agent.py

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.distributions import Independent
from torch.distributions import Normal
from torch.distributions import TransformedDistribution
from torch.distributions import Transform
from torch.distributions import constraints

from gym.spaces import flatdim
from lagom import BaseAgent
from lagom.utils import tensorify
from lagom.utils import numpify
from lagom.networks import Module
from lagom.networks import make_fc
from lagom.networks import ortho_init
from lagom.transform import describe


# TODO: import from PyTorch when PR merged: https://github.com/pytorch/pytorch/pull/19785
class TanhTransform(Transform):
    r"""Transform via the mapping :math:`y = \tanh(x)`."""
    domain = constraints.real
    codomain = constraints.interval(-1.0, 1.0)
    bijective = True
    sign = +1

    @staticmethod
    def atanh(x):
        return 0.5 * (x.log1p() - (-x).log1p())

    def __eq__(self, other):
        return isinstance(other, TanhTransform)

    def _call(self, x):
        return x.tanh()

    def _inverse(self, y):
        return self.atanh(y)

    def log_abs_det_jacobian(self, x, y):
        # We use a formula that is more numerically stable, see details in the following link
        # https://github.com/tensorflow/probability/commit/ef6bb176e0ebd1cf6e25c6b5cecdd2428c22963f#diff-e120f70e92e6741bca649f04fcd907b7
        return 2. * (np.log(2.) - x - F.softplus(-2. * x))
    

class Actor(Module):
    LOGSTD_MAX = 2
    LOGSTD_MIN = -20

    def __init__(self, config, env, device, **kwargs):
        super().__init__(**kwargs)
        self.config = config
        self.env = env
        self.device = device
        
        self.feature_layers = make_fc(flatdim(env.observation_space), [256, 256])
        self.mean_head = nn.Linear(256, flatdim(env.action_space))
        self.logstd_head = nn.Linear(256, flatdim(env.action_space))
        
        self.to(device)

    def forward(self, x):
        for layer in self.feature_layers:
            x = F.relu(layer(x))
        mean = self.mean_head(x)
        logstd = self.logstd_head(x)
        logstd = torch.tanh(logstd)
        logstd = self.LOGSTD_MIN + 0.5*(self.LOGSTD_MAX - self.LOGSTD_MIN)*(1 + logstd)
        std = torch.exp(logstd)
        dist = TransformedDistribution(Independent(Normal(mean, std), 1), [TanhTransform(cache_size=1)])
        return dist
    
    def mean_forward(self, x):
        for layer in self.feature_layers:
            x = F.relu(layer(x))
        mean = self.mean_head(x)
        return mean


class Critic(Module):
    def __init__(self, config, env, device, **kwargs):
        super().__init__(**kwargs)
        self.config = config
        self.env = env
        self.device = device
        
        # Q1
        self.first_feature_layers = make_fc(flatdim(env.observation_space) + flatdim(env.action_space), [256, 256])
        self.first_Q_head = nn.Linear(256, 1)
        
        # Q2
        self.second_feature_layers = make_fc(flatdim(env.observation_space) + flatdim(env.action_space), [256, 256])
        self.second_Q_head = nn.Linear(256, 1)
        
        self.to(self.device)
        
    def Q1(self, x, action):
        x = torch.cat([x, action], dim=-1)
        for layer in self.first_feature_layers:
            x = F.relu(layer(x))
        x = self.first_Q_head(x)
        return x
    
    def Q2(self, x, action):
        x = torch.cat([x, action], dim=-1)
        for layer in self.second_feature_layers:
            x = F.relu(layer(x))
        x = self.second_Q_head(x)
        return x
        
    def forward(self, x, action):
        return self.Q1(x, action), self.Q2(x, action)


class Agent(BaseAgent):
    def __init__(self, config, env, device, **kwargs):
        super().__init__(config, env, device, **kwargs)
        
        self.actor = Actor(config, env, device, **kwargs)
        self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=config['agent.actor.lr'])
        
        self.critic = Critic(config, env, device, **kwargs)
        self.critic_target = Critic(config, env, device, **kwargs)
        self.critic_target.load_state_dict(self.critic.state_dict())
        self.critic_optimizer = optim.Adam(self.critic.parameters(), lr=config['agent.critic.lr'])
        
        self.target_entropy = -float(flatdim(env.action_space))
        self.log_alpha = nn.Parameter(torch.tensor(np.log(config['agent.initial_temperature'])).to(device))
        self.log_alpha_optimizer = optim.Adam([self.log_alpha], lr=config['agent.actor.lr'])
        
        self.optimizer_zero_grad = lambda: [opt.zero_grad() for opt in [self.actor_optimizer, 
                                                                        self.critic_optimizer, 
                                                                        self.log_alpha_optimizer]]
        self.total_timestep = 0
        
    @property
    def alpha(self):
        return self.log_alpha.exp()

    def polyak_update_target(self):
        p = self.config['agent.polyak']
        for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()):
            target_param.data.copy_(p*target_param.data + (1 - p)*param.data)

    def choose_action(self, x, **kwargs):
        obs = tensorify(x.observation, self.device).unsqueeze(0)
        with torch.no_grad():
            if kwargs['mode'] == 'train':
                action = numpify(self.actor(obs).sample(), 'float')
            elif kwargs['mode'] == 'eval':
                action = numpify(torch.tanh(self.actor.mean_forward(obs)), 'float')
        out = {}
        out['raw_action'] = action.squeeze(0)
        return out

    def learn(self, D, **kwargs):
        replay = kwargs['replay']
        T = kwargs['T']
        list_actor_loss = []
        list_critic_loss = []
        list_alpha_loss = []
        Q1_vals = []
        Q2_vals = []
        logprob_vals = []
        for i in range(T):
            observations, actions, rewards, next_observations, masks = replay.sample(self.config['replay.batch_size'])
            
            Qs1, Qs2 = self.critic(observations, actions)
            with torch.no_grad():
                action_dist = self.actor(next_observations)
                next_actions = action_dist.rsample()
                next_actions_logprob = action_dist.log_prob(next_actions).unsqueeze(-1)
                next_Qs1, next_Qs2 = self.critic_target(next_observations, next_actions)
                next_Qs = torch.min(next_Qs1, next_Qs2) - self.alpha.detach()*next_actions_logprob
                targets = rewards + self.config['agent.gamma']*masks*next_Qs
            critic_loss = F.mse_loss(Qs1, targets.detach()) + F.mse_loss(Qs2, targets.detach())
            self.optimizer_zero_grad()
            critic_loss.backward()
            critic_grad_norm = nn.utils.clip_grad_norm_(self.critic.parameters(), self.config['agent.max_grad_norm'])
            self.critic_optimizer.step()
            
            action_dist = self.actor(observations)
            policy_actions = action_dist.rsample()
            policy_actions_logprob = action_dist.log_prob(policy_actions).unsqueeze(-1)
            actor_Qs1, actor_Qs2 = self.critic(observations, policy_actions)
            actor_Qs = torch.min(actor_Qs1, actor_Qs2)
            actor_loss = torch.mean(self.alpha.detach()*policy_actions_logprob - actor_Qs)
            self.optimizer_zero_grad()
            actor_loss.backward()
            actor_grad_norm = nn.utils.clip_grad_norm_(self.actor.parameters(), self.config['agent.max_grad_norm'])
            self.actor_optimizer.step()
            
            alpha_loss = torch.mean(self.log_alpha*(-policy_actions_logprob - self.target_entropy).detach())
            self.optimizer_zero_grad()
            alpha_loss.backward()
            self.log_alpha_optimizer.step()

            self.polyak_update_target()
            list_actor_loss.append(actor_loss)
            list_critic_loss.append(critic_loss)
            list_alpha_loss.append(alpha_loss)
            Q1_vals.append(Qs1)
            Q2_vals.append(Qs2)
            logprob_vals.append(policy_actions_logprob)
        self.total_timestep += T
        
        out = {}
        out['actor_loss'] = torch.tensor(list_actor_loss).mean(0).item()
        out['actor_grad_norm'] = actor_grad_norm
        out['critic_loss'] = torch.tensor(list_critic_loss).mean(0).item()
        out['critic_grad_norm'] = critic_grad_norm
        describe_it = lambda x: describe(numpify(torch.cat(x), 'float').squeeze(), axis=-1, repr_indent=1, repr_prefix='\n')
        out['Q1'] = describe_it(Q1_vals)
        out['Q2'] = describe_it(Q2_vals)
        out['logprob'] = describe_it(logprob_vals)
        out['alpha_loss'] = torch.tensor(list_alpha_loss).mean(0).item()
        out['alpha'] = self.alpha.item()
        return out
    
    def checkpoint(self, logdir, num_iter):
        self.save(logdir/f'agent_{num_iter}.pth')