Spaces:
Build error
Build error
File size: 15,138 Bytes
7101384 15d57a8 7101384 6b92232 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 |
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.distributions import Categorical
import numpy as np
import ale_py
import gymnasium as gym
import matplotlib.pyplot as plt
from collections import deque
# Register ALE environments
gym.register_envs(ale_py)
# Set random seeds for reproducibility
torch.manual_seed(42)
np.random.seed(42)
# Check if GPU is available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")
# ==================== Policy Networks ====================
class CartPolePolicy(nn.Module):
"""Policy network for CartPole environment"""
def __init__(self, state_dim, action_dim, hidden_dim=128):
super(CartPolePolicy, self).__init__()
self.fc1 = nn.Linear(state_dim, hidden_dim)
self.fc2 = nn.Linear(hidden_dim, action_dim)
# Initialize weights
self._initialize_weights()
def _initialize_weights(self):
"""Initialize network weights"""
for m in self.modules():
if isinstance(m, nn.Linear):
nn.init.xavier_uniform_(m.weight)
nn.init.constant_(m.bias, 0.0)
def forward(self, x):
x = F.relu(self.fc1(x))
x = self.fc2(x)
return F.softmax(x, dim=-1)
class PongPolicy(nn.Module):
"""Policy network for Pong with CNN architecture"""
def __init__(self, action_dim=2):
super(PongPolicy, self).__init__()
# CNN layers for processing 80x80 images
self.conv1 = nn.Conv2d(1, 16, kernel_size=8, stride=4)
self.conv2 = nn.Conv2d(16, 32, kernel_size=4, stride=2)
# Calculate size after convolutions: 80 -> 19 -> 8
self.fc1 = nn.Linear(32 * 8 * 8, 256)
self.fc2 = nn.Linear(256, action_dim)
# Initialize weights for better training stability
self._initialize_weights()
def _initialize_weights(self):
"""Initialize network weights with proper initialization"""
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_uniform_(m.weight, nonlinearity='relu')
if m.bias is not None:
nn.init.constant_(m.bias, 0.0)
elif isinstance(m, nn.Linear):
nn.init.xavier_uniform_(m.weight)
nn.init.constant_(m.bias, 0.0)
def forward(self, x):
# x shape: (batch, 80, 80) -> add channel dimension
if len(x.shape) == 2:
x = x.unsqueeze(0).unsqueeze(0)
elif len(x.shape) == 3:
x = x.unsqueeze(1)
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
x = x.view(x.size(0), -1)
x = F.relu(self.fc1(x))
x = self.fc2(x)
return F.softmax(x, dim=-1)
# ==================== Helper Functions ====================
def preprocess(image):
"""Prepro 210x160x3 uint8 frame into 6400 (80x80) 2D float array"""
image = image[35:195] # crop
image = image[::2, ::2, 0] # downsample by factor of 2
image[image == 144] = 0 # erase background (background type 1)
image[image == 109] = 0 # erase background (background type 2)
image[image != 0] = 1 # everything else (paddles, ball) just set to 1
return np.reshape(image.astype(float).ravel(), [80, 80])
def compute_returns(rewards, gamma):
"""Compute discounted returns for each timestep"""
returns = []
R = 0
for r in reversed(rewards):
R = r + gamma * R
returns.insert(0, R)
returns = torch.tensor(returns, dtype=torch.float32).to(device)
# Normalize returns for more stable training
if len(returns) > 1:
returns = (returns - returns.mean()) / (returns.std() + 1e-8)
return returns
def moving_average(data, window_size):
"""Compute moving average"""
if len(data) < window_size:
return np.array([np.mean(data[:i+1]) for i in range(len(data))])
moving_avg = []
for i in range(len(data)):
if i < window_size:
moving_avg.append(np.mean(data[:i+1]))
else:
moving_avg.append(np.mean(data[i-window_size+1:i+1]))
return np.array(moving_avg)
# ==================== Policy Gradient Algorithm ====================
def train_policy_gradient(env_name, policy, optimizer, gamma, num_episodes,
max_steps=None, is_pong=False, action_map=None):
"""
Train policy using REINFORCE algorithm
Args:
env_name: Name of the gym environment
policy: Policy network
optimizer: PyTorch optimizer
gamma: Discount factor
num_episodes: Number of training episodes
max_steps: Maximum steps per episode (None for default)
is_pong: Whether this is Pong environment
action_map: Mapping from policy action to env action (for Pong)
"""
env = gym.make(env_name)
episode_rewards = []
for episode in range(num_episodes):
state, _ = env.reset()
# Preprocess state for Pong
if is_pong:
state = preprocess(state)
prev_frame = None # Track previous frame for motion
log_probs = []
rewards = []
done = False
step = 0
while not done:
# For Pong, use frame difference (motion signal)
if is_pong:
cur_frame = state
if prev_frame is not None:
state_input = cur_frame - prev_frame
else:
state_input = np.zeros_like(cur_frame, dtype=np.float32)
prev_frame = cur_frame
state_tensor = torch.FloatTensor(state_input).to(device)
else:
# Convert state to tensor
state_tensor = torch.FloatTensor(state).to(device)
# Get action probabilities
action_probs = policy(state_tensor)
# Sample action from the distribution
dist = Categorical(action_probs)
action = dist.sample()
log_prob = dist.log_prob(action)
# Map action for Pong (0,1 -> 2,3)
if is_pong:
env_action = action_map[action.item()]
else:
env_action = action.item()
# Take action in environment
next_state, reward, terminated, truncated, _ = env.step(env_action)
done = terminated or truncated
# Preprocess next state for Pong
if is_pong:
next_state = preprocess(next_state)
# Store log probability and reward
log_probs.append(log_prob)
rewards.append(reward)
state = next_state
step += 1
if max_steps and step >= max_steps:
break
# Compute returns
returns = compute_returns(rewards, gamma)
# Compute policy gradient loss
policy_loss = []
for log_prob, R in zip(log_probs, returns):
policy_loss.append(-log_prob * R)
# Optimize policy
optimizer.zero_grad()
policy_loss = torch.stack(policy_loss).sum()
policy_loss.backward()
# Gradient clipping for training stability
torch.nn.utils.clip_grad_norm_(policy.parameters(), max_norm=1.0)
optimizer.step()
# Record episode reward
episode_reward = sum(rewards)
episode_rewards.append(episode_reward)
# Print progress
if (episode + 1) % 100 == 0:
avg_reward = np.mean(episode_rewards[-100:])
print(f"Episode {episode + 1}/{num_episodes}, "
f"Avg Reward (last 100): {avg_reward:.2f}")
# Save checkpoint for Pong every 500 episodes
if is_pong and (episode + 1) % 500 == 0:
checkpoint_path = f'pong_checkpoint_ep{episode + 1}.pth'
torch.save({
'episode': episode + 1,
'policy_state_dict': policy.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'episode_rewards': episode_rewards,
}, checkpoint_path)
print(f" → Checkpoint saved: {checkpoint_path}")
env.close()
return episode_rewards
def evaluate_policy(env_name, policy, num_episodes=500, is_pong=False, action_map=None):
"""Evaluate trained policy over multiple episodes"""
env = gym.make(env_name)
eval_rewards = []
for episode in range(num_episodes):
state, _ = env.reset()
if is_pong:
state = preprocess(state)
prev_frame = None # Track previous frame for motion
episode_reward = 0
done = False
while not done:
# For Pong, use frame difference (motion signal)
if is_pong:
cur_frame = state
if prev_frame is not None:
state_input = cur_frame - prev_frame
else:
state_input = np.zeros_like(cur_frame, dtype=np.float32)
prev_frame = cur_frame
state_tensor = torch.FloatTensor(state_input).to(device)
else:
state_tensor = torch.FloatTensor(state).to(device)
with torch.no_grad():
action_probs = policy(state_tensor)
action = torch.argmax(action_probs).item()
if is_pong:
env_action = action_map[action]
else:
env_action = action
next_state, reward, terminated, truncated, _ = env.step(env_action)
done = terminated or truncated
if is_pong:
next_state = preprocess(next_state)
episode_reward += reward
state = next_state
eval_rewards.append(episode_reward)
if (episode + 1) % 100 == 0:
print(f"Evaluated {episode + 1}/{num_episodes} episodes")
env.close()
return eval_rewards
def plot_results(episode_rewards, eval_rewards, title, save_prefix):
"""Plot training curve and evaluation histogram"""
fig, axes = plt.subplots(1, 2, figsize=(15, 5))
# Plot training curve
ax1 = axes[0]
episodes = np.arange(1, len(episode_rewards) + 1)
ma = moving_average(episode_rewards, 100)
ax1.plot(episodes, episode_rewards, alpha=0.3, label='Episode Reward')
ax1.plot(episodes, ma, linewidth=2, label='Moving Average (100 episodes)')
ax1.set_xlabel('Episode')
ax1.set_ylabel('Reward')
ax1.set_title(f'{title} - Training Curve')
ax1.legend()
ax1.grid(True, alpha=0.3)
# Plot evaluation histogram
ax2 = axes[1]
mean_reward = np.mean(eval_rewards)
std_reward = np.std(eval_rewards)
ax2.hist(eval_rewards, bins=30, edgecolor='black', alpha=0.7)
ax2.axvline(mean_reward, color='red', linestyle='--', linewidth=2,
label=f'Mean: {mean_reward:.2f}')
ax2.set_xlabel('Episode Reward')
ax2.set_ylabel('Frequency')
ax2.set_title(f'{title} - Evaluation Histogram (500 episodes)\n'
f'Mean: {mean_reward:.2f}, Std: {std_reward:.2f}')
ax2.legend()
ax2.grid(True, alpha=0.3, axis='y')
plt.tight_layout()
plt.savefig(f'{save_prefix}_results.png', dpi=150, bbox_inches='tight')
plt.show()
print(f"\n{title} Evaluation Results:")
print(f"Mean Reward: {mean_reward:.2f}")
print(f"Std Reward: {std_reward:.2f}")
# ==================== Main Training Scripts ====================
def train_cartpole():
"""Train CartPole-v1"""
print("\n" + "="*60)
print("Training CartPole-v1")
print("="*60 + "\n")
# Environment parameters
env = gym.make('CartPole-v1')
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.n
env.close()
# Hyperparameters
gamma = 0.95
learning_rate = 0.01
num_episodes = 1000
# Initialize policy and optimizer
policy = CartPolePolicy(state_dim, action_dim).to(device)
optimizer = optim.Adam(policy.parameters(), lr=learning_rate)
# Train
episode_rewards = train_policy_gradient(
'CartPole-v1', policy, optimizer, gamma, num_episodes
)
# Evaluate
print("\nEvaluating trained policy...")
eval_rewards = evaluate_policy('CartPole-v1', policy, num_episodes=500)
# Plot results
plot_results(episode_rewards, eval_rewards, 'CartPole-v1', 'cartpole')
# Save model
torch.save(policy.state_dict(), 'cartpole_policy.pth')
print("\nModel saved as 'cartpole_policy.pth'")
return policy, episode_rewards, eval_rewards
def train_pong():
"""Train Pong-v5"""
print("\n" + "="*60)
print("Training Pong-v5")
print("="*60 + "\n")
# Hyperparameters
gamma = 0.99
learning_rate = 0.001 # Lower learning rate for stability
num_episodes = 1000 # Pong requires more episodes
# Action mapping: policy outputs 0 or 1, map to RIGHT(2) or LEFT(3)
action_map = [2, 3] # Index 0->RIGHT(2), Index 1->LEFT(3)
# Initialize policy and optimizer
policy = PongPolicy(action_dim=2).to(device)
optimizer = optim.Adam(policy.parameters(), lr=learning_rate)
print(f"Using learning rate: {learning_rate} (reduced for stability)")
print(f"Action mapping: 0->RIGHT(2), 1->LEFT(3)")
print(f"Gradient clipping: max_norm=1.0")
print(f"Weight initialization: Kaiming (Conv) + Xavier (FC)\n")
# Train with periodic checkpointing
print("Starting training (checkpoints saved every 500 episodes)...\n")
episode_rewards = train_policy_gradient(
'ALE/Pong-v5', policy, optimizer, gamma, num_episodes,
is_pong=True, action_map=action_map
)
print("\nTraining completed!")
# Evaluate
print("\nEvaluating trained policy...")
eval_rewards = evaluate_policy(
'ALE/Pong-v5', policy, num_episodes=500,
is_pong=True, action_map=action_map
)
# Plot results
plot_results(episode_rewards, eval_rewards, 'Pong-v5', 'pong')
# Save model
torch.save(policy.state_dict(), 'pong_policy.pth')
print("\nModel saved as 'pong_policy.pth'")
return policy, episode_rewards, eval_rewards
# ==================== Run Training ====================
if __name__ == "__main__":
# Train CartPole
#cartpole_policy, cartpole_train_rewards, cartpole_eval_rewards = train_cartpole()
# Train Pong (this will take longer)
#print("\n\nNote: Pong training will take significantly longer (may take hours)")
#print("You may want to reduce num_episodes if just testing the code.\n")
# Uncomment the line below to train Pong
pong_policy, pong_train_rewards, pong_eval_rewards = train_pong() |