss_x.backward()
    loss_o.backward()

    # 更新权重
    optimizer_x.step()
    optimizer_o.step()
    print(game)
    return loss_x.item(), loss_o.item()


# 训练多个回合
def train(num_epochs,n):
    k=0
    for epoch in range(num_epochs):
        loss_x, loss_o = train_game()
        print(f"epoch [{epoch+1}/{num_epochs}], loss x: {loss_x}, loss o: {loss_o}")
        k+=1
        if k==n:
            modo.save('mo')
            modx.save('mx')
            print('saved')
            k=0

# 开始训练
train(50000,1000)

from game import game
from mod import mymod,input_qi


#玩家下x ai下o
def playx():
    m=mymod()
    m.load('mo')
    g=game(15,15)
    while 1:
        print(g)
        if g.heqi() or g.win_user:
            break
        while 1:
            try:
                r=input('x:')
                y,x=r.split(',')
                y=int(y)
                x=int(x)
                if g.set(y,x):
                    break
            except exception as e:
                print(e)
        if g.heqi() or g.win_user:
            break
        while 1:
            (y,x),_=input_qi(g,m)
            if g.set(y,x):
                break
    print(g)
    print('winner',g.win_user)


#玩家下o ai下x
def playo():
    m=mymod()
    m.load('mx')
    g=game(15,15)
    while 1:
        
        if g.heqi() or g.win_user:
            break
        while 1:
            (y,x),_=input_qi(g,m)
            if g.set(y,x):
                break
        if g.heqi() or g.win_user:
            break
        print(g)
        while 1:
            try:
                r=input('o:')
                y,x=r.split(',')
                y=int(y)
                x=int(x)
                if g.set(y,x):
                    break
            except exception as e:
                print(e)
    print(g)
    print('winner',g.win_user)

playx()