ppo_gae_continuous.py

import gym
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.distributions import Categorical
from torch.distributions import Normal
import numpy as np

#Hyperparameters
learning_rate = 1e-4
gamma         = 0.99
lmbda         = 0.95
eps_clip      = 0.1
batch_size    = 1280
K_epoch       = 10
T_horizon     = 10000

class NormalizedActions(gym.ActionWrapper):
    def _action(self, action):
        low  = self.action_space.low
        high = self.action_space.high
        
        action = low + (action + 1.0) * 0.5 * (high - low)
        action = np.clip(action, low, high)
        
        return action

    def _reverse_action(self, action):
        low  = self.action_space.low
        high = self.action_space.high
        
        action = 2 * (action - low) / (high - low) - 1
        action = np.clip(action, low, high)
        
        return action

class PPO(nn.Module):
    def __init__(self, num_inputs, num_actions, hidden_size, action_range = 1.):
        super(PPO, self).__init__()
        self.data = []
        self.action_range = action_range

        self.linear1 = nn.Linear(num_inputs, hidden_size)
        self.linear2 = nn.Linear(hidden_size, hidden_size)
        self.linear3 = nn.Linear(hidden_size, hidden_size)
        self.linear4 = nn.Linear(hidden_size, hidden_size)
        self.linear5 = nn.Linear(hidden_size, hidden_size)
        self.linear6 = nn.Linear(hidden_size, hidden_size)

        self.mean_linear = nn.Linear(hidden_size, num_actions)      
        self.log_std_linear = nn.Linear(hidden_size, num_actions)
        # self.log_std_param = nn.Parameter(torch.zeros(num_actions, requires_grad=True))

        self.v_linear = nn.Linear(hidden_size, 1)

        self.optimizer = optim.Adam(self.parameters(), lr=learning_rate)

    def pi(self, x):

        x = F.tanh(self.linear1(x))
        x = F.tanh(self.linear2(x))
        x1 = F.tanh(self.linear3(x)) 
        x2 = F.tanh(self.linear4(x.detach())) # std learning not BP to the feature

        mean    = F.tanh(self.mean_linear(x1))
        log_std = self.log_std_linear(x2)
        # log_std = self.log_std_param.expand_as(mean)

        return mean, log_std
    
    def v(self, x):
        x = F.tanh(self.linear1(x))
        x = F.tanh(self.linear2(x))
        x = F.tanh(self.linear5(x))
        x = F.tanh(self.linear6(x))

        v = self.v_linear(x)
        return v
      
    def get_action(self, x):
        mean, log_std = self.pi(x)
        std = log_std.exp()
        normal = Normal(mean, std)
        action = normal.sample()
        log_prob = normal.log_prob(action).sum(-1)
        prob = log_prob.exp()

        ## The following way of generating action seems not correct. 
        ## All dimensions of action depends on the same hidden variable z.
        ## In some envs like Ant-v2, it may let the agent not fall easity due to the correlation of actions.
        ## But this does not in general holds true, and may cause numerical problem (nan) in update.
        # normal = Normal(0, 1)  
        # z      = normal.sample()
        # action = mean + std*z
        # log_prob = Normal(mean, std).log_prob(action)
        # log_prob = log_prob.sum(dim=-1, keepdim=True)  # reduce dim
        # prob = log_prob.exp()

        action = self.action_range*action # scale the action

        return action.detach().numpy(), prob

    def get_log_prob(self, mean, log_std, action):
        action = action/self.action_range
        log_prob = Normal(mean, log_std.exp()).log_prob(action)
        log_prob = log_prob.sum(dim=-1, keepdim=True)  # reduce dim
        return log_prob

    def put_data(self, transition):
        self.data.append(transition)
        
    def make_batch(self):
        s_lst, a_lst, r_lst, s_prime_lst, prob_a_lst, done_lst = [], [], [], [], [], []
        for transition in self.data:
            s, a, r, s_prime, prob_a, done = transition
            
            s_lst.append(s)
            a_lst.append(a)
            r_lst.append([r])
            s_prime_lst.append(s_prime)
            prob_a_lst.append([prob_a])
            done_mask = 0 if done else 1
            done_lst.append([done_mask])
        s,a,r,s_prime,done_mask, prob_a = torch.tensor(s_lst, dtype=torch.float), torch.tensor(a_lst), \
                                          torch.tensor(r_lst, dtype=torch.float), torch.tensor(s_prime_lst, dtype=torch.float), \
                                          torch.tensor(done_lst, dtype=torch.float), torch.tensor(prob_a_lst)
        self.data = []
        return s, a, r, s_prime, done_mask, prob_a
        
    def train_net(self):
        s, a, r, s_prime, done_mask, prob_a = self.make_batch()
        done_mask_ = torch.flip(done_mask, dims=(0,))

        # put target value computation before epoch update reduce computation and stabilize training
        td_target = r + gamma * self.v(s_prime) * done_mask
        delta = td_target - self.v(s)
        delta = delta.detach().numpy()

        advantage_lst = []
        advantage = 0.0
        for delta_t, mask in zip(delta[::-1], done_mask_):
            advantage = gamma * lmbda * advantage * mask + delta_t[0]
            advantage_lst.append([advantage])
        advantage_lst.reverse()
        advantage = torch.tensor(advantage_lst, dtype=torch.float)

        if not np.isnan(advantage.std()):
            advantage = (advantage - advantage.mean()) / (advantage.std() + 1e-8) 

        # td_target = advantage + self.v(s)

        for i in range(K_epoch):            
            mean, log_std = self.pi(s)
            log_pi_a = self.get_log_prob(mean, log_std, a)
            # pi = self.pi(s, softmax_dim=1)
            # pi_a = pi.gather(1,a)
            ratio = torch.exp(log_pi_a - torch.log(prob_a))  # a/b == exp(log(a)-log(b))
            surr1 = ratio * advantage
            surr2 = torch.clamp(ratio, 1-eps_clip, 1+eps_clip) * advantage
            loss = -torch.min(surr1, surr2) + F.smooth_l1_loss(self.v(s) , td_target.detach())

            self.optimizer.zero_grad()
            loss.mean().backward()
            self.optimizer.step()
        
def main():
    # env = gym.make('HalfCheetah-v2')
    env = gym.make('Ant-v2')
    state_dim = env.observation_space.shape[0]
    action_dim =  env.action_space.shape[0]
    hidden_dim = 64
    model = PPO(state_dim, action_dim, hidden_dim)
    score = 0.0
    print_interval = 2
    step = 0

    for n_epi in range(10000):
        s = env.reset()
        done = False
        # while not done:
        for t in range(T_horizon):
            step += 1
            a, prob = model.get_action(torch.from_numpy(s).float())
            s_prime, r, done, info = env.step(a)
            # print(a)
            # env.render()

            model.put_data((s, a, r, s_prime, prob, done))
            s = s_prime

            score += r

            if (step+1) % batch_size == 0:
                model.train_net()

            if done:
                break
        if n_epi%print_interval==0 and n_epi!=0:
            print("# of episode :{}, avg score : {:.1f}".format(n_epi, score/print_interval))
            score = 0.0

    env.close()

if __name__ == '__main__':
    main()