Spaces:
Sleeping
Sleeping
File size: 9,870 Bytes
e8a0fd8 |
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 |
import torch
from collections import deque
import numpy as np
from typing import List, Iterator, Tuple, Optional
import chess
class Game:
"""
Represents a single chess game trajectory with all relevant data for RL training.
Acts as a *temporary* buffer inside loop
Handles:
- Storing trajectory data (fens, reps, actions, log_probs, values, invalid_masks)
- Tracking game status (active/complete)
"""
def __init__(self):
self.active = True
self.valid = True
self.completion_reason = None
self.game_result = None
self.fens = []
self.repetition_counts = []
self.actions = []
self.values = []
self.log_probs = []
self.invalid_masks = []
def update_trajectory(self, fen, rep, act, val, logp, inv_m):
self.fens.append(fen)
self.repetition_counts.append(rep)
self.actions.append(act)
self.values.append(val)
self.log_probs.append(logp)
self.invalid_masks.append(inv_m)
def update_game_status(self, done, reason):
if done == True:
self.active = False
if reason in ["1-0","0-1","1/2-1/2"]:
self.completion_reason = reason
self.game_result = reason
else:
self.completion_reason = reason
self.game_result = None
self.valid = False
def get_white_trajectory(self):
"""Extract the trajectory for white"""
indices = []
for i in range(len(self.fens) - 1):
board = chess.Board(self.fens[i])
if board.turn: # True if white to move
indices.append(i)
return {
'fens': [self.fens[i] for i in indices],
'repetition_counts': [self.repetition_counts[i] for i in indices],
'actions': [self.actions[i] for i in indices],
'values': [self.values[i] for i in indices],
'log_probs': [self.log_probs[i] for i in indices],
'invalid_masks': [self.invalid_masks[i] for i in indices]
}
def get_black_trajectory(self):
"""Extract the trajectory for black pieces."""
indices = []
for i in range(len(self.fens) - 1):
board = chess.Board(self.fens[i])
if not board.turn: # False if black to move
indices.append(i)
return {
'fens': [self.fens[i] for i in indices],
'repetition_counts': [self.repetition_counts[i] for i in indices],
'actions': [self.actions[i] for i in indices],
'values': [self.values[i] for i in indices],
'log_probs': [self.log_probs[i] for i in indices],
'invalid_masks': [self.invalid_masks[i] for i in indices]
}
class ReplayBuffer:
"""
The buffer class for PPO reinforcement learning.
Handles:
- store samples including:
1. fens
2. reps
3. actions
4. log_probs
5. values
6. invalid_masks
- calculate advantage based on reward and value (7. advantage)
- output samples in batches
Since PPO is largely on-policy, so the replay buffer will not be so large that deque is not appropriate
"""
def __init__(self,
capacity: int,
batch_size: int,
gamma: float,
gae_lambda: float,
shuffle: bool=True
):
self.gamma = gamma
self.gae_lambda = gae_lambda
self.fens = deque(maxlen=capacity)
self.repetition_counts = deque(maxlen=capacity)
self.actions = deque(maxlen=capacity)
self.log_probs = deque(maxlen=capacity)
self.values = deque(maxlen=capacity)
self.invalid_masks = deque(maxlen=capacity)
self.advantages = deque(maxlen=capacity)
self.batch_size = batch_size
self.shuffle = shuffle
def push_game(self, game: Game):
"""
Process a completed game and add its trajectories to the buffer.
Handles reward computation for both white and black players.
"""
if not game.valid:
return
result = game.game_result
if result not in ["1-0","0-1","1/2-1/2"]:
raise ValueError(f"Result not recognized: {result}. Either an incompleted game was passed in, or game.update_game_status() method is wrong.")
if result == "1-0": result = 1
elif result == "0-1": result = -1
elif result == "1/2-1/2": result = 0
white_traj = game.get_white_trajectory()
if white_traj["fens"]:
self._process_trajectory(
white_traj["fens"],
white_traj["repetition_counts"],
white_traj["actions"],
white_traj["log_probs"],
white_traj["values"],
white_traj["invalid_masks"],
result
)
black_traj = game.get_black_trajectory()
if black_traj["fens"]:
self._process_trajectory(
black_traj["fens"],
black_traj["repetition_counts"],
black_traj["actions"],
black_traj["log_probs"],
black_traj["values"],
black_traj["invalid_masks"],
-result # flip reward for black's perspective
)
def _process_trajectory(self, fens, reps, actions, log_probs, values, invalid_masks, final_reward):
"""Process a trajectory for one player, compute advantages and add to buffer"""
values_tensor = torch.tensor(values) if not torch.is_tensor(values) else values
advantages = self._compute_advantage(values_tensor, final_reward)
for i in range(len(fens)):
self.fens.append(fens[i])
self.repetition_counts.append(reps[i])
self.actions.append(actions[i])
self.log_probs.append(log_probs[i])
self.values.append(values[i])
self.invalid_masks.append(invalid_masks[i])
self.advantages.append(advantages[i])
def _compute_advantage(self, value_traj: torch.Tensor, final_reward: float) -> torch.Tensor:
"""
Calculate GAE with only a terminal reward: r_t = 0 for t < T-1 and r_{T-1} = final_reward
Args:
value_traj: value prediction of the model
final_reward: game result
Returns:
advantage, torch.Tensor of shape same with value_traj
"""
vals = value_traj.detach().cpu().float()
T = vals.shape[0] if vals.dim() > 0 else 1
adv = torch.zeros(T)
next_value = 0.0
gae = 0.0
for t in reversed(range(T)):
reward = final_reward if t == T-1 else 0.0
delta = reward + self.gamma * next_value - vals[t]
gae = delta + self.gamma * self.gae_lambda * gae
adv[t] = gae
next_value = vals[t]
return adv
def sample(self) -> Iterator[Tuple[List[str], # fen
torch.Tensor,# rep
torch.Tensor,# act
torch.Tensor,# logp
torch.Tensor,# val
torch.Tensor,# inv_m
torch.Tensor]]: # adv
"""Yield minibatches of size batch_size for training"""
n = len(self.fens)
if n < self.batch_size:
return
idxs = np.arange(n)
if self.shuffle:
np.random.shuffle(idxs)
for start in range(0, n, self.batch_size):
batch_idx = idxs[start:start+self.batch_size]
if len(batch_idx) < self.batch_size:
break
fens_b = [self.fens[i] for i in batch_idx]
reps_b = torch.stack([
self.repetition_counts[i].detach().clone() if torch.is_tensor(self.repetition_counts[i])
else torch.tensor(self.repetition_counts[i])
for i in batch_idx
])
acts_b = torch.stack([
self.actions[i].detach().clone() if torch.is_tensor(self.actions[i])
else torch.tensor(self.actions[i])
for i in batch_idx
])
logps_b = torch.stack([
self.log_probs[i].detach().clone() if torch.is_tensor(self.log_probs[i])
else torch.tensor(self.log_probs[i])
for i in batch_idx
])
vals_b = torch.stack([
self.values[i].detach().clone() if torch.is_tensor(self.values[i])
else torch.tensor(self.values[i])
for i in batch_idx
])
advs_b = torch.stack([
self.advantages[i].detach().clone() if torch.is_tensor(self.advantages[i])
else torch.tensor(self.advantages[i])
for i in batch_idx
])
invs_b = torch.stack([
self.invalid_masks[i] if torch.is_tensor(self.invalid_masks[i])
else torch.tensor(self.invalid_masks[i])
for i in batch_idx
])
yield fens_b, reps_b, acts_b, logps_b, vals_b, invs_b, advs_b
def __len__(self) -> int:
return len(self.fens)
def clear(self) -> None:
self.fens.clear()
self.repetition_counts.clear()
self.actions.clear()
self.log_probs.clear()
self.values.clear()
self.invalid_masks.clear()
self.advantages.clear()
|