vim /etc/ssh/sshd_config

PasswordAuthentication no
ChallengeResponseAuthentication no
PermitRootLogin prohibit-password
PubkeyAuthentication yes
sudo systemctl restart sshd

ssh-copy-id -i id_rsa.pub root@192.168.1.1

创建虚拟环境指定版本：virtualenv venv --python=pythonx.x.x

CPU使用率高
./build/examples/alpha_zero_torch_example --game=new_game --actors=28 --evaluators=4 --inference_threads=4 --inference_batch_size=1 --train_batch_size=1024 --inference_cache=2621440 --max_simulations=100 --path=./point --checkpoint_freq=10 --max_steps=100 --verbose=false --devices=cuda:0,cpu --replay_buffer_size=655360 --explicit_learning=true

GPU利用率高
./build/examples/alpha_zero_torch_example --game=new_game --actors=20 --inference_batch_size=6 --inference_threads=3 --evaluators=4 --inference_cache=2621440 --max_simulations=100 --path=./point --checkpoint_freq=10 --max_steps=100 --devices=cuda:0,cpu --replay_buffer_size=655360 --explicit_learning=true

查看GPU驱动
watch -n 1 nvidia-smi
watch -n 1 gpustat

安装gpu驱动
sudo apt purge '^nvidia-.*'
sudo apt purge '^cuda-.*'
sudo apt purge '^libcuda.*'
sudo apt autoremove
sudo apt clean
ubuntu-drivers devices
nvcc -V
export PATH=/usr/local/cuda-12.1/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda-12.1/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

开启Cmake 调试日志
-DCMAKE_CXX_FLAGS="-fsanitize=address -g" \
-DCMAKE_CXX_FLAGS_DEBUG="-O0" \

在alpha_zero训练时无法显示错误，如何定位错误
1.检查coredump设置，如果输出为0，则表示coredump被禁用了。
ulimit -c

2.启用coredump
ulimit -c unlimited

3.配置coredump文件的保存位置
echo "/tmp/core.%e.%p" | sudo tee /proc/sys/kernel/core_pattern

4.修改open_spiel/scripts/build_and_run_tests.sh 199行-DBUILD_TYPE=Debug

5.修改open_spiel/CMakeLists.txt 46行，set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -O0")

6.运行gdb ./build/examples/alpha_zero_torch_example /tmp/core.alpha_zero_torc.444708

7.bt显示错误信息

8.frame 0定位错误

9.list 上下文错误位置

import tensorflow as tf

# 指定checkpoint文件和meta文件的路径
checkpoint_path = '/Users/mac/Desktop/model/checkpoint-25'
meta_path = '/Users/mac/Desktop/model/checkpoint-25.meta'

# 创建一个新的图
graph = tf.Graph()
with graph.as_default():
    # 使用tf.compat.v1来访问旧的API
    with tf.compat.v1.Session() as sess:
        # 导入meta文件中的图结构
        saver = tf.compat.v1.train.import_meta_graph(meta_path, clear_devices=True)
        
        # 恢复模型的权重
        saver.restore(sess, checkpoint_path)
        
        # 指定输出节点
        output_node_names = 'input,value_targets'  # 你需要指定你的模型的输出节点名称
        
        # 将变量转换为常量，并保存为pb文件
        output_graph_def = tf.compat.v1.graph_util.convert_variables_to_constants(
            sess,
            sess.graph_def,
            output_node_names.split(',')
        )
        
        # 保存为.pb文件
        with tf.io.gfile.GFile('/Users/mac/Desktop/model/frozen_model.pb', 'wb') as f:
            f.write(output_graph_def.SerializeToString())

pb预测

import tensorflow.compat.v1 as tf
# 定义一个用于加载模型的函数
def load_model(model_path):
    with graph.as_default():
        # 创建一个新的tf.Session
        sess = tf.compat.v1.Session(graph=graph)
        with tf.io.gfile.GFile(model_path, 'rb') as f:
            graph_def = tf.compat.v1.GraphDef()
            # 解析GraphDef
            graph_def.ParseFromString(f.read())
            # 导入GraphDef到图
            tf.import_graph_def(graph_def, name='')
        return sess  
model_session = load_model('/home/keras/open_spiel/model/aaa.pb')
import pyspiel
import numpy as np
game = pyspiel.load_game("tic_tac_toe")
state=game.new_initial_state()
obs = np.expand_dims(state.observation_tensor(), 0)
mask = np.expand_dims(state.legal_actions_mask(), 0)
def get_var(name):
    return model_session.graph.get_tensor_by_name(name + ":0")
input = get_var("input")
legals_mask = get_var("legals_mask")
training = get_var("training")
value_out = get_var("value_out")
policy_softmax = get_var("policy_softmax")
value, policy=model_session.run(
        [value_out, policy_softmax],
        feed_dict={input: np.array(obs, dtype=np.float32),
                   legals_mask: np.array(mask, dtype=bool),
                   training: False})

例如：Y=w11+w22+b

import pandas as pd

data=pd.read_csv('/Users/mac/Desktop/Keras/000.数据+课件+代码【下部分】/小型数据集/Advertising.csv')

data.head()

x=data[data.columns[1:-1]]

x

y=data.iloc[:,-1]

y

import keras

import keras.layers as layers

model=keras.Sequential()

model.add(layers.Dense(1,input_dim=3))

model.summary()

model.compile(optimizer='adam',loss='mse')

model.fit(x,y,epochs=2000)

使用梯度下降算法

keras.Sequential() #顺序模型

model.add(layers.Dense(1,input_dim=1)) #输入数据1维，输出也是1维

import keras

import numpy as np

import matplotlib.pyplot as plt

x=np.linspace(0,100,30)

y=3*x+7+np.random.randn(30)*8

x

y

plt.scatter(x,y)

model=keras.Sequential()

from keras import layers

model.add(layers.Dense(1,input_dim=1)) 

model.summary()

#编译模型，损失函数，使得损失函数越小越好,adam:梯度下降算法，mse：均方差

model.compile(optimizer='adam',loss='mse')

#训练模型

model.fit(x,y,epochs=3000)

model.predict(x)

plt.scatter(x,y,c='r')
plt.plot(x,model.predict(x))

model.predict([150])