update

codes/TD3/agent.py

@@ -0,0 +1,170 @@
+import copy
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from TD3.memory import ReplayBuffer
+
+
+
+# Implementation of Twin Delayed Deep Deterministic Policy Gradients (TD3)
+# Paper: https://arxiv.org/abs/1802.09477
+
+
+class Actor(nn.Module):
+	def __init__(self, state_dim, action_dim, max_action):
+		super(Actor, self).__init__()
+
+		self.l1 = nn.Linear(state_dim, 256)
+		self.l2 = nn.Linear(256, 256)
+		self.l3 = nn.Linear(256, action_dim)
+		
+		self.max_action = max_action
+		
+
+	def forward(self, state):
+		a = F.relu(self.l1(state))
+		a = F.relu(self.l2(a))
+		return self.max_action * torch.tanh(self.l3(a))
+
+
+class Critic(nn.Module):
+	def __init__(self, state_dim, action_dim):
+		super(Critic, self).__init__()
+
+		# Q1 architecture
+		self.l1 = nn.Linear(state_dim + action_dim, 256)
+		self.l2 = nn.Linear(256, 256)
+		self.l3 = nn.Linear(256, 1)
+
+		# Q2 architecture
+		self.l4 = nn.Linear(state_dim + action_dim, 256)
+		self.l5 = nn.Linear(256, 256)
+		self.l6 = nn.Linear(256, 1)
+
+
+	def forward(self, state, action):
+		sa = torch.cat([state, action], 1)
+
+		q1 = F.relu(self.l1(sa))
+		q1 = F.relu(self.l2(q1))
+		q1 = self.l3(q1)
+
+		q2 = F.relu(self.l4(sa))
+		q2 = F.relu(self.l5(q2))
+		q2 = self.l6(q2)
+		return q1, q2
+
+
+	def Q1(self, state, action):
+		sa = torch.cat([state, action], 1)
+
+		q1 = F.relu(self.l1(sa))
+		q1 = F.relu(self.l2(q1))
+		q1 = self.l3(q1)
+		return q1
+
+
+class TD3(object):
+	def __init__(
+		self,
+		state_dim,
+		action_dim,
+		max_action,
+		cfg,
+	):
+		self.max_action = max_action
+		self.gamma = cfg.gamma
+		self.lr = cfg.lr
+		self.policy_noise = cfg.policy_noise
+		self.noise_clip = cfg.noise_clip
+		self.policy_freq = cfg.policy_freq
+		self.batch_size =  cfg.batch_size 
+		self.device = cfg.device
+		self.total_it = 0
+
+		self.actor = Actor(state_dim, action_dim, max_action).to(self.device)
+		self.actor_target = copy.deepcopy(self.actor)
+		self.actor_optimizer = torch.optim.Adam(self.actor.parameters(), lr=3e-4)
+
+		self.critic = Critic(state_dim, action_dim).to(self.device)
+		self.critic_target = copy.deepcopy(self.critic)
+		self.critic_optimizer = torch.optim.Adam(self.critic.parameters(), lr=3e-4)
+		self.memory = ReplayBuffer(state_dim, action_dim)
+
+		
+
+
+	def choose_action(self, state):
+		state = torch.FloatTensor(state.reshape(1, -1)).to(self.device)
+		return self.actor(state).cpu().data.numpy().flatten()
+
+
+	def update(self):
+		self.total_it += 1
+
+		# Sample replay buffer 
+		state, action, next_state, reward, not_done = self.memory.sample(self.batch_size)
+
+		with torch.no_grad():
+			# Select action according to policy and add clipped noise
+			noise = (
+				torch.randn_like(action) * self.policy_noise
+			).clamp(-self.noise_clip, self.noise_clip)
+			
+			next_action = (
+				self.actor_target(next_state) + noise
+			).clamp(-self.max_action, self.max_action)
+
+			# Compute the target Q value
+			target_Q1, target_Q2 = self.critic_target(next_state, next_action)
+			target_Q = torch.min(target_Q1, target_Q2)
+			target_Q = reward + not_done * self.gamma * target_Q
+
+		# Get current Q estimates
+		current_Q1, current_Q2 = self.critic(state, action)
+
+		# Compute critic loss
+		critic_loss = F.mse_loss(current_Q1, target_Q) + F.mse_loss(current_Q2, target_Q)
+
+		# Optimize the critic
+		self.critic_optimizer.zero_grad()
+		critic_loss.backward()
+		self.critic_optimizer.step()
+
+		# Delayed policy updates
+		if self.total_it % self.policy_freq == 0:
+
+			# Compute actor losse
+			actor_loss = -self.critic.Q1(state, self.actor(state)).mean()
+			
+			# Optimize the actor 
+			self.actor_optimizer.zero_grad()
+			actor_loss.backward()
+			self.actor_optimizer.step()
+
+			# Update the frozen target models
+			for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()):
+				target_param.data.copy_(self.lr * param.data + (1 - self.lr) * target_param.data)
+
+			for param, target_param in zip(self.actor.parameters(), self.actor_target.parameters()):
+				target_param.data.copy_(self.lr * param.data + (1 - self.lr) * target_param.data)
+
+
+	def save(self, path):
+		torch.save(self.critic.state_dict(), path + "td3_critic")
+		torch.save(self.critic_optimizer.state_dict(), path + "td3_critic_optimizer")
+		
+		torch.save(self.actor.state_dict(), path + "td3_actor")
+		torch.save(self.actor_optimizer.state_dict(), path + "td3_actor_optimizer")
+
+
+	def load(self, path):
+		self.critic.load_state_dict(torch.load(path + "td3_critic"))
+		self.critic_optimizer.load_state_dict(torch.load(path + "td3_critic_optimizer"))
+		self.critic_target = copy.deepcopy(self.critic)
+
+		self.actor.load_state_dict(torch.load(path + "td3_actor"))
+		self.actor_optimizer.load_state_dict(torch.load(path + "td3_actor_optimizer"))
+		self.actor_target = copy.deepcopy(self.actor)
+		

codes/TD3/main.py

@@ -1,14 +1,169 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-@Author: John
-@Email: johnjim0816@gmail.com
-@Date: 2020-06-11 23:38:13
-@LastEditor: John
-@LastEditTime: 2020-06-11 23:38:31
-@Discription: 
-@Environment: python 3.7.7
-'''
+import sys,os
+curr_path = os.path.dirname(__file__)
+parent_path=os.path.dirname(curr_path) 
+sys.path.append(parent_path) # add current terminal path to sys.path
+
 import torch
+import gym
+import numpy as np
+import datetime
+
+
+from TD3.agent import TD3
+from common.plot import plot_rewards
+from common.utils import save_results,make_dir
+
+curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
+	
+
+class TD3Config:
+	def __init__(self) -> None:
+		self.algo = 'TD3'
+		self.env = 'HalfCheetah-v2'
+		self.seed = 0
+		self.result_path = curr_path+"/results/" +self.env+'/'+curr_time+'/'  # path to save results
+		self.start_timestep = 25e3 # Time steps initial random policy is used
+		self.eval_freq = 5e3 # How often (time steps) we evaluate
+		# self.train_eps = 800
+		self.max_timestep = 1600000 # Max time steps to run environment
+		self.expl_noise = 0.1 # Std of Gaussian exploration noise
+		self.batch_size = 256 # Batch size for both actor and critic
+		self.gamma = 0.99 # gamma factor
+		self.lr = 0.0005 # Target network update rate 
+		self.policy_noise = 0.2 # Noise added to target policy during critic update
+		self.noise_clip = 0.5  # Range to clip target policy noise
+		self.policy_freq = 2 # Frequency of delayed policy updates
+		self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+		
+# Runs policy for X episodes and returns average reward
+# A fixed seed is used for the eval environment
+def eval(env,agent, seed, eval_episodes=10):
+	eval_env = gym.make(env)
+	eval_env.seed(seed + 100)
+	avg_reward = 0.
+	for _ in range(eval_episodes):
+		state, done = eval_env.reset(), False
+		while not done:
+			# eval_env.render()
+			action = agent.choose_action(np.array(state))
+			state, reward, done, _ = eval_env.step(action)
+			avg_reward += reward
+	avg_reward /= eval_episodes
+	print("---------------------------------------")
+	print(f"Evaluation over {eval_episodes} episodes: {avg_reward:.3f}")
+	print("---------------------------------------")
+	return avg_reward
+
+def train(cfg,env,agent):
+	# Evaluate untrained policy
+	evaluations = [eval(cfg.env,agent, cfg.seed)]
+	state, done = env.reset(), False
+	ep_reward = 0
+	ep_timesteps = 0
+	episode_num = 0
+	rewards = []
+	ma_rewards = [] # moveing average reward
+	for t in range(int(cfg.max_timestep)):
+		ep_timesteps += 1
+		# Select action randomly or according to policy
+		if t < cfg.start_timestep:
+			action = env.action_space.sample()
+		else:
+			action = (
+				agent.choose_action(np.array(state))
+				+ np.random.normal(0, max_action * cfg.expl_noise, size=action_dim)
+			).clip(-max_action, max_action)
+		# Perform action
+		next_state, reward, done, _ = env.step(action) 
+		done_bool = float(done) if ep_timesteps < env._max_episode_steps else 0
+		# Store data in replay buffer
+		agent.memory.push(state, action, next_state, reward, done_bool)
+		state = next_state
+		ep_reward += reward
+		# Train agent after collecting sufficient data
+		if t >= cfg.start_timestep:
+			agent.update()
+		if done: 
+			# +1 to account for 0 indexing. +0 on ep_timesteps since it will increment +1 even if done=True
+			print(f"Episode:{episode_num+1}, Episode T:{ep_timesteps}, Reward:{ep_reward:.3f}")
+			# Reset environment
+			state, done = env.reset(), False
+			rewards.append(ep_reward)
+			# 计算滑动窗口的reward
+			if ma_rewards:
+				ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
+			else:
+				ma_rewards.append(ep_reward) 
+			ep_reward = 0
+			ep_timesteps = 0
+			episode_num += 1 
+		# Evaluate episode
+		if (t + 1) % cfg.eval_freq == 0:
+			evaluations.append(eval(cfg.env,agent, cfg.seed))
+	return rewards, ma_rewards
+# def train(cfg,env,agent):
+# 	evaluations = [eval(cfg.env,agent,cfg.seed)]
+# 	ep_reward = 0
+# 	tot_timestep = 0
+# 	rewards = []
+# 	ma_rewards = [] # moveing average reward
+# 	for i_ep in range(int(cfg.train_eps)):
+# 		state, done = env.reset(), False
+# 		ep_reward = 0
+# 		ep_timestep = 0
+# 		while not done:
+# 			ep_timestep += 1
+# 			tot_timestep +=1
+# 			# Select action randomly or according to policy
+# 			if tot_timestep < cfg.start_timestep:
+# 				action = env.action_space.sample()
+# 			else:
+# 				action = (
+# 					agent.choose_action(np.array(state))
+# 					+ np.random.normal(0, max_action * cfg.expl_noise, size=action_dim)
+# 				).clip(-max_action, max_action)
+# 			# action = (
+# 			# 		agent.choose_action(np.array(state))
+# 			# 		+ np.random.normal(0, max_action * cfg.expl_noise, size=action_dim)
+# 			# 	).clip(-max_action, max_action)
+# 			# Perform action
+# 			next_state, reward, done, _ = env.step(action) 
+# 			done_bool = float(done) if ep_timestep < env._max_episode_steps else 0
+
+# 			# Store data in replay buffer
+# 			agent.memory.push(state, action, next_state, reward, done_bool)
+# 			state = next_state
+# 			ep_reward += reward
+# 			# Train agent after collecting sufficient data
+# 			if tot_timestep >= cfg.start_timestep:
+# 				agent.update()
+# 		print(f"Episode:{i_ep}/{cfg.train_eps}, Episode Timestep:{ep_timestep}, Reward:{ep_reward:.3f}")
+# 		rewards.append(ep_reward)
+# 		# 计算滑动窗口的reward
+# 		if ma_rewards:
+# 			ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
+# 		else:
+# 			ma_rewards.append(ep_reward) 
+# 		# Evaluate episode
+# 		if (i_ep+1) % cfg.eval_freq == 0:
+# 			evaluations.append(eval(cfg.env,agent, cfg.seed))
+# 	return rewards,ma_rewards
+
+
 if __name__ == "__main__":
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+	cfg  = TD3Config()
+	env = gym.make(cfg.env)
+	env.seed(cfg.seed) # Set seeds
+	torch.manual_seed(cfg.seed)
+	np.random.seed(cfg.seed)
+	state_dim = env.observation_space.shape[0]
+	action_dim = env.action_space.shape[0] 
+	max_action = float(env.action_space.high[0])
+	agent = TD3(state_dim,action_dim,max_action,cfg)
+	rewards,ma_rewards = train(cfg,env,agent)
+	make_dir(cfg.result_path)
+	agent.save(path=cfg.result_path)
+	save_results(rewards,ma_rewards,tag='train',path=cfg.result_path)
+	plot_rewards(rewards,ma_rewards,tag="train",env=cfg.env,algo = cfg.algo,path=cfg.result_path)
+	
+		

codes/TD3/memory.py

@@ -1,34 +1,44 @@
 #!/usr/bin/env python
 # coding=utf-8
 '''
-@Author: John
-@Email: johnjim0816@gmail.com
-@Date: 2020-06-10 15:27:16
-@LastEditor: John
-@LastEditTime: 2020-06-11 21:04:50
-@Discription: 
-@Environment: python 3.7.7
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2021-04-13 11:00:13
+LastEditor: John
+LastEditTime: 2021-04-15 01:25:14
+Discription: 
+Environment: 
 '''
-import random
 import numpy as np
+import torch
 
-class ReplayBuffer:
-    
-    def __init__(self, capacity):
-        self.capacity = capacity
-        self.buffer = []
-        self.position = 0
-    
-    def push(self, state, action, reward, next_state, done):
-        if len(self.buffer) < self.capacity:
-            self.buffer.append(None)
-        self.buffer[self.position] = (state, action, reward, next_state, done)
-        self.position = (self.position + 1) % self.capacity
-    
-    def sample(self, batch_size):
-        batch = random.sample(self.buffer, batch_size)
-        state, action, reward, next_state, done = map(np.stack, zip(*batch))
-        return state, action, reward, next_state, done
-    
-    def __len__(self):
-        return len(self.buffer)
+
+class ReplayBuffer(object):
+	def __init__(self, state_dim, action_dim, max_size=int(1e6)):
+		self.max_size = max_size
+		self.ptr = 0
+		self.size = 0
+		self.state = np.zeros((max_size, state_dim))
+		self.action = np.zeros((max_size, action_dim))
+		self.next_state = np.zeros((max_size, state_dim))
+		self.reward = np.zeros((max_size, 1))
+		self.not_done = np.zeros((max_size, 1))
+		self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+	def push(self, state, action, next_state, reward, done):
+		self.state[self.ptr] = state
+		self.action[self.ptr] = action
+		self.next_state[self.ptr] = next_state
+		self.reward[self.ptr] = reward
+		self.not_done[self.ptr] = 1. - done
+		self.ptr = (self.ptr + 1) % self.max_size
+		self.size = min(self.size + 1, self.max_size)
+
+	def sample(self, batch_size):
+		ind = np.random.randint(0, self.size, size=batch_size)
+		return (
+			torch.FloatTensor(self.state[ind]).to(self.device),
+			torch.FloatTensor(self.action[ind]).to(self.device),
+			torch.FloatTensor(self.next_state[ind]).to(self.device),
+			torch.FloatTensor(self.reward[ind]).to(self.device),
+			torch.FloatTensor(self.not_done[ind]).to(self.device)
+		)