Create selfchess-colab.py
Browse files- selfchess-colab.py +224 -0
selfchess-colab.py
ADDED
@@ -0,0 +1,224 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
os.system('pip install chess')
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import torch.optim as optim
|
6 |
+
import chess
|
7 |
+
import os
|
8 |
+
import chess.engine as eng
|
9 |
+
import torch.multiprocessing as mp
|
10 |
+
|
11 |
+
# CONFIGURATION
|
12 |
+
CONFIG = {
|
13 |
+
"stockfish_path": "/usr/games/stockfish",
|
14 |
+
"model_path": "NeoChess/chessy_model.pth",
|
15 |
+
"backup_model_path": "NeoChess/chessy_modelt-1.pth",
|
16 |
+
"device": torch.device("cuda"),
|
17 |
+
"learning_rate": 1e-4,
|
18 |
+
"num_games": 30,
|
19 |
+
"num_epochs": 10,
|
20 |
+
"stockfish_time_limit": 1.0,
|
21 |
+
"search_depth": 1,
|
22 |
+
"epsilon": 4
|
23 |
+
}
|
24 |
+
|
25 |
+
device = CONFIG["device"]
|
26 |
+
|
27 |
+
def board_to_tensor(board):
|
28 |
+
piece_encoding = {
|
29 |
+
'P': 1, 'N': 2, 'B': 3, 'R': 4, 'Q': 5, 'K': 6,
|
30 |
+
'p': 7, 'n': 8, 'b': 9, 'r': 10, 'q': 11, 'k': 12
|
31 |
+
}
|
32 |
+
|
33 |
+
tensor = torch.zeros(64, dtype=torch.long)
|
34 |
+
for square in chess.SQUARES:
|
35 |
+
piece = board.piece_at(square)
|
36 |
+
if piece:
|
37 |
+
tensor[square] = piece_encoding[piece.symbol()]
|
38 |
+
else:
|
39 |
+
tensor[square] = 0
|
40 |
+
|
41 |
+
return tensor.unsqueeze(0)
|
42 |
+
|
43 |
+
class NN1(nn.Module):
|
44 |
+
def __init__(self):
|
45 |
+
super().__init__()
|
46 |
+
self.embedding = nn.Embedding(13, 64)
|
47 |
+
self.attention = nn.MultiheadAttention(embed_dim=64, num_heads=16)
|
48 |
+
self.neu = 512
|
49 |
+
self.neurons = nn.Sequential(
|
50 |
+
nn.Linear(4096, self.neu),
|
51 |
+
nn.ReLU(),
|
52 |
+
nn.Linear(self.neu, self.neu),
|
53 |
+
nn.ReLU(),
|
54 |
+
nn.Linear(self.neu, self.neu),
|
55 |
+
nn.ReLU(),
|
56 |
+
nn.Linear(self.neu, self.neu),
|
57 |
+
nn.ReLU(),
|
58 |
+
nn.Linear(self.neu, self.neu),
|
59 |
+
nn.ReLU(),
|
60 |
+
nn.Linear(self.neu, self.neu),
|
61 |
+
nn.ReLU(),
|
62 |
+
nn.Linear(self.neu, self.neu),
|
63 |
+
nn.ReLU(),
|
64 |
+
nn.Linear(self.neu, self.neu),
|
65 |
+
nn.ReLU(),
|
66 |
+
nn.Linear(self.neu, self.neu),
|
67 |
+
nn.ReLU(),
|
68 |
+
nn.Linear(self.neu, self.neu),
|
69 |
+
nn.ReLU(),
|
70 |
+
nn.Linear(self.neu, self.neu),
|
71 |
+
nn.ReLU(),
|
72 |
+
nn.Linear(self.neu, self.neu),
|
73 |
+
nn.ReLU(),
|
74 |
+
nn.Linear(self.neu, self.neu),
|
75 |
+
nn.ReLU(),
|
76 |
+
nn.Linear(self.neu, 64),
|
77 |
+
nn.ReLU(),
|
78 |
+
nn.Linear(64, 4)
|
79 |
+
)
|
80 |
+
|
81 |
+
def forward(self, x):
|
82 |
+
x = self.embedding(x)
|
83 |
+
x = x.permute(1, 0, 2)
|
84 |
+
attn_output, _ = self.attention(x, x, x)
|
85 |
+
x = attn_output.permute(1, 0, 2).contiguous()
|
86 |
+
x = x.view(x.size(0), -1)
|
87 |
+
x = self.neurons(x)
|
88 |
+
return x
|
89 |
+
|
90 |
+
model = NN1().to(device)
|
91 |
+
optimizer = optim.Adam(model.parameters(), lr=CONFIG["learning_rate"])
|
92 |
+
|
93 |
+
try:
|
94 |
+
model.load_state_dict(torch.load(CONFIG["model_path"], map_location=device))
|
95 |
+
print(f"Loaded model from {CONFIG['model_path']}")
|
96 |
+
except FileNotFoundError:
|
97 |
+
try:
|
98 |
+
model.load_state_dict(torch.load(CONFIG["backup_model_path"], map_location=device))
|
99 |
+
print(f"Loaded backup model from {CONFIG['backup_model_path']}")
|
100 |
+
except FileNotFoundError:
|
101 |
+
print("No model file found, starting from scratch.")
|
102 |
+
|
103 |
+
model.train()
|
104 |
+
criterion = nn.MSELoss()
|
105 |
+
engine = eng.SimpleEngine.popen_uci(CONFIG["stockfish_path"])
|
106 |
+
lim = eng.Limit(time=CONFIG["stockfish_time_limit"])
|
107 |
+
|
108 |
+
def get_evaluation(board):
|
109 |
+
"""
|
110 |
+
Returns the evaluation of the board from the perspective of the current player.
|
111 |
+
The model's output is from White's perspective.
|
112 |
+
"""
|
113 |
+
tensor = board_to_tensor(board).to(device)
|
114 |
+
with torch.no_grad():
|
115 |
+
evaluation = model(tensor)[0][0].item()
|
116 |
+
|
117 |
+
if board.turn == chess.WHITE:
|
118 |
+
return evaluation
|
119 |
+
else:
|
120 |
+
return -evaluation
|
121 |
+
|
122 |
+
def search(board, depth, alpha, beta):
|
123 |
+
"""
|
124 |
+
A negamax search function.
|
125 |
+
"""
|
126 |
+
if depth == 0 or board.is_game_over():
|
127 |
+
return get_evaluation(board)
|
128 |
+
|
129 |
+
max_eval = float('-inf')
|
130 |
+
for move in board.legal_moves:
|
131 |
+
board.push(move)
|
132 |
+
eval = -search(board, depth - 1, -beta, -alpha)
|
133 |
+
board.pop()
|
134 |
+
max_eval = max(max_eval, eval)
|
135 |
+
alpha = max(alpha, eval)
|
136 |
+
if alpha >= beta:
|
137 |
+
break
|
138 |
+
return max_eval
|
139 |
+
|
140 |
+
|
141 |
+
|
142 |
+
|
143 |
+
def game_gen(engine_side):
|
144 |
+
data = []
|
145 |
+
mc = 0
|
146 |
+
board = chess.Board()
|
147 |
+
while not board.is_game_over():
|
148 |
+
is_bot_turn = board.turn != engine_side
|
149 |
+
|
150 |
+
if is_bot_turn:
|
151 |
+
evaling = {}
|
152 |
+
for move in board.legal_moves:
|
153 |
+
board.push(move)
|
154 |
+
evaling[move] = -search(board, depth=CONFIG["search_depth"], alpha=float('-inf'), beta=float('inf'))
|
155 |
+
board.pop()
|
156 |
+
|
157 |
+
if not evaling:
|
158 |
+
break
|
159 |
+
|
160 |
+
keys = list(evaling.keys())
|
161 |
+
logits = torch.tensor(list(evaling.values())).to(device)
|
162 |
+
probs = torch.softmax(logits,dim=0)
|
163 |
+
epsilon = min(CONFIG["epsilon"],len(keys))
|
164 |
+
bests = torch.multinomial(probs,num_samples=epsilon,replacement=False)
|
165 |
+
best_idx = bests[torch.argmax(logits[bests])]
|
166 |
+
move = keys[best_idx.item()]
|
167 |
+
|
168 |
+
else:
|
169 |
+
result = engine.play(board, lim)
|
170 |
+
move = result.move
|
171 |
+
|
172 |
+
if is_bot_turn:
|
173 |
+
data.append({
|
174 |
+
'fen': board.fen(),
|
175 |
+
'move_number': mc,
|
176 |
+
})
|
177 |
+
|
178 |
+
board.push(move)
|
179 |
+
mc += 1
|
180 |
+
|
181 |
+
result = board.result()
|
182 |
+
c = 0
|
183 |
+
if result == '1-0':
|
184 |
+
c = 10.0
|
185 |
+
elif result == '0-1':
|
186 |
+
c = -10.0
|
187 |
+
return data, c, mc
|
188 |
+
def train(data, c, mc):
|
189 |
+
for entry in data:
|
190 |
+
tensor = board_to_tensor(chess.Board(entry['fen'])).to(device)
|
191 |
+
target = torch.tensor(c * entry['move_number'] / mc, dtype=torch.float32).to(device)
|
192 |
+
output = model(tensor)[0][0]
|
193 |
+
loss = criterion(output, target)
|
194 |
+
optimizer.zero_grad()
|
195 |
+
loss.backward()
|
196 |
+
optimizer.step()
|
197 |
+
|
198 |
+
print(f"Saving model to {CONFIG['model_path']}")
|
199 |
+
torch.save(model.state_dict(), CONFIG["model_path"])
|
200 |
+
return
|
201 |
+
def main():
|
202 |
+
for i in range(CONFIG["num_epochs"]):
|
203 |
+
mp.set_start_method('spawn', force=True)
|
204 |
+
num_games = CONFIG['num_games']
|
205 |
+
num_instances = mp.cpu_count()
|
206 |
+
print(f"Saving backup model to {CONFIG['backup_model_path']}")
|
207 |
+
torch.save(model.state_dict(), CONFIG["backup_model_path"])
|
208 |
+
with mp.Pool(processes=num_instances) as pool:
|
209 |
+
results_self = pool.starmap(game_gen, [(None,) for _ in range(num_games // 3)])
|
210 |
+
results_white = pool.starmap(game_gen, [(chess.WHITE,) for _ in range(num_games // 3)])
|
211 |
+
results_black = pool.starmap(game_gen, [(chess.BLACK,) for _ in range(num_games // 3)])
|
212 |
+
results = []
|
213 |
+
for s, w, b in zip(results_self, results_white, results_black):
|
214 |
+
results.extend([s, w, b])
|
215 |
+
for batch in results:
|
216 |
+
data, c, mc = batch
|
217 |
+
print(f"Saving backup model to {CONFIG['backup_model_path']}")
|
218 |
+
torch.save(model.state_dict(), CONFIG["backup_model_path"])
|
219 |
+
if data:
|
220 |
+
train(data, c, mc)
|
221 |
+
print("Training complete.")
|
222 |
+
engine.quit()
|
223 |
+
if __name__ == "__main__":
|
224 |
+
main()
|