A3C——pytorch

这一部分的代码都来自莫凡，由于所看的书《白话强化学习》里面的代码块有一部分看不懂，转而看其他老师的代码，感觉莫老师的代码通俗易懂，但是语法可能和书上所学的有所不一样，所以还是读了这个代码，作了点注释，以供之后翻阅
main.py

"""
Reinforcement Learning (A3C) using Pytroch + multiprocessing.
The most simple implementation for continuous action.

View more on my Chinese tutorial page [莫烦Python](https://morvanzhou.github.io/).
"""

import torch
import torch.nn as nn
from utils import v_wrap, set_init, push_and_pull, record
import torch.nn.functional as F
import torch.multiprocessing as mp
from shared_adam import SharedAdam
import gym
import math, os
os.environ["OMP_NUM_THREADS"] = "1"

UPDATE_GLOBAL_ITER = 5
GAMMA = 0.9
MAX_EP = 3000
MAX_EP_STEP = 200

env = gym.make('Pendulum-v0')
N_S = env.observation_space.shape[0]
# 环境观测空间
N_A = env.action_space.shape[0]
# 环境动作空间


class Net(nn.Module):
    def __init__(self, s_dim, a_dim):
        super(Net, self).__init__()
        self.s_dim = s_dim
        self.a_dim = a_dim
        self.a1 = nn.Linear(s_dim, 200)
        self.mu = nn.Linear(200, a_dim)
        self.sigma = nn.Linear(200, a_dim)
        self.c1 = nn.Linear(s_dim, 100)
        self.v = nn.Linear(100, 1)
        set_init([self.a1, self.mu, self.sigma, self.c1, self.v])
        # 将参数初始化
        self.distribution = torch.distributions.Normal

    def forward(self, x):
        a1 = F.relu6(self.a1(x))
        mu = 2 * F.tanh(self.mu(a1))
        sigma = F.softplus(self.sigma(a1)) + 0.001      # avoid 0
        c1 = F.relu6(self.c1(x))
        values = self.v(c1)
        return mu, sigma, values

    def choose_action(self, s):
        self.training = False
        mu, sigma, _ = self.forward(s)
        m = self.distribution(mu.view(1, ).data, sigma.view(1, ).data)
        # distribution分布中选取动作，对应后面的动作的概率，可以直接得到
        return m.sample().numpy()

    def loss_func(self, s, a, v_t):
        self.train()
        mu, sigma, values = self.forward(s)
        td = v_t - values
        c_loss = td.pow(2)   # critic 的误差函数

        m = self.distribution(mu, sigma)
        # 生成正态分布
        log_prob = m.log_prob(a)
        # 动作的可能性
        #log_prob*td代表动作价值期望
        entropy = 0.5 + 0.5 * math.log(2 * math.pi) + torch.log(m.scale)  #添加误差，增加探索性，相当于采用了贪心算法 exploration
        exp_v = log_prob * td.detach() + 0.005 * entropy
        a_loss = -exp_v
        total_loss = (a_loss + c_loss).mean()
        return total_loss


class Worker(mp.Process):
    def __init__(self, gnet, opt, global_ep, global_ep_r, res_queue, name):
        super(Worker, self).__init__()
        self.name = 'w%i' % name
        self.g_ep, self.g_ep_r, self.res_queue = global_ep, global_ep_r, res_queue
        self.gnet, self.opt = gnet, opt
        self.lnet = Net(N_S, N_A)           # local network
        self.env = gym.make('Pendulum-v0').unwrapped

    def run(self):
        # 在mp文件中有run这个函数，应该是被改写了，进行进程时自动开启运行这个子类。
        # 具体的可以参考这个博客：https://blog.csdn.net/zhangxuelong461/article/details/104059149，解释的很清楚，以后都可以这样用
        total_step = 1
        while self.g_ep.value < MAX_EP:
            s = self.env.reset()
            buffer_s, buffer_a, buffer_r = [], [], []
            ep_r = 0.
            for t in range(MAX_EP_STEP):
                if self.name == 'w0':
                    self.env.render()
                a = self.lnet.choose_action(v_wrap(s[None, :]))
                # [None, :]代表增加1个维度
                s_, r, done, _ = self.env.step(a.clip(-2, 2))
                if t == MAX_EP_STEP - 1:
                    done = True
                ep_r += r
                buffer_a.append(a)
                buffer_s.append(s)
                buffer_r.append((r+8.1)/8.1)    # normalize

                if total_step % UPDATE_GLOBAL_ITER == 0 or done:  # update global and assign to local net
                    # 当游戏结束或这到达更新全局参数次数整数倍时进行更新参数
                    # sync
                    push_and_pull(self.opt, self.lnet, self.gnet, done, s_, buffer_s, buffer_a, buffer_r, GAMMA)
                    buffer_s, buffer_a, buffer_r = [], [], []

                    if done:  # done and print information
                        record(self.g_ep, self.g_ep_r, ep_r, self.res_queue, self.name)
                        break
                s = s_
                total_step += 1

        self.res_queue.put(None)


if __name__ == "__main__":
    gnet = Net(N_S, N_A)        # global network
    gnet.share_memory()         # share the global parameters in multiprocessing
    opt = SharedAdam(gnet.parameters(), lr=1e-4, betas=(0.95, 0.999))  # global optimizer
    global_ep, global_ep_r, res_queue = mp.Value('i', 0), mp.Value('d', 0.), mp.Queue()

    # parallel training
    workers = [Worker(gnet, opt, global_ep, global_ep_r, res_queue, i) for i in range(mp.cpu_count())]

    [w.start() for w in workers]
    res = []                    # record episode reward to plot
    while True:
        r = res_queue.get()
        if r is not None:
            res.append(r)
        else:
            break
    [w.join() for w in workers]

    import matplotlib.pyplot as plt
    plt.plot(res)
    plt.ylabel('Moving average ep reward')
    plt.xlabel('Step')
    plt.show()

utils.py

"""
Functions that use multiple times
"""

from torch import nn
import torch
import numpy as np


def v_wrap(np_array, dtype=np.float32):
    if np_array.dtype != dtype:
        np_array = np_array.astype(dtype)
    return torch.from_numpy(np_array)


def set_init(layers):
    for layer in layers:
        # 将参数初始化
        nn.init.normal_(layer.weight, mean=0., std=0.1)
        nn.init.constant_(layer.bias, 0.)


def push_and_pull(opt, lnet, gnet, done, s_, bs, ba, br, gamma):
    if done:
        v_s_ = 0.               # terminal
    else:
        v_s_ = lnet.forward(v_wrap(s_[None, :]))[-1].data.numpy()[0, 0]
        # 取出下一个状态s'下values价值数值

    buffer_v_target = []
    for r in br[::-1]:    # reverse buffer r
        v_s_ = r + gamma * v_s_
        # n步折扣价值函数
        buffer_v_target.append(v_s_)
    buffer_v_target.reverse()   # 反向列表元素

    loss = lnet.loss_func(
        v_wrap(np.vstack(bs)),
        v_wrap(np.array(ba), dtype=np.int64) if ba[0].dtype == np.int64 else v_wrap(np.vstack(ba)),
        v_wrap(np.array(buffer_v_target)[:, None]))
    # v_wrap()函数是用来将类型转换

    # calculate local gradients and push local parameters to global
    opt.zero_grad()
    loss.backward()
    for lp, gp in zip(lnet.parameters(), gnet.parameters()):
        gp._grad = lp.grad
        # 将刚刚计算的梯度传递给全局网路
    opt.step()

    # pull global parameters
    lnet.load_state_dict(gnet.state_dict())


def record(global_ep, global_ep_r, ep_r, res_queue, name):
    with global_ep.get_lock():
        global_ep.value += 1
    with global_ep_r.get_lock():
        if global_ep_r.value == 0.:
            global_ep_r.value = ep_r
        else:
            global_ep_r.value = global_ep_r.value * 0.99 + ep_r * 0.01
    res_queue.put(global_ep_r.value)
    print(
        name,
        "Ep:", global_ep.value,
        "| Ep_r: %.0f" % global_ep_r.value,
    )