C51.py

import math, random, os

import gym
import numpy as np

import torch
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable
import torch.nn.functional as F

from common.replay_buffer import ReplayBuffer

from IPython.display import clear_output
import matplotlib.pyplot as plt

### Use Cuda ###
device = "cuda:1"


### Cart Pole Environment ###
env_id = "CartPole-v0"
env = gym.make(env_id)

### Noisy Networks for Exploration ###
'''
register_buffer: Adds a persistent buffer to the module. 
                 This is typically used to register a buffer that 
                 should not to be considered a model parameter.
'''
class NoisyLinear(nn.Module):
    def __init__(self, in_features, out_features, std_init=0.4):
        super(NoisyLinear, self).__init__()
        
        self.in_features  = in_features
        self.out_features = out_features
        self.std_init     = std_init
        
        self.weight_mu    = nn.Parameter(torch.FloatTensor(out_features, in_features))
        self.weight_sigma = nn.Parameter(torch.FloatTensor(out_features, in_features))
        self.register_buffer('weight_epsilon', torch.FloatTensor(out_features, in_features))
        
        self.bias_mu    = nn.Parameter(torch.FloatTensor(out_features))
        self.bias_sigma = nn.Parameter(torch.FloatTensor(out_features))
        self.register_buffer('bias_epsilon', torch.FloatTensor(out_features))
        
        self.reset_parameters()
        self.reset_noise()
    
    def forward(self, x):
        if self.training: 
            weight = self.weight_mu + self.weight_sigma.mul(Variable(self.weight_epsilon))
            bias   = self.bias_mu   + self.bias_sigma.mul(Variable(self.bias_epsilon))
        else:
            weight = self.weight_mu
            bias   = self.bias_mu
        
        return F.linear(x, weight, bias)
    
    def reset_parameters(self):
        mu_range = 1 / math.sqrt(self.weight_mu.size(1))
        
        self.weight_mu.data.uniform_(-mu_range, mu_range)
        self.weight_sigma.data.fill_(self.std_init / math.sqrt(self.weight_sigma.size(1)))
        
        self.bias_mu.data.uniform_(-mu_range, mu_range)
        self.bias_sigma.data.fill_(self.std_init / math.sqrt(self.bias_sigma.size(0)))
    
    def reset_noise(self):
        epsilon_in  = self._scale_noise(self.in_features)
        epsilon_out = self._scale_noise(self.out_features)
        
        self.weight_epsilon.copy_(epsilon_out.ger(epsilon_in)) # ger: Outer product
        self.bias_epsilon.copy_(self._scale_noise(self.out_features))
    
    def _scale_noise(self, size):
        x = torch.randn(size)
        x = x.sign().mul(x.abs().sqrt())
        return x

### C-51 Algorithm ###
### A Distributional Perspective on Reinforcement Learning ###
class CategoricalDQN(nn.Module):
    def __init__(self, num_inputs, num_actions, num_atoms, Vmin, Vmax):
        super(CategoricalDQN, self).__init__()
        
        self.num_inputs  = num_inputs
        self.num_actions = num_actions
        self.num_atoms   = num_atoms
        
        self.Vmin = Vmin
        self.Vmax = Vmax
        
        self.linear1 = nn.Linear(num_inputs, 128)
        self.linear2 = nn.Linear(128, 128)
        self.noisy1 = NoisyLinear(128, 512)
        self.noisy2 = NoisyLinear(512, self.num_actions * self.num_atoms)
        
    def forward(self, x):
        x = F.relu(self.linear1(x))
        x = F.relu(self.linear2(x))
        x = F.relu(self.noisy1(x))
        x = self.noisy2(x)
        x = nn.Softmax(dim=1)(x.view(-1, self.num_atoms)).view(-1, self.num_actions, self.num_atoms)
        return x
    
    def act(self, state):
        state = Variable(torch.FloatTensor(state).unsqueeze(0)).to(device)
        dist = self.forward(state).data.cpu()
        dist = dist * torch.linspace(Vmin, Vmax, num_atoms)
        action = dist.sum(2).max(1)[1].numpy()[0]
        return action

    def reset_noise(self):
        self.noisy1.reset_noise()
        self.noisy2.reset_noise()

def projection_distribution(next_state, rewards, dones):
    batch_size = next_state.size(0)
    
    delta_z = float(Vmax - Vmin) / (num_atoms - 1)
    support = torch.linspace(Vmin, Vmax, num_atoms).to(device)
    
    next_dist = target_model(next_state) * support
    next_action = next_dist.sum(2).max(1)[1]
    next_action = next_action.unsqueeze(1).unsqueeze(1).expand(next_dist.size(0), 1, next_dist.size(2))
    next_dist = next_dist.gather(1, next_action).squeeze(1)
        
    rewards = rewards.unsqueeze(1).expand_as(next_dist)
    dones   = dones.unsqueeze(1).expand_as(next_dist)
    support = support.unsqueeze(0).expand_as(next_dist)
    
    Tz = rewards + (1 - dones) * 0.99 * support
    Tz = Tz.clamp(min=Vmin, max=Vmax)
    b  = (Tz - Vmin) / delta_z
    l  = b.floor().long() # Lower bound
    u  = b.ceil().long()  # Upper bound
        
    offset = torch.linspace(0, (batch_size - 1) * num_atoms, batch_size).long()\
                    .unsqueeze(1).expand(batch_size, num_atoms).to(device)

    # Equation (7)
    proj_dist = torch.zeros(next_dist.size()).to(device)
    proj_dist.view(-1).index_add_(0, (l + offset).view(-1), (next_dist * (u.float() - b)).view(-1))
    proj_dist.view(-1).index_add_(0, (u + offset).view(-1), (next_dist * (b - l.float())).view(-1))
        
    return proj_dist

num_atoms = 51
Vmin = -10
Vmax = 10

current_model = CategoricalDQN(env.observation_space.shape[0], env.action_space.n, num_atoms, Vmin, Vmax).to(device)
target_model  = CategoricalDQN(env.observation_space.shape[0], env.action_space.n, num_atoms, Vmin, Vmax).to(device)
    
optimizer = optim.Adam(current_model.parameters())

replay_buffer = ReplayBuffer(10000)

### Synchronize current policy net and target net ###
def update_target(current_model, target_model):
    target_model.load_state_dict(current_model.state_dict())
    
update_target(current_model, target_model)

### Computing Temporal Difference Loss ###
# torch.gather: Gathers values along an axis specified by dim
def compute_td_loss(batch_size):
    state, action, reward, next_state, done = replay_buffer.sample(batch_size) 

    state      = Variable(torch.FloatTensor(np.float32(state))).to(device)
    next_state = Variable(torch.FloatTensor(np.float32(next_state))).to(device)
    action     = Variable(torch.LongTensor(action)).to(device)
    reward     = torch.FloatTensor(reward).to(device)
    done       = torch.FloatTensor(np.float32(done)).to(device)

    proj_dist = projection_distribution(next_state, reward, done)
    
    dist = current_model(state)
    action = action.unsqueeze(1).unsqueeze(1).expand(batch_size, 1, num_atoms)
    dist = dist.gather(1, action).squeeze(1)
    dist.data.clamp_(0.01, 0.99)
    loss = - (proj_dist * dist.log()).sum(1).mean()
        
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()
    
    current_model.reset_noise()
    target_model.reset_noise()
    
    return loss

def CartPole_plot(frame_idx, rewards, losses):
    clear_output(True)
    plt.figure(figsize=(20,5))
    plt.subplot(131)
    plt.title('frame %s. reward: %s' % (frame_idx, np.mean(rewards[-10:])))
    plt.plot(rewards)
    plt.subplot(132)
    plt.title('loss')
    plt.plot(losses)
    plt.savefig('img/C51_CartPole_%s.png' % (frame_idx))
    plt.cla()
    plt.close("all")


### Training CartPole ###
num_frames = 20000
batch_size = 32
gamma = 0.99

losses = []
all_rewards = []
episode_reward = 0

state = env.reset()
for frame_idx in range(1, num_frames + 1):
    action = current_model.act(state)
    
    next_state, reward, done, _ = env.step(action)
    replay_buffer.push(state, action, reward, next_state, done)
    
    state = next_state
    episode_reward += reward
    
    if done:
        state = env.reset()
        all_rewards.append(episode_reward)
        episode_reward = 0
        
    if len(replay_buffer) > batch_size:
        loss = compute_td_loss(batch_size)
        losses.append(loss.item())
        
    if frame_idx % 1000 == 0:
        CartPole_plot(frame_idx, all_rewards, losses)
        if frame_idx > 1000:
            os.system('rm img/C51_CartPole_%s.png' % (frame_idx - 1000))
        
    if frame_idx % 100 == 0:
        update_target(current_model, target_model)


### Atari Environment ###
from common.wrappers import make_atari, wrap_deepmind, wrap_pytorch

env_id = "PongNoFrameskip-v4"
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)

class CategoricalCnnDQN(nn.Module):
    def __init__(self, input_shape, num_actions, num_atoms, Vmin, Vmax):
        super(CategoricalCnnDQN, self).__init__()
        
        self.input_shape  = input_shape
        self.num_actions  = num_actions
        self.num_atoms    = num_atoms
        self.Vmin         = Vmin
        self.Vmax         = Vmax
        
        self.features = nn.Sequential(
            nn.Conv2d(input_shape[0], 32, kernel_size=8, stride=4),
            nn.ReLU(),
            nn.Conv2d(32, 64, kernel_size=4, stride=2),
            nn.ReLU(),
            nn.Conv2d(64, 64, kernel_size=3, stride=1),
            nn.ReLU()
        )
        
        self.noisy1 = NoisyLinear(self.feature_size(), 512)
        self.noisy2 = NoisyLinear(512, self.num_actions * self.num_atoms)
        
    def forward(self, x):
        batch_size = x.size(0)
        
        x = x / 255.
        x = self.features(x)
        x = x.view(batch_size, -1)
        
        x = F.relu(self.noisy1(x))
        x = self.noisy2(x)
        x = nn.Softmax(dim=1)(x.view(-1, self.num_atoms)).view(-1, self.num_actions, self.num_atoms)
        return x
        
    def reset_noise(self):
        self.noisy1.reset_noise()
        self.noisy2.reset_noise()
        
    def feature_size(self):
        return self.features(Variable(torch.zeros(1, *self.input_shape))).view(1, -1).size(1)
    
    def act(self, state):
        state = Variable(torch.FloatTensor(np.float32(state)).unsqueeze(0)).to(device)
        dist = self.forward(state).data.cpu()
        dist = dist * torch.linspace(Vmin, Vmax, num_atoms)
        action = dist.sum(2).max(1)[1].numpy()[0]
        return action

def Atari_plot(frame_idx, rewards, losses):
    clear_output(True)
    plt.figure(figsize=(20,5))
    plt.subplot(131)
    plt.title('frame %s. reward: %s' % (frame_idx, np.mean(rewards[-10:])))
    plt.plot(rewards)
    plt.subplot(132)
    plt.title('loss')
    plt.plot(losses)
    plt.savefig('img/C51_Atari_%s.png' % (frame_idx))
    plt.cla()
    plt.close("all")

num_atoms = 51
Vmin = -10
Vmax = 10

current_model = CategoricalCnnDQN(env.observation_space.shape, env.action_space.n, num_atoms, Vmin, Vmax).to(device)
target_model  = CategoricalCnnDQN(env.observation_space.shape, env.action_space.n, num_atoms, Vmin, Vmax).to(device)
    
optimizer = optim.Adam(current_model.parameters(), 0.0001)
update_target(current_model, target_model)

replay_initial = 10000
replay_buffer = ReplayBuffer(100000)


### Training Atari ###
num_frames = 1000000
batch_size = 32
gamma = 0.99

losses = []
all_rewards = []
episode_reward = 0

state = env.reset()
for frame_idx in range(1, num_frames + 1):
    action = current_model.act(state)
    
    next_state, reward, done, _ = env.step(action)
    replay_buffer.push(state, action, reward, next_state, done)
    
    state = next_state
    episode_reward += reward
    
    if done:
        state = env.reset()
        all_rewards.append(episode_reward)
        episode_reward = 0
        
    if len(replay_buffer) > replay_initial:
        loss = compute_td_loss(batch_size)
        losses.append(loss.item())
        
    if frame_idx % 10000 == 0:
        Atari_plot(frame_idx, all_rewards, losses)
        if frame_idx > 10000:
            os.system('rm img/C51_Atari_%s.png' % (frame_idx - 10000))
        
    if frame_idx % 1000 == 0:
        update_target(current_model, target_model)