learnBasePolicy.py

import torch
import torch.nn as nn
import torch.optim as optim
import numpy as np
from dm_control import suite
from torch.distributions import Normal
import matplotlib.pyplot as plt
import os
import torch.nn.functional as F
import time
import random
from torch.autograd import Variable
from collections import deque, OrderedDict


domainName = "walker" # Name of a environment (set it to any Continous environment you want)
taskName = "stand" # Name of a environment (set it to any Continous environment you want)
env_name = domainName+ "_"+taskName
seed = 0 # Random seed number
start_timesteps = 1e4 # Number of iterations/timesteps before which the model randomly chooses an action, and after which it starts to use the policy network
eval_freq = 5e3 # How often the evaluation step is performed (after how many timesteps)
max_timesteps = 5e5 # Total number of iterations/timesteps
save_models = True # Boolean checker whether or not to save the pre-trained model
expl_noise = 0.1 # Exploration noise - STD value of exploration Gaussian noise
batch_size = 100 # Size of the batch
discount = 0.99 # Discount factor gamma, used in the calculation of the total discounted reward
tau = 0.005 # Target network update rate
policy_noise = 0.2 # STD of Gaussian noise added to the actions for the exploration purposes
noise_clip = 0.5 # Maximum value of the Gaussian noise added to the actions (policy)
policy_freq = 2 # Number of iterations to wait before the policy network (Actor model) is updated
     
file_name = "%s_%s_%s" % ("TD3", env_name, str(seed))
print ("---------------------------------------")
print ("Settings: %s" % (file_name))
print ("---------------------------------------")


if not os.path.exists("./results"):
  os.makedirs("./results")
if save_models and not os.path.exists("./pytorch_models"):
  os.makedirs("./pytorch_models")


class ReplayBuffer(object):

  def __init__(self, max_size=1e6):
    self.storage = []
    self.max_size = max_size
    self.ptr = 0

  def add(self, transition):
    if len(self.storage) == self.max_size:
      self.storage[int(self.ptr)] = transition
      self.ptr = (self.ptr + 1) % self.max_size
    else:
      self.storage.append(transition)

  def sample(self, batch_size):
    ind = np.random.randint(0, len(self.storage), size=batch_size)
    batch_states, batch_next_states, batch_actions, batch_rewards, batch_dones = [], [], [], [], []
    for i in ind: 
      state, next_state, action, reward, done = self.storage[i]
      batch_states.append(np.array(state, copy=False))
      batch_next_states.append(np.array(next_state, copy=False))
      batch_actions.append(np.array(action, copy=False))
      batch_rewards.append(np.array(reward, copy=False))
      batch_dones.append(np.array(done, copy=False))
    return np.array(batch_states), np.array(batch_next_states), np.array(batch_actions), np.array(batch_rewards).reshape(-1, 1), np.array(batch_dones).reshape(-1, 1)
  

class Actor(nn.Module):
  
  def __init__(self, state_dim, action_dim, max_action):
    super(Actor, self).__init__()
    self.layer_1 = nn.Linear(state_dim, 400)
    self.layer_2 = nn.Linear(400, 300)
    self.layer_3 = nn.Linear(300, action_dim)
    self.max_action = max_action

  def forward(self, x):
    x = F.relu(self.layer_1(x))
    x = F.relu(self.layer_2(x))
    x = self.max_action * torch.tanh(self.layer_3(x))
    return x
  
class Critic(nn.Module):
  
  def __init__(self, state_dim, action_dim):
    super(Critic, self).__init__()
    # Defining the first Critic neural network
    self.layer_1 = nn.Linear(state_dim + action_dim, 400)
    self.layer_2 = nn.Linear(400, 300)
    self.layer_3 = nn.Linear(300, 1)
    # Defining the second Critic neural network
    self.layer_4 = nn.Linear(state_dim + action_dim, 400)
    self.layer_5 = nn.Linear(400, 300)
    self.layer_6 = nn.Linear(300, 1)

  def forward(self, x, u):
    xu = torch.cat([x, u], 1)
    # Forward-Propagation on the first Critic Neural Network
    x1 = F.relu(self.layer_1(xu))
    x1 = F.relu(self.layer_2(x1))
    x1 = self.layer_3(x1)
    # Forward-Propagation on the second Critic Neural Network
    x2 = F.relu(self.layer_4(xu))
    x2 = F.relu(self.layer_5(x2))
    x2 = self.layer_6(x2)
    return x1, x2

  def Q1(self, x, u):
    xu = torch.cat([x, u], 1)
    x1 = F.relu(self.layer_1(xu))
    x1 = F.relu(self.layer_2(x1))
    x1 = self.layer_3(x1)
    return x1

class TD3(object):
  
  def __init__(self, state_dim, action_dim, max_action):
    self.actor = Actor(state_dim, action_dim, max_action).to(device)
    self.actor_target = Actor(state_dim, action_dim, max_action).to(device)
    self.actor_target.load_state_dict(self.actor.state_dict())
    self.actor_optimizer = torch.optim.Adam(self.actor.parameters())
    self.critic = Critic(state_dim, action_dim).to(device)
    self.critic_target = Critic(state_dim, action_dim).to(device)
    self.critic_target.load_state_dict(self.critic.state_dict())
    self.critic_optimizer = torch.optim.Adam(self.critic.parameters())
    self.max_action = max_action

  def select_action(self, state):
    #print(state)
    state = torch.Tensor(state.reshape(1, -1)).to(device)
    return self.actor(state).cpu().data.numpy().flatten()

  def train(self, replay_buffer, iterations, batch_size=100, discount=0.99, tau=0.005, policy_noise=0.2, noise_clip=0.5, policy_freq=2):
    
    for it in range(iterations):
      
      # Step 4: We sample a batch of transitions (s, s’, a, r) from the memory
      batch_states, batch_next_states, batch_actions, batch_rewards, batch_dones = replay_buffer.sample(batch_size)
      state = torch.Tensor(batch_states).to(device)
      next_state = torch.Tensor(batch_next_states).to(device)
      action = torch.Tensor(batch_actions).to(device)
      reward = torch.Tensor(batch_rewards).to(device)
      done = torch.Tensor(batch_dones).to(device)
      
      # Step 5: From the next state s’, the Actor target plays the next action a’
      next_action = self.actor_target(next_state)
      
      # Step 6: We add Gaussian noise to this next action a’ and we clamp it in a range of values supported by the environment
      noise = torch.Tensor(batch_actions).data.normal_(0, policy_noise).to(device)
      noise = noise.clamp(-noise_clip, noise_clip)
      next_action = (next_action + noise).clamp(-self.max_action, self.max_action)
      
      # Step 7: The two Critic targets take each the couple (s’, a’) as input and return two Q-values Qt1(s’,a’) and Qt2(s’,a’) as outputs
      target_Q1, target_Q2 = self.critic_target(next_state, next_action)
      
      # Step 8: We keep the minimum of these two Q-values: min(Qt1, Qt2)
      target_Q = torch.min(target_Q1, target_Q2)
      
      # Step 9: We get the final target of the two Critic models, which is: Qt = r + γ * min(Qt1, Qt2), where γ is the discount factor
      target_Q = reward + ((1 - done) * discount * target_Q).detach()
      
      # Step 10: The two Critic models take each the couple (s, a) as input and return two Q-values Q1(s,a) and Q2(s,a) as outputs
      current_Q1, current_Q2 = self.critic(state, action)
      
      # Step 11: We compute the loss coming from the two Critic models: Critic Loss = MSE_Loss(Q1(s,a), Qt) + MSE_Loss(Q2(s,a), Qt)
      critic_loss = F.mse_loss(current_Q1, target_Q) + F.mse_loss(current_Q2, target_Q)
      
      # Step 12: We backpropagate this Critic loss and update the parameters of the two Critic models with a SGD optimizer
      self.critic_optimizer.zero_grad()
      critic_loss.backward()
      self.critic_optimizer.step()
      
      # Step 13: Once every two iterations, we update our Actor model by performing gradient ascent on the output of the first Critic model
      if it % policy_freq == 0:
        actor_loss = -self.critic.Q1(state, self.actor(state)).mean()
        self.actor_optimizer.zero_grad()
        actor_loss.backward()
        self.actor_optimizer.step()
        
        # Step 14: Still once every two iterations, we update the weights of the Actor target by polyak averaging
        for param, target_param in zip(self.actor.parameters(), self.actor_target.parameters()):
          target_param.data.copy_(tau * param.data + (1 - tau) * target_param.data)
        
        # Step 15: Still once every two iterations, we update the weights of the Critic target by polyak averaging
        for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()):
          target_param.data.copy_(tau * param.data + (1 - tau) * target_param.data)
  
  # Making a save method to save a trained model
  def save(self, filename, directory):
    torch.save(self.actor.state_dict(), '%s/%s_actor.pth' % (directory, filename))
    torch.save(self.critic.state_dict(), '%s/%s_critic.pth' % (directory, filename))
  
  # Making a load method to load a pre-trained model
  def load(self, filename, directory):
    self.actor.load_state_dict(torch.load('%s/%s_actor.pth' % (directory, filename)))
    self.critic.load_state_dict(torch.load('%s/%s_critic.pth' % (directory, filename)))


def evaluate_policy(policy, eval_episodes=10):
  avg_reward = 0.
  for _ in range(eval_episodes):
    #obs = env.reset()

    state = env.reset()
    obs = process_state(state)
    

    done = False
    while not done:
      action = policy.select_action(np.array(obs))
      #obs, reward, done, _ = env.step(action)
      state = env.step(action)
      obs = process_state(state)
      reward = state.reward
      done = state.last()

      avg_reward += reward
  avg_reward /= eval_episodes
  print ("---------------------------------------")
  print ("Average Reward over the Evaluation Step: %f" % (avg_reward))
  print ("---------------------------------------")
  return avg_reward

def process_state(state):
    if isinstance(state, OrderedDict):
        if 'orientations' in state and 'height' in state and 'velocity' in state:
            orient = state['orientations']
            height = state['height']
            velocity = state['velocity']
            if np.isscalar(height):
                height = np.array([height])
            out = np.concatenate((orient, height, velocity))
            return out
    elif isinstance(state, np.ndarray) and state.shape == (24,):
        return state
    elif hasattr(state, 'observation') and isinstance(state.observation, OrderedDict):
        observation = state.observation
        if 'orientations' in observation and 'height' in observation and 'velocity' in observation:
            orient = observation['orientations']
            height = observation['height']
            velocity = observation['velocity']
            if np.isscalar(height):
                height = np.array([height])
            out = np.concatenate((orient, height, velocity))
            return out
    else:
        raise ValueError("Input state must be either an OrderedDict with keys 'orientations', 'height', and 'velocity', a numpy ndarray of shape (24,), or a TimeStep object with a valid observation.")



logginWB = True
if logginWB:
   import wandb
   wandb.init(project="dmWalker_train",name=f"TrainBasepolicy",save_code=True)

env = suite.load(domain_name=domainName, task_name=taskName, task_kwargs={'random': seed})
torch.manual_seed(seed)
np.random.seed(seed)
random_state = np.random.RandomState(seed)


obsSpec = env.observation_spec()
action_spec = env.action_spec()

orientDim = obsSpec['orientations'].shape[0]
heightDim = len(obsSpec['height'].shape) + 1 
velocityDim = obsSpec['velocity'].shape[0]
input_dim = orientDim + heightDim + velocityDim

hidden_dim = 128
output_dim = env.action_spec().shape[0]


state_dim = input_dim
action_dim = output_dim
max_action = float(1)

# Selecting the device (CPU or GPU)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")


policy = TD3(state_dim, action_dim, max_action)
replay_buffer = ReplayBuffer()
evaluations = [evaluate_policy(policy)]


total_timesteps = 0
timesteps_since_eval = 0
episode_num = 0
done = True
t0 = time.time()
episode_reward = 0
episode_timesteps = 0
ii = 0
# We start the main loop over 500,000 timesteps
while total_timesteps < max_timesteps:
  # If the episode is done
    if done:
        # If we are not at the very beginning, we start the training process of the model
        if total_timesteps != 0:
            print("Total Timesteps: {} Episode Num: {} Reward: {}".format(total_timesteps, episode_num, episode_reward))
            if logginWB:
              wandb.log({'episode_reward':episode_reward,
                    },step=ii)
              ii += 1 
            policy.train(replay_buffer, episode_timesteps, batch_size, discount, tau, policy_noise, noise_clip, policy_freq)

        # We evaluate the episode and we save the policy
        if timesteps_since_eval >= eval_freq:
            timesteps_since_eval %= eval_freq
            evaluations.append(evaluate_policy(policy))
            policy.save(file_name, directory="./pytorch_models")
            np.save("./results/%s" % (file_name), evaluations)

        # When the training step is done, we reset the state of the environment
        state = env.reset()
        obs = process_state(state)

        # Set the Done to False
        done = False

        # Set rewards and episode timesteps to zero
        episode_reward = 0
        episode_timesteps = 0
        episode_num += 1
    
    # Before 10000 timesteps, we play random actions
    if total_timesteps < start_timesteps:
        action = random_state.uniform(action_spec.minimum, action_spec.maximum, action_spec.shape)
    else: # After 10000 timesteps, we switch to the model
        action = policy.select_action(np.array(obs))
        # If the explore_noise parameter is not 0, we add noise to the action and we clip it
        if expl_noise != 0:
            action = (action + np.random.normal(0, expl_noise, size=action_spec.shape[0])).clip(action_spec.minimum, action_spec.maximum)

    # The agent performs the action in the environment, then reaches the next state and receives the reward
    state = env.step(action)
    new_obs = process_state(state)
    reward = state.reward
    done = state.last()

    # We check if the episode is done
    done_bool = float(done)
    
    # We increase the total reward
    episode_reward += reward

    # We store the new transition into the Experience Replay memory (ReplayBuffer)
    replay_buffer.add((obs, new_obs, action, reward, done_bool))

    # We update the state, the episode timestep, the total timesteps, and the timesteps since the evaluation of the policy
    obs = new_obs
    episode_timesteps += 1
    total_timesteps += 1
    timesteps_since_eval += 1
    
    
    

wandb.finish()

# We add the last policy evaluation to our list of evaluations and we save our model
evaluations.append(evaluate_policy(policy))

if save_models: policy.save("%s" % (file_name), directory="./pytorch_models")
np.save("./results/%s" % (file_name), evaluations)
  
if logginWB:
    wandb.init(project="dmWalker_test",name=f"evaluation_basepolicy",save_code=True)
    for idx, evaluation in enumerate(evaluations):
        wandb.log({'evaluation_reward': evaluation}, step=idx)
        
# Finalize WandB run
wandb.finish()