File size: 24,281 Bytes
6292460 769f59f 6292460 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 | import torch
import torch.nn as nn
import torch.nn.functional as F
from model import ChessTransformer
from data_loader import VOCAB
import os
import random
import chess
import chess.engine
import gradio as gr
import threading
import time
import collections
import numpy as np
# Global variables for UI monitoring
current_game_pgn = ""
current_eval = 0.0
current_sf_eval = 0.0
champion_wins = 0
challenger_wins = 0
draws = 0
training_stats = {"loss": 0.0, "reward": 0.0, "epoch": 0}
current_challenger_is_white = True
last_promoted_step = 0
# Thread-safe queues and locks
class ReplayBuffer:
def __init__(self, capacity=50000):
self.buffer = []
self.capacity = capacity
self.ptr = 0
self.lock = threading.Lock()
def append(self, item):
with self.lock:
if len(self.buffer) < self.capacity:
self.buffer.append(item)
else:
self.buffer[self.ptr] = item
self.ptr = (self.ptr + 1) % self.capacity
def __len__(self):
with self.lock:
return len(self.buffer)
def sample(self, batch_size):
with self.lock:
return random.sample(self.buffer, batch_size)
replay_buffer = ReplayBuffer(50000)
recent_outcomes = collections.deque(maxlen=100)
ui_lock = threading.Lock()
stats_lock = threading.Lock()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")
INV_VOCAB = {v: k for k, v in VOCAB.items()}
# Global Models
champion = None
challenger = None
optimizer = None
def get_stockfish_engine():
stockfish_path = os.path.join("stockfish", "stockfish-windows-x86-64-avx2.exe")
if os.path.exists(stockfish_path):
sf_engine = chess.engine.SimpleEngine.popen_uci(stockfish_path)
sf_engine.configure({"Hash": 64, "Threads": 1})
return sf_engine
return None
def encode_history(history, max_length=120):
seq = []
for tok in history:
seq.append(VOCAB.get(tok, VOCAB.get("<unk>", 0)))
if len(seq) > max_length:
seq = seq[-max_length:]
else:
seq = seq + [0] * (max_length - len(seq))
return torch.tensor(seq, dtype=torch.long, device=device).unsqueeze(0)
def sample_move(policy_logits, board, temperature=1.0):
logits = policy_logits / temperature
legal_moves = list(board.legal_moves)
legal_ucis = [m.uci() for m in legal_moves]
mask = torch.full_like(logits, float('-inf'))
for idx, token in INV_VOCAB.items():
if token in legal_ucis:
mask[0, idx] = 0.0
masked_logits = logits + mask
probs = F.softmax(masked_logits, dim=-1)
if torch.isnan(probs).any() or probs.sum() == 0:
return random.choice(legal_moves).uci()
m = torch.multinomial(probs[0], 1).item()
action = INV_VOCAB.get(m)
if action not in legal_ucis:
return random.choice(legal_moves).uci()
return action
def actor_worker(worker_id):
"""Background thread playing BATCH_SIZE games concurrently."""
global current_game_pgn, current_eval, current_sf_eval, current_challenger_is_white
global champion_wins, challenger_wins, draws
BATCH_SIZE = 16
sf_engine = get_stockfish_engine()
sf_limit = chess.engine.Limit(time=0.05)
def evaluate_position(board):
if sf_engine is None: return 0.0
info = sf_engine.analyse(board, sf_limit)
score = info["score"].white()
if score.is_mate(): return 10000 if score.mate() > 0 else -10000
return score.score()
print(f"Actor {worker_id} started with Batch Size {BATCH_SIZE}.")
while True:
boards = [chess.Board() for _ in range(BATCH_SIZE)]
histories = [["<bos>"] for _ in range(BATCH_SIZE)]
active = [True for _ in range(BATCH_SIZE)]
evals_current = []
for b in boards:
evals_current.append(evaluate_position(b))
import random
challenger_is_white = [random.choice([True, False]) for _ in range(BATCH_SIZE)]
if worker_id == 0:
with ui_lock:
current_challenger_is_white = challenger_is_white[0]
game_data = [{"states": [], "actions": [], "advantages": [], "sf_values": []} for _ in range(BATCH_SIZE)]
turn_count = 0
while any(active) and turn_count < 200:
turn_count += 1
challenger_indices = []
champion_indices = []
for i in range(BATCH_SIZE):
if not active[i]:
continue
turn_white = boards[i].turn
if turn_white == challenger_is_white[i]:
challenger_indices.append(i)
else:
champion_indices.append(i)
p_chal, v_chal = None, None
p_champ, v_champ = None, None
with torch.no_grad():
with torch.autocast(device_type='cuda', dtype=torch.float16):
if len(challenger_indices) > 0:
x_chal = torch.cat([encode_history(histories[i]) for i in challenger_indices], dim=0)
p_chal, v_chal = challenger(x_chal)
if len(champion_indices) > 0:
x_champ = torch.cat([encode_history(histories[i]) for i in champion_indices], dim=0)
p_champ, v_champ = champion(x_champ)
# Process Challenger moves
for idx, i in enumerate(challenger_indices):
p_logits = p_chal[idx, -1, :]
v = v_chal[idx, -1].item()
action = sample_move(p_logits.unsqueeze(0), boards[i])
if worker_id == 0 and i == 0:
with ui_lock:
current_game_pgn = boards[0].fen()
turn_white_ui = boards[0].turn
current_eval = v if turn_white_ui else -v
current_sf_eval = max(-1.0, min(1.0, evals_current[0] / 500.0))
import time
time.sleep(0.05)
game_data[i]["states"].append(encode_history(histories[i]).cpu())
game_data[i]["actions"].append(VOCAB.get(action, 0))
boards[i].push_uci(action)
eval_next_cp = evaluate_position(boards[i])
turn_white_val = not boards[i].turn
adv = (eval_next_cp - evals_current[i]) if turn_white_val else (evals_current[i] - eval_next_cp)
game_data[i]["advantages"].append(adv)
sf_val = max(-1.0, min(1.0, evals_current[i] / 500.0))
if not turn_white_val: sf_val = -sf_val
game_data[i]["sf_values"].append(sf_val)
evals_current[i] = eval_next_cp
histories[i].append(action)
if abs(eval_next_cp) > 800 or boards[i].is_game_over():
active[i] = False
# Process Champion moves
for idx, i in enumerate(champion_indices):
p_logits = p_champ[idx, -1, :]
action = sample_move(p_logits.unsqueeze(0), boards[i])
boards[i].push_uci(action)
eval_next_cp = evaluate_position(boards[i])
evals_current[i] = eval_next_cp
histories[i].append(action)
if abs(eval_next_cp) > 800 or boards[i].is_game_over():
active[i] = False
# Calculate rewards and push to replay buffer
for i in range(BATCH_SIZE):
outcome = boards[i].outcome()
if outcome is None:
if evals_current[i] >= 500: reward = 1.0 if challenger_is_white[i] else -1.0
elif evals_current[i] <= -500: reward = -1.0 if challenger_is_white[i] else 1.0
else: reward = 0.0
elif outcome.winner is None:
reward = 0.0
else:
reward = 1.0 if outcome.winner == challenger_is_white[i] else -1.0
with stats_lock:
if reward > 0:
challenger_wins += 1
recent_outcomes.append(1)
elif reward < 0:
champion_wins += 1
recent_outcomes.append(-1)
else:
draws += 1
recent_outcomes.append(0)
states = game_data[i]["states"]
if len(states) > 0:
adv_tensor = torch.tensor(game_data[i]["advantages"], dtype=torch.float32)
if adv_tensor.std() > 0:
adv_tensor = (adv_tensor - adv_tensor.mean()) / (adv_tensor.std() + 1e-8)
else:
adv_tensor = adv_tensor - adv_tensor.mean()
for j in range(len(states)):
replay_buffer.append((states[j], game_data[i]["actions"][j], adv_tensor[j].item(), game_data[i]["sf_values"][j]))
def learner_worker():
"""Background thread that continuously samples the Replay Buffer and updates the Neural Network."""
global training_stats
print("Learner started.")
batch_size = 128
scaler = torch.amp.GradScaler('cuda')
import time
while True:
if len(replay_buffer) < batch_size:
time.sleep(1)
continue
batch = replay_buffer.sample(batch_size)
# Move batch to GPU
s_batch = torch.cat([b[0] for b in batch]).to(device)
a_batch = torch.tensor([b[1] for b in batch], dtype=torch.long, device=device)
adv_batch = torch.tensor([b[2] for b in batch], dtype=torch.float32, device=device)
sf_val_batch = torch.tensor([b[3] for b in batch], dtype=torch.float32, device=device)
optimizer.zero_grad()
with torch.autocast(device_type='cuda', dtype=torch.float16):
p, v_pred = challenger(s_batch)
p_logits = p[:, -1, :]
v_pred = v_pred[:, -1].squeeze(-1)
log_prob = F.log_softmax(p_logits, dim=-1)
action_log_probs = log_prob[torch.arange(batch_size), a_batch]
# CRITICAL FIX: Only train on positive advantages (Advantage-Weighted Behavioral Cloning).
# If we allow negative advantages, the optimizer pushes log_prob to negative infinity,
# causing the policy_loss to explode into massive negative numbers (e.g. -59.7) and
# completely destroying the Neural Network's weights (including the Value Head).
positive_adv = torch.clamp(adv_batch, min=0.0)
policy_loss = -(action_log_probs * positive_adv).mean()
value_loss = F.mse_loss(v_pred, sf_val_batch)
loss = policy_loss + 0.5 * value_loss
scaler.scale(loss).backward()
scaler.unscale_(optimizer)
torch.nn.utils.clip_grad_norm_(challenger.parameters(), 1.0)
scaler.step(optimizer)
scaler.update()
with stats_lock:
training_stats["epoch"] += 1
training_stats["loss"] = loss.item()
if len(recent_outcomes) > 0:
win_rate = sum(1 for x in recent_outcomes if x == 1) / len(recent_outcomes)
training_stats["reward"] = win_rate
# If Challenger is consistently crushing the Champion, promote it!
if len(recent_outcomes) >= 50 and win_rate >= 0.55:
print(f"\\n>>> PROMOTING CHALLENGER! Win rate: {win_rate:.2f} <<<\\n")
global last_promoted_step
last_promoted_step = training_stats["epoch"]
champion.load_state_dict(challenger.state_dict())
recent_outcomes.clear()
torch.save({
"epoch": training_stats["epoch"],
"model_state_dict": challenger.state_dict()
}, "rl_weights/champion_latest.pth")
if training_stats["epoch"] % 100 == 0:
print(f"Step {training_stats['epoch']} | Loss: {loss.item():.4f} | Win Rate: {training_stats['reward']:.2f} | Buffer: {len(replay_buffer)}")
torch.save({
"epoch": training_stats["epoch"],
"model_state_dict": challenger.state_dict()
}, "rl_weights/challenger_latest.pth")
def init_models():
global champion, challenger, optimizer
os.makedirs("rl_weights", exist_ok=True)
champion = ChessTransformer(vocab_size=len(VOCAB), d_model=512, nhead=8, num_layers=6, max_length=120).to(device)
challenger = ChessTransformer(vocab_size=len(VOCAB), d_model=512, nhead=8, num_layers=6, max_length=120).to(device)
latest_rl_weights = "rl_weights/champion_latest.pth"
fast_weights = "weights/chess_fast_best.pth"
start_epoch = 0
if os.path.exists(latest_rl_weights):
print(f"Loading latest RL champion from {latest_rl_weights}")
ckpt = torch.load(latest_rl_weights, map_location=device)
if "epoch" in ckpt and isinstance(ckpt, dict):
start_epoch = ckpt["epoch"]
training_stats["epoch"] = start_epoch
global last_promoted_step
last_promoted_step = start_epoch
if "model_state_dict" in ckpt:
ckpt = ckpt["model_state_dict"]
champion.load_state_dict(ckpt)
challenger.load_state_dict(ckpt)
elif os.path.exists(fast_weights):
print(f"Loading base Fast weights from {fast_weights}")
ckpt = torch.load(fast_weights, map_location=device)
if "epoch" in ckpt and isinstance(ckpt, dict):
start_epoch = ckpt["epoch"]
training_stats["epoch"] = start_epoch
if "model_state_dict" in ckpt:
ckpt = ckpt["model_state_dict"]
champion.load_state_dict(ckpt)
challenger.load_state_dict(ckpt)
champion.eval()
challenger.train()
optimizer = torch.optim.AdamW(challenger.parameters(), lr=1e-5)
def build_ui():
def get_state():
try:
import json
with ui_lock:
b = chess.Board(current_game_pgn)
eval_val = current_eval
sf_eval_val = current_sf_eval
white_name = "Challenger" if current_challenger_is_white else "Champion"
black_name = "Champion" if current_challenger_is_white else "Challenger"
return json.dumps({
"fen": b.fen(),
"eval": eval_val,
"sf_eval": sf_eval_val,
"white_name": white_name,
"black_name": black_name
})
except:
import json
return json.dumps({"fen": chess.STARTING_FEN, "eval": 0.0, "sf_eval": 0.0, "white_name": "", "black_name": ""})
def get_stats():
with stats_lock:
return [
["Training Steps", str(training_stats['epoch'])],
["Last Promoted Step", str(last_promoted_step)],
["Loss", f"{training_stats['loss']:.4f}"],
["Recent Win Rate", f"{training_stats['reward']:.2f}"],
["Replay Buffer Size", str(len(replay_buffer))],
["Challenger Wins", str(challenger_wins)],
["Champion Wins", str(champion_wins)],
["Draws", str(draws)]
]
custom_css = """
@import url('https://fonts.googleapis.com/css2?family=Outfit:wght@400;700&display=swap');
body, .gradio-container {
font-family: 'Outfit', sans-serif !important;
background: linear-gradient(135deg, #0f2027, #203a43, #2c5364) !important;
background-attachment: fixed !important;
color: white !important;
}
.gradio-container { border: none !important; }
.glass-panel {
background: rgba(255, 255, 255, 0.1) !important;
backdrop-filter: blur(10px) !important;
border-radius: 12px !important;
border: 1px solid rgba(255, 255, 255, 0.18) !important;
padding: 20px !important;
box-shadow: 0 8px 32px 0 rgba(31, 38, 135, 0.37) !important;
}
.eval-bar-container {
width: 30px;
height: 400px;
background-color: #333;
border-radius: 4px;
border: 4px solid #fff;
position: relative;
overflow: hidden;
display: flex;
flex-direction: column-reverse;
box-shadow: 0 15px 35px rgba(0,0,0,0.5);
}
.eval-bar-fill {
width: 100%;
height: 50%;
background-color: #fff;
transition: height 0.5s cubic-bezier(0.4, 0, 0.2, 1);
}
.eval-marker {
position: absolute;
top: 50%;
left: 0;
width: 100%;
height: 2px;
background-color: #ff5e7e;
z-index: 10;
}
"""
with gr.Blocks(title="Neurex RL Dashboard", css=custom_css) as demo:
gr.HTML("<h1 style='text-align: center; color: white; font-weight: 700; font-size: 2.5rem; text-shadow: 2px 2px 10px rgba(0,0,0,0.5);'>🧠 Neurex RL Self-Play Dashboard</h1>")
gr.HTML("<p style='text-align: center; color: #ddd; font-size: 1.1rem;'><b>ASYNCHRONOUS ALPHA-ZERO MODE</b> | Real-time Actor-Learner Architecture</p>")
with gr.Row():
with gr.Column(elem_classes=["glass-panel"]):
board_html = """
<div style="display: flex; flex-direction: column; align-items: center;">
<div id="blackName" style="font-size: 1.3rem; font-weight: bold; margin-bottom: 12px; color: #ff5e7e; text-shadow: 1px 1px 5px rgba(0,0,0,0.5);">Black</div>
<div style="display: flex; align-items: center; gap: 15px; justify-content: center;">
<div class="eval-bar-container" title="Neural Network Evaluation">
<div class="eval-bar-fill" id="evalBar"></div>
<div class="eval-marker"></div>
</div>
<div id="board" style="width: 400px; box-shadow: 0 15px 35px rgba(0,0,0,0.5); border: 4px solid #fff; border-radius: 4px; overflow: hidden;"></div>
<div class="eval-bar-container" title="Stockfish Evaluation">
<div class="eval-bar-fill" id="sfEvalBar" style="background-color: #00ff88;"></div>
<div class="eval-marker"></div>
</div>
</div>
<div id="whiteName" style="font-size: 1.3rem; font-weight: bold; margin-top: 12px; color: #00ff88; text-shadow: 1px 1px 5px rgba(0,0,0,0.5);">White</div>
</div>
"""
board_view = gr.HTML(board_html)
current_state_box = gr.Textbox(visible=False)
with gr.Column(elem_classes=["glass-panel"]):
stats_view = gr.Dataframe(headers=["Metric", "Value"], interactive=False)
timer = gr.Timer(0.5)
timer.tick(get_state, inputs=[], outputs=[current_state_box])
timer.tick(get_stats, inputs=[], outputs=[stats_view])
js_callback = """
(state_str) => {
try {
let state = JSON.parse(state_str);
if (window.my_board) window.my_board.position(state.fen);
let evalBar = document.getElementById('evalBar');
if (evalBar) {
let heightPercent = ((state.eval + 1.0) / 2.0) * 100;
heightPercent = Math.max(0, Math.min(100, heightPercent));
evalBar.style.height = heightPercent + '%';
}
let sfEvalBar = document.getElementById('sfEvalBar');
if (sfEvalBar) {
let sfHeightPercent = ((state.sf_eval + 1.0) / 2.0) * 100;
sfHeightPercent = Math.max(0, Math.min(100, sfHeightPercent));
sfEvalBar.style.height = sfHeightPercent + '%';
}
let blackName = document.getElementById('blackName');
if (blackName && state.black_name) {
let dot = state.black_name === "Challenger" ? "🟢" : "🔴";
let color = state.black_name === "Challenger" ? "#00ff88" : "#ff5e7e";
blackName.innerText = dot + " " + state.black_name + " (Black)";
blackName.style.color = color;
}
let whiteName = document.getElementById('whiteName');
if (whiteName && state.white_name) {
let dot = state.white_name === "Challenger" ? "🟢" : "🔴";
let color = state.white_name === "Challenger" ? "#00ff88" : "#ff5e7e";
whiteName.innerText = dot + " " + state.white_name + " (White)";
whiteName.style.color = color;
}
} catch(e) {}
return state_str;
}
"""
current_state_box.change(None, inputs=[current_state_box], js=js_callback)
init_js = """
function() {
var jq = document.createElement('script');
jq.src = "https://code.jquery.com/jquery-3.5.1.min.js";
document.head.appendChild(jq);
var css = document.createElement('link');
css.rel = "stylesheet";
css.href = "https://unpkg.com/@chrisoakman/chessboardjs@1.0.0/dist/chessboard-1.0.0.min.css";
document.head.appendChild(css);
jq.onload = function() {
var cb = document.createElement('script');
cb.src = "https://unpkg.com/@chrisoakman/chessboardjs@1.0.0/dist/chessboard-1.0.0.min.js";
document.head.appendChild(cb);
cb.onload = function() {
let checkExist = setInterval(function() {
if (document.getElementById('board')) {
window.my_board = Chessboard('board', {
position: 'start',
pieceTheme: 'https://chessboardjs.com/img/chesspieces/wikipedia/{piece}.png'
});
clearInterval(checkExist);
}
}, 100);
};
};
}
"""
demo.load(None, None, None, js=init_js)
return demo
if __name__ == "__main__":
init_models()
# Spawn 3 Actor Threads to play games using 3 Stockfish instances
for i in range(4):
t = threading.Thread(target=actor_worker, args=(i,), daemon=True)
t.start()
# Spawn 1 Learner Thread to aggressively train the GPU
t_learner = threading.Thread(target=learner_worker, daemon=True)
t_learner.start()
# Launch Gradio UI in main thread
demo = build_ui()
demo.launch(server_name="0.0.0.0", prevent_thread_lock=False)
|