pytorch-soft-actor-critic/sac.py

import sys
import os
import numpy as np
import torch
import torch.nn.functional as F
from torch.optim import Adam
from utils import soft_update, hard_update
from model import GaussianPolicy, QNetwork, ValueNetwork, DeterministicPolicy


class SAC(object):
    def __init__(self, num_inputs, action_space, args):

        self.num_inputs = num_inputs
        self.action_space = action_space.shape[0]
        self.gamma = args.gamma
        self.tau = args.tau

        self.policy_type = args.policy
        self.target_update_interval = args.target_update_interval
        self.automatic_entropy_tuning = args.automatic_entropy_tuning

        self.device = torch.device("cuda" if args.cuda else "cpu") 

        self.critic = QNetwork(self.num_inputs, self.action_space, args.hidden_size).to(device=self.device)
        self.critic_optim = Adam(self.critic.parameters(), lr=args.lr)

        if self.policy_type == "Gaussian":
            self.alpha = args.alpha
            # Target Entropy = −dim(A) (e.g. , -6 for HalfCheetah-v2) as given in the paper
            if self.automatic_entropy_tuning == True:
                self.target_entropy = -torch.prod(torch.Tensor(action_space.shape).to(self.device)).item()
                self.log_alpha = torch.zeros(1, requires_grad=True, device=self.device)
                self.alpha_optim = Adam([self.log_alpha], lr=args.lr)
            else:
                pass


            self.policy = GaussianPolicy(self.num_inputs, self.action_space, args.hidden_size).to(self.device)
            self.policy_optim = Adam(self.policy.parameters(), lr=args.lr)

            self.value = ValueNetwork(self.num_inputs, args.hidden_size).to(self.device)
            self.value_target = ValueNetwork(self.num_inputs, args.hidden_size).to(self.device)
            self.value_optim = Adam(self.value.parameters(), lr=args.lr)
            hard_update(self.value_target, self.value)
        else:
            self.policy = DeterministicPolicy(self.num_inputs, self.action_space, args.hidden_size).to(self.device)
            self.policy_optim = Adam(self.policy.parameters(), lr=args.lr)

            self.critic_target = QNetwork(self.num_inputs, self.action_space, args.hidden_size).to(self.device)
            hard_update(self.critic_target, self.critic)



    def select_action(self, state, eval=False):
        state = torch.FloatTensor(state).to(self.device).unsqueeze(0)
        if eval == False:
            self.policy.train()
            action, _, _, _, _ = self.policy.sample(state)
        else:
            self.policy.eval()
            _, _, _, action, _ = self.policy.sample(state)
            if self.policy_type == "Gaussian":
                action = torch.tanh(action)
            else:
                pass
        #action = torch.tanh(action)
        action = action.detach().cpu().numpy()
        return action[0]



    def update_parameters(self, state_batch, action_batch, reward_batch, next_state_batch, mask_batch, updates):
        state_batch = torch.FloatTensor(state_batch).to(self.device)
        next_state_batch = torch.FloatTensor(next_state_batch).to(self.device)
        action_batch = torch.FloatTensor(action_batch).to(self.device)
        reward_batch = torch.FloatTensor(reward_batch).to(self.device).unsqueeze(1)
        mask_batch = torch.FloatTensor(np.float32(mask_batch)).to(self.device).unsqueeze(1)

        """
        Use two Q-functions to mitigate positive bias in the policy improvement step that is known
        to degrade performance of value based methods. Two Q-functions also significantly speed
        up training, especially on harder task.
        """
        expected_q1_value, expected_q2_value = self.critic(state_batch, action_batch)
        new_action, log_prob, _, mean, log_std = self.policy.sample(state_batch)

        if self.policy_type == "Gaussian":
            if self.automatic_entropy_tuning:
                """
                Alpha Loss
                """
                alpha_loss = -(self.log_alpha * (log_prob + self.target_entropy).detach()).mean()
                self.alpha_optim.zero_grad()
                alpha_loss.backward()
                self.alpha_optim.step()
                self.alpha = self.log_alpha.exp()
                alpha_logs = self.alpha.clone() # For TensorboardX logs
            else:
                alpha_loss = torch.tensor(0.).to(self.device)
                alpha_logs = self.alpha # For TensorboardX logs


            """
            Including a separate function approximator for the soft value can stabilize training.
            """
            expected_value = self.value(state_batch)
            target_value = self.value_target(next_state_batch)
            next_q_value = reward_batch + mask_batch * self.gamma * (target_value).detach()
        else:
            """
            There is no need in principle to include a separate function approximator for the state value.
            We use a target critic network for deterministic policy and eradicate the value value network completely.
            """
            alpha_loss = torch.tensor(0.).to(self.device)
            alpha_logs = self.alpha  # For TensorboardX logs
            next_state_action, _, _, _, _, = self.policy.sample(next_state_batch)
            target_critic_1, target_critic_2 = self.critic_target(next_state_batch, next_state_action)
            target_critic = torch.min(target_critic_1, target_critic_2)
            next_q_value = reward_batch + mask_batch * self.gamma * (target_critic).detach()
        
        
        """
        Soft Q-function parameters can be trained to minimize the soft Bellman residual
        JQ = 𝔼(st,at)~D[0.5(Q1(st,at) - r(st,at) - γ(𝔼st+1~p[V(st+1)]))^2]
        ∇JQ = ∇Q(st,at)(Q(st,at) - r(st,at) - γV(target)(st+1))
        """
        q1_value_loss = F.mse_loss(expected_q1_value, next_q_value)
        q2_value_loss = F.mse_loss(expected_q2_value, next_q_value)
        q1_new, q2_new = self.critic(state_batch, new_action)
        expected_new_q_value = torch.min(q1_new, q2_new)

        if self.policy_type == "Gaussian":
            """
            Including a separate function approximator for the soft value can stabilize training and is convenient to 
            train simultaneously with the other networks
            Update the V towards the min of two Q-functions in order to reduce overestimation bias from function approximation error.
            JV = 𝔼st~D[0.5(V(st) - (𝔼at~π[Qmin(st,at) - α * log π(at|st)]))^2]
            ∇JV = ∇V(st)(V(st) - Q(st,at) + (α * logπ(at|st)))
            """
            next_value = expected_new_q_value - (self.alpha * log_prob)
            value_loss = F.mse_loss(expected_value, next_value.detach())
        else:
            pass

        """
        Reparameterization trick is used to get a low variance estimator
        f(εt;st) = action sampled from the policy
        εt is an input noise vector, sampled from some fixed distribution
        Jπ = 𝔼st∼D,εt∼N[α * logπ(f(εt;st)|st) − Q(st,f(εt;st))]
        ∇Jπ = ∇log π + ([∇at (α * logπ(at|st)) − ∇at Q(st,at)])∇f(εt;st)
        """
        policy_loss = ((self.alpha * log_prob) - expected_new_q_value).mean()

        # Regularization Loss
        mean_loss = 0.001 * mean.pow(2).mean()
        std_loss = 0.001 * log_std.pow(2).mean()

        policy_loss += mean_loss + std_loss

        self.critic_optim.zero_grad()
        q1_value_loss.backward()
        self.critic_optim.step()

        self.critic_optim.zero_grad()
        q2_value_loss.backward()
        self.critic_optim.step()

        if self.policy_type == "Gaussian":
            self.value_optim.zero_grad()
            value_loss.backward()
            self.value_optim.step()
        else:
            value_loss = torch.tensor(0.).to(self.device)

        self.policy_optim.zero_grad()
        policy_loss.backward()
        self.policy_optim.step()
        
        
        """
        We update the target weights to match the current value function weights periodically
        Update target parameter after every n(args.target_update_interval) updates
        """
        if updates % self.target_update_interval == 0 and self.policy_type == "Deterministic":
            soft_update(self.critic_target, self.critic, self.tau)

        elif updates % self.target_update_interval == 0 and self.policy_type == "Gaussian":
            soft_update(self.value_target, self.value, self.tau)
        return value_loss.item(), q1_value_loss.item(), q2_value_loss.item(), policy_loss.item(), alpha_loss.item(), alpha_logs

    # Save model parameters
    def save_model(self, env_name, suffix="", actor_path=None, critic_path=None, value_path=None):
        if not os.path.exists('models/'):
            os.makedirs('models/')

        if actor_path is None:
            actor_path = "models/sac_actor_{}_{}".format(env_name, suffix)
        if critic_path is None:
            critic_path = "models/sac_critic_{}_{}".format(env_name, suffix)
        if value_path is None:
            value_path = "models/sac_value_{}_{}".format(env_name, suffix)
        print('Saving models to {}, {} and {}'.format(actor_path, critic_path, value_path))
        torch.save(self.value.state_dict(), value_path)
        torch.save(self.policy.state_dict(), actor_path)
        torch.save(self.critic.state_dict(), critic_path)
    
    # Load model parameters
    def load_model(self, actor_path, critic_path, value_path):
        print('Loading models from {}, {} and {}'.format(actor_path, critic_path, value_path))
        if actor_path is not None:
            self.policy.load_state_dict(torch.load(actor_path))
        if critic_path is not None:
            self.critic.load_state_dict(torch.load(critic_path))
        if value_path is not None:
            self.value.load_state_dict(torch.load(value_path))