Spaces:
Running
Running
File size: 12,021 Bytes
40ee6b4 |
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 |
"""
Policy-Value Network using ResNet Architecture.
Implements the dual-head neural network used in AlphaZero:
- Policy Head: Outputs action probabilities
- Value Head: Outputs state value estimation
Based on:
- "Mastering Chess and Shogi by Self-Play with a General RL Algorithm" (AlphaZero)
- Deep Residual Learning for Image Recognition (ResNet)
"""
from __future__ import annotations
import torch
import torch.nn as nn
import torch.nn.functional as F
from ..training.system_config import NeuralNetworkConfig
class ResidualBlock(nn.Module):
"""
Residual block with batch normalization and skip connections.
Architecture:
Conv -> BN -> ReLU -> Conv -> BN -> Add -> ReLU
"""
def __init__(self, channels: int, use_batch_norm: bool = True):
super().__init__()
self.conv1 = nn.Conv2d(channels, channels, kernel_size=3, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(channels) if use_batch_norm else nn.Identity()
self.conv2 = nn.Conv2d(channels, channels, kernel_size=3, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(channels) if use_batch_norm else nn.Identity()
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""Apply residual block transformation."""
residual = x
out = self.conv1(x)
out = self.bn1(out)
out = F.relu(out)
out = self.conv2(out)
out = self.bn2(out)
# Skip connection
out = out + residual
out = F.relu(out)
return out
class PolicyHead(nn.Module):
"""
Policy head for outputting action probabilities.
Architecture:
Conv -> BN -> ReLU -> FC -> LogSoftmax
"""
def __init__(
self,
input_channels: int,
policy_conv_channels: int,
action_size: int,
board_size: int = 19,
):
super().__init__()
self.conv = nn.Conv2d(input_channels, policy_conv_channels, kernel_size=1, bias=False)
self.bn = nn.BatchNorm2d(policy_conv_channels)
# Assuming square board
fc_input_size = policy_conv_channels * board_size * board_size
self.fc = nn.Linear(fc_input_size, action_size)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Compute policy (action probabilities).
Args:
x: [batch, channels, height, width]
Returns:
Log probabilities: [batch, action_size]
"""
batch_size = x.size(0)
out = self.conv(x)
out = self.bn(out)
out = F.relu(out)
# Flatten spatial dimensions
out = out.view(batch_size, -1)
# Fully connected layer
out = self.fc(out)
# Log probabilities for numerical stability
return F.log_softmax(out, dim=1)
class ValueHead(nn.Module):
"""
Value head for estimating state value.
Architecture:
Conv -> BN -> ReLU -> FC -> ReLU -> FC -> Tanh
"""
def __init__(
self,
input_channels: int,
value_conv_channels: int,
value_fc_hidden: int,
board_size: int = 19,
):
super().__init__()
self.conv = nn.Conv2d(input_channels, value_conv_channels, kernel_size=1, bias=False)
self.bn = nn.BatchNorm2d(value_conv_channels)
# Assuming square board
fc_input_size = value_conv_channels * board_size * board_size
self.fc1 = nn.Linear(fc_input_size, value_fc_hidden)
self.fc2 = nn.Linear(value_fc_hidden, 1)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Compute value estimation.
Args:
x: [batch, channels, height, width]
Returns:
Value: [batch, 1] in range [-1, 1]
"""
batch_size = x.size(0)
out = self.conv(x)
out = self.bn(out)
out = F.relu(out)
# Flatten spatial dimensions
out = out.view(batch_size, -1)
# Fully connected layers
out = self.fc1(out)
out = F.relu(out)
out = self.fc2(out)
# Tanh to bound value in [-1, 1]
return torch.tanh(out)
class PolicyValueNetwork(nn.Module):
"""
Combined policy-value network with ResNet backbone.
This is the core neural network used in AlphaZero-style learning.
"""
def __init__(self, config: NeuralNetworkConfig, board_size: int = 19):
super().__init__()
self.config = config
self.board_size = board_size
# Initial convolution
self.conv_input = nn.Conv2d(
config.input_channels,
config.num_channels,
kernel_size=3,
padding=1,
bias=False,
)
self.bn_input = nn.BatchNorm2d(config.num_channels) if config.use_batch_norm else nn.Identity()
# Residual blocks (shared feature extractor)
self.res_blocks = nn.ModuleList(
[ResidualBlock(config.num_channels, config.use_batch_norm) for _ in range(config.num_res_blocks)]
)
# Policy head
self.policy_head = PolicyHead(
input_channels=config.num_channels,
policy_conv_channels=config.policy_conv_channels,
action_size=config.action_size,
board_size=board_size,
)
# Value head
self.value_head = ValueHead(
input_channels=config.num_channels,
value_conv_channels=config.value_conv_channels,
value_fc_hidden=config.value_fc_hidden,
board_size=board_size,
)
def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
"""
Forward pass through the network.
Args:
x: Input state [batch, channels, height, width]
Returns:
(policy_logits, value) tuple
- policy_logits: [batch, action_size] log probabilities
- value: [batch, 1] state value in [-1, 1]
"""
# Initial convolution
out = self.conv_input(x)
out = self.bn_input(out)
out = F.relu(out)
# Residual blocks
for res_block in self.res_blocks:
out = res_block(out)
# Split into policy and value heads
policy = self.policy_head(out)
value = self.value_head(out)
return policy, value
def predict(self, state: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
"""
Inference mode prediction.
Args:
state: Input state tensor
Returns:
(policy_probs, value) tuple with probabilities (not log)
"""
with torch.no_grad():
policy_log_probs, value = self.forward(state)
policy_probs = torch.exp(policy_log_probs)
return policy_probs, value
def get_parameter_count(self) -> int:
"""Return total number of trainable parameters."""
return sum(p.numel() for p in self.parameters() if p.requires_grad)
class AlphaZeroLoss(nn.Module):
"""
Combined loss function for AlphaZero training.
Loss = (z - v)^2 - π^T log(p) + c||θ||^2
Where:
- z: actual game outcome
- v: value prediction
- π: MCTS visit count distribution
- p: policy prediction
- c: L2 regularization coefficient
"""
def __init__(self, value_loss_weight: float = 1.0):
super().__init__()
self.value_loss_weight = value_loss_weight
def forward(
self,
policy_logits: torch.Tensor,
value: torch.Tensor,
target_policy: torch.Tensor,
target_value: torch.Tensor,
) -> tuple[torch.Tensor, dict]:
"""
Compute AlphaZero loss.
Args:
policy_logits: Predicted policy log probabilities [batch, action_size]
value: Predicted values [batch, 1]
target_policy: Target policy from MCTS [batch, action_size]
target_value: Target value from game outcome [batch, 1]
Returns:
(total_loss, loss_dict) tuple
"""
# Value loss: MSE between predicted and actual outcome
value_loss = F.mse_loss(value.squeeze(-1), target_value)
# Policy loss: Cross-entropy between MCTS policy and network policy
# Target policy is already normalized, policy_logits are log probabilities
policy_loss = -torch.sum(target_policy * policy_logits, dim=1).mean()
# Combined loss
total_loss = self.value_loss_weight * value_loss + policy_loss
loss_dict = {
"total": total_loss.item(),
"value": value_loss.item(),
"policy": policy_loss.item(),
}
return total_loss, loss_dict
def create_policy_value_network(
config: NeuralNetworkConfig,
board_size: int = 19,
device: str = "cpu",
) -> PolicyValueNetwork:
"""
Factory function to create and initialize policy-value network.
Args:
config: Network configuration
board_size: Board/grid size (for games)
device: Device to place model on
Returns:
Initialized PolicyValueNetwork
"""
network = PolicyValueNetwork(config, board_size)
# He initialization for convolutional layers
def init_weights(m):
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
if m.bias is not None:
nn.init.constant_(m.bias, 0)
network.apply(init_weights)
network = network.to(device)
return network
# Example: Simpler MLP-based policy-value network for non-spatial tasks
class MLPPolicyValueNetwork(nn.Module):
"""
MLP-based policy-value network for non-spatial state representations.
Useful for tasks where state is not naturally represented as an image.
"""
def __init__(
self,
state_dim: int,
action_size: int,
hidden_dims: list[int] | None = None,
use_batch_norm: bool = True,
dropout: float = 0.1,
):
super().__init__()
self.state_dim = state_dim
self.action_size = action_size
if hidden_dims is None:
hidden_dims = [512, 256]
# Shared feature extractor
layers = []
prev_dim = state_dim
for hidden_dim in hidden_dims:
layers.append(nn.Linear(prev_dim, hidden_dim))
if use_batch_norm:
layers.append(nn.BatchNorm1d(hidden_dim))
layers.append(nn.ReLU())
if dropout > 0:
layers.append(nn.Dropout(dropout))
prev_dim = hidden_dim
self.shared_network = nn.Sequential(*layers)
# Policy head
self.policy_head = nn.Sequential(
nn.Linear(prev_dim, prev_dim // 2),
nn.ReLU(),
nn.Linear(prev_dim // 2, action_size),
)
# Value head
self.value_head = nn.Sequential(
nn.Linear(prev_dim, prev_dim // 2),
nn.ReLU(),
nn.Linear(prev_dim // 2, 1),
nn.Tanh(),
)
def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
"""
Forward pass.
Args:
x: Input state [batch, state_dim]
Returns:
(policy_log_probs, value) tuple
"""
# Shared features
features = self.shared_network(x)
# Policy
policy_logits = self.policy_head(features)
policy_log_probs = F.log_softmax(policy_logits, dim=1)
# Value
value = self.value_head(features)
return policy_log_probs, value
def get_parameter_count(self) -> int:
"""Return total number of trainable parameters."""
return sum(p.numel() for p in self.parameters() if p.requires_grad)
|