dqn.py

import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from collections import namedtuple, deque
from pathlib import Path
import random
from tqdm import trange
from itertools import count
import matplotlib.pyplot as plt
import matplotlib
import time
import math
from fta import FTA
import pickle

is_ipython = 'inline' in matplotlib.get_backend()
if is_ipython:
    from IPython import display

plt.ion()

Transition = namedtuple('Transition',
                        ('state', 'action', 'next_state', 'reward', 'done'))

class ReplayMemory(object):

    def __init__(self, capacity):
        self.capacity = capacity
        self.memory = []
        self.position = 0
        self.rng = np.random.default_rng()

    def push(self, *args):
        """Saves a transition."""
        if len(self.memory) < self.capacity:
            self.memory.append(None)
        self.memory[self.position] = Transition(*args)
        self.position = (self.position + 1) % self.capacity

    def sample(self, batch_size):
        idx = self.rng.choice(np.arange(len(self.memory)), batch_size, replace=False)
        res = []
        for i in idx:
            res.append(self.memory[i])
        return res
        # return self.rng.choice(self.memory, batch_size, replace=False)

    def __len__(self):
        return len(self.memory)

class DQN(nn.Module):

    def __init__(self, input, outputs, activation='relu', device='cpu'):
        super(DQN, self).__init__()
        self.tiles = 10
        self.emb = nn.Embedding(input, 4)
        self.l1 = nn.Linear(4, 50)
        self.l2 = nn.Linear(50, 50)
        self.l3 = nn.Linear(50*self.tiles, outputs)
        if activation == 'relu':
            self.activation = F.relu
        elif activation == 'fta':
            self.activation = FTA(tiles=self.tiles, bound_low=-1, bound_high=1, eta=0.2, input_dim=50, device=device)

    def forward(self, x):
        x = F.relu(self.l1(self.emb(x)))
        x = self.activation(self.l2(x))
        x = self.l3(x)
        return x
           
class QAgent():
    def __init__(self, env):
        self.activation = 'fta'
        self.env = env
        self.num_episodes = 10000
        self.model_dir = Path('.models')
        self.reward_dir = Path('.rewards')

        self.save_ratio = 250
        self.device = torch.device("cuda")
        self.batch_size = 128
        self.gamma = 0.99
        self.eps_start = 1
        self.eps_end = 0.1
        self.eps_decay = 10000
        self.target_update = 100
        self.learning_rate = 1e-3
        self.max_episode = 100
        self.id = time.time()
        
        if not os.path.exists(self.model_dir):
            os.makedirs(self.model_dir)
        if not os.path.exists(self.reward_dir):
            os.makedirs(self.reward_dir)
            
        self.action_space = env.action_space.n
        self.observation_space = env.observation_space.n
        
        self.policy_net = DQN(self.observation_space, self.action_space, activation=self.activation, device=self.device).to(self.device)
        self.target_net = DQN(self.observation_space, self.action_space, activation=self.activation, device=self.device).to(self.device)
        self.target_net.load_state_dict(self.policy_net.state_dict())
        
        self.loss_fn = nn.SmoothL1Loss()
        self.optimizer = torch.optim.AdamW(self.policy_net.parameters(), lr=self.learning_rate, amsgrad=True)
        
        self.memory = ReplayMemory(100000)
        
        self.steps_done = 0
        self.reward_in_episode = []
    
    
    def _save(self):
        torch.save({
            'model_state_dics': self.policy_net.state_dict(), 
            'optimizer_state_dict': self.optimizer.state_dict(), 
            'reward_in_episode': self.reward_in_episode
        }, f'{self.model_dir}/pytorch_{self.id}.pt')
        
    def select_action(self, state):
        sample = random.random()
        eps_threshold = self.eps_end + (self.eps_start - self.eps_end) * math.exp(-1. * self.steps_done / self.eps_decay)
        self.steps_done += 1
        if sample < eps_threshold:
            return self.env.action_space.sample()
        else:
            with torch.no_grad():
                return self.policy_net(torch.tensor([state], device=self.device)).max(1)[1].item()
        
    def plot_rewards(self, show_result=False):
        plt.figure(1)
        rewards_t = torch.tensor(self.reward_in_episode, dtype=torch.float)
        if show_result:
            plt.title('Result')
        else:
            plt.clf()
            plt.title('Training...')
        plt.xlabel('Episode')
        plt.ylabel('Rewards')
        plt.plot(rewards_t.numpy())
        # Take 100 episode averages and plot them too
        if len(rewards_t) >= 100:
            means = rewards_t.unfold(0, 100, 1).mean(1).view(-1)
            means = torch.cat((torch.zeros(99), means))
            plt.plot(means.numpy())

        plt.pause(0.001)  # pause a bit so that plots are updated
        if is_ipython:
            if not show_result:
                display.display(plt.gcf())
                display.clear_output(wait=True)
            else:
                display.display(plt.gcf())
                
                
    def optimize(self):
        if len(self.memory) < self.batch_size:
            return
        transitions = self.memory.sample(self.batch_size)
        batch = Transition(*zip(*transitions))
        
        state_batch = torch.cat(batch.state)
        action_batch = torch.cat(batch.action)
        reward_batch = torch.cat(batch.reward)
        next_state_batch = torch.cat(batch.next_state)
        done_batch = torch.cat(batch.done)
        
        action_values = self.policy_net(state_batch).gather(1, action_batch.unsqueeze(1))
        
        next_values = self.target_net(next_state_batch).max(1)[0]
        expected_action_values = (~done_batch * next_values * self.gamma) + reward_batch
        
        loss = self.loss_fn(action_values, expected_action_values.unsqueeze(1))
        
        self.optimizer.zero_grad()
        loss.backward()
        torch.nn.utils.clip_grad_value_(self.policy_net.parameters(), 100)
        self.optimizer.step()
      
      
    def _remember(self, state, action, next_state, reward, done):
        self.memory.push(torch.tensor([state], device=self.device),
                        torch.tensor([action], device=self.device, dtype=torch.long),
                        torch.tensor([next_state], device=self.device),
                        torch.tensor([reward], device=self.device),
                        torch.tensor([done], device=self.device, dtype=torch.bool))
  
    def train(self):
        for i in trange(self.num_episodes):
            
            state, _ = self.env.reset()
            done = False
            reward_in_episode = 0
            for t in count():
                
                action = self.select_action(state=state)
                next_state, reward, terminated, truncated, _ = self.env.step(action)
                done = terminated or truncated
                
                self._remember(state, action, next_state, reward, done)
                
                self.optimize()
                state = next_state
                reward_in_episode += reward

                # target_net_state_dict = self.target_net.state_dict()
                # policy_net_state_dict = self.policy_net.state_dict()
                # for key in policy_net_state_dict:
                #     target_net_state_dict[key] = policy_net_state_dict[key]*self.tau + target_net_state_dict[key]*(1-self.tau)
                # self.target_net.load_state_dict(target_net_state_dict)
                
                if done or t > self.max_episode:
                    self.reward_in_episode.append(reward_in_episode)
                    self.plot_rewards()
                    break
                
            if i % self.target_update == 0:
                self.target_net.load_state_dict(self.policy_net.state_dict())
            
            if i % self.save_ratio == 0:
                # self._save()
                # torch.save(self, f'{self.model_dir}/pytorch_{self.id}.pt')
                
                with open(f'{self.reward_dir}/rewards_{self.id}.pkl', 'wb') as fp:
                    pickle.dump(self.reward_in_episode, fp)
        self.plot_rewards(show_result=True)
        plt.ioff()
        plt.show()