philipp-zettl
commited on
Commit
•
ece90a2
1
Parent(s):
ee89e6e
update train script
Browse files
train.py
CHANGED
@@ -1,9 +1,15 @@
|
|
|
|
1 |
import argparse
|
|
|
2 |
import torch
|
3 |
import torch.nn as nn
|
|
|
4 |
from torch.nn import functional as F
|
5 |
from gpt_p.model import DecoderTransformer
|
|
|
|
|
6 |
from datasets import load_dataset
|
|
|
7 |
|
8 |
|
9 |
torch.manual_seed(420) # 1337
|
@@ -12,7 +18,7 @@ base_name = 'gpt-p_CHARS_CHAT_'
|
|
12 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
13 |
context_size = 256 # how many tokens to consider while generating the next
|
14 |
batch_size = 128 # how many independent sequences will we process in parallel
|
15 |
-
max_iters =
|
16 |
learning_rate = 3e-5
|
17 |
eval_interval = 100
|
18 |
eval_iters = 20 # number evaluation iterations
|
@@ -21,28 +27,304 @@ n_layer = 6 # number of transformer layers
|
|
21 |
n_head = 6
|
22 |
dropout = 0.2 # dropout factor
|
23 |
|
24 |
-
|
25 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
|
27 |
## BUILD DATA SET ##
|
|
|
|
|
|
|
|
|
28 |
book = content
|
29 |
-
|
|
|
|
|
|
|
30 |
vocab_size = len(characters)
|
31 |
|
32 |
# convert
|
33 |
-
|
34 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
35 |
|
36 |
-
|
37 |
-
|
38 |
|
|
|
|
|
|
|
|
|
|
|
39 |
|
40 |
-
|
41 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
train_data = data[:n]
|
43 |
val_data = data[n:]
|
44 |
|
45 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
46 |
def get_batch(split):
|
47 |
data = train_data if split == 'train' else val_data
|
48 |
idx = torch.randint(len(data) - context_size, (batch_size,))
|
@@ -50,6 +332,9 @@ def get_batch(split):
|
|
50 |
y = torch.stack([data[i+1:i+context_size+1] for i in idx])
|
51 |
return x.to(device), y.to(device)
|
52 |
|
|
|
|
|
|
|
53 |
## END BUILD DATA SET ##
|
54 |
## MODEL DEFINITION ##
|
55 |
|
@@ -72,15 +357,58 @@ def estimate_loss():
|
|
72 |
for k in range(eval_iters):
|
73 |
X, Y = get_batch(split)
|
74 |
logits, loss = model(X, Y)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
75 |
losses[k] = loss.item()
|
76 |
out[split] = losses.mean()
|
77 |
|
78 |
-
input_string = '1. e4 g6'
|
79 |
print_sample(torch.tensor(encode(input_string), dtype=torch.long, device=device).view((1, len(input_string))))
|
80 |
model.train()
|
81 |
return out
|
82 |
|
83 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
84 |
if __name__ == "__main__":
|
85 |
args = argparse.ArgumentParser()
|
86 |
args.add_argument('--load', '-l', action='store_true', default=False, help='Load model state.')
|
@@ -91,28 +419,62 @@ if __name__ == "__main__":
|
|
91 |
params = {'vocab_size': vocab_size, 'n_embed': n_embed, 'context_size': context_size, 'n_layer': n_layer, 'n_head': n_head, 'dropout': dropout}
|
92 |
if args.load:
|
93 |
m = DecoderTransformer(vocab_size, n_embed, context_size, n_layer, n_head, dropout)
|
94 |
-
m.load_state_dict(torch.load(f'./models/{base_name}' + ''.join(f'{key}={v}' for key, v in params.items())))
|
95 |
else:
|
96 |
m = DecoderTransformer(vocab_size, n_embed, context_size, n_layer, n_head, dropout)
|
97 |
model = m.to(device)
|
98 |
|
99 |
if args.inference:
|
|
|
|
|
|
|
|
|
|
|
|
|
100 |
exit()
|
101 |
## END MODEL ##
|
102 |
## START TRAINING ##
|
|
|
|
|
|
|
103 |
optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
|
|
|
|
|
104 |
|
105 |
-
for step in range(max_iters):
|
106 |
if step % eval_interval == 0:
|
107 |
losses = estimate_loss()
|
108 |
-
|
|
|
|
|
|
|
|
|
109 |
|
110 |
xb, yb = get_batch('train')
|
111 |
|
112 |
logits, loss = model(xb, yb)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
113 |
optimizer.zero_grad(set_to_none=True)
|
114 |
loss.backward()
|
115 |
optimizer.step()
|
|
|
|
|
116 |
|
117 |
print()
|
118 |
print('Loss:')
|
@@ -124,6 +486,8 @@ if __name__ == "__main__":
|
|
124 |
## END VALIDATION ##
|
125 |
|
126 |
# save model weights
|
127 |
-
torch.save(model.state_dict(), f'./models/{base_name}'
|
|
|
|
|
128 |
with open('train.log', 'a') as f:
|
129 |
f.write(f'{max_iters},{learning_rate}\n')
|
|
|
1 |
+
import re
|
2 |
import argparse
|
3 |
+
import json
|
4 |
import torch
|
5 |
import torch.nn as nn
|
6 |
+
from tqdm import tqdm
|
7 |
from torch.nn import functional as F
|
8 |
from gpt_p.model import DecoderTransformer
|
9 |
+
from torch.optim.lr_scheduler import _LRScheduler
|
10 |
+
import math
|
11 |
from datasets import load_dataset
|
12 |
+
import wandb
|
13 |
|
14 |
|
15 |
torch.manual_seed(420) # 1337
|
|
|
18 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
19 |
context_size = 256 # how many tokens to consider while generating the next
|
20 |
batch_size = 128 # how many independent sequences will we process in parallel
|
21 |
+
max_iters = 50_000
|
22 |
learning_rate = 3e-5
|
23 |
eval_interval = 100
|
24 |
eval_iters = 20 # number evaluation iterations
|
|
|
27 |
n_head = 6
|
28 |
dropout = 0.2 # dropout factor
|
29 |
|
30 |
+
mask_all_data = True
|
31 |
+
use_scheduler = False
|
32 |
+
|
33 |
+
dataset = load_dataset('Lichess/standard-chess-games', '2014-08', split='train')
|
34 |
+
og_samples = list(filter(lambda x: 'eval' not in x, dataset['movetext']))
|
35 |
+
|
36 |
+
|
37 |
+
new_dataset = load_dataset('Lichess/standard-chess-games', '2024-07', split='train', data_files=[f'data/year=2024/month=07/train-{str(i).zfill(5)}-of-00384.parquet' for i in range(10)])
|
38 |
+
|
39 |
+
new_dataset = [re.sub('[0-9]+\.\.\.', '', re.sub('{[^\}]*}', '', foo)).replace(' ', ' ').replace(' ', ' ') for foo in dataset['movetext']]
|
40 |
+
|
41 |
+
og_samples += new_dataset
|
42 |
+
|
43 |
+
if mask_all_data:
|
44 |
+
content = '\n'.join(list(filter(lambda x: 'eval' not in x, dataset['movetext'])))
|
45 |
+
else:
|
46 |
+
content = og_samples
|
47 |
+
|
48 |
+
print('Data loaded')
|
49 |
+
print('Training on ', len(content), 'characters. Good luck!')
|
50 |
|
51 |
## BUILD DATA SET ##
|
52 |
+
# load data
|
53 |
+
#with open('data.txt', 'r') as f:
|
54 |
+
# content = f.read()
|
55 |
+
|
56 |
book = content
|
57 |
+
if mask_all_data:
|
58 |
+
characters = sorted(list(set(book)))
|
59 |
+
else:
|
60 |
+
characters = sorted(list(set('\n'.join(book))))
|
61 |
vocab_size = len(characters)
|
62 |
|
63 |
# convert
|
64 |
+
class Tokenizer:
|
65 |
+
def __init__(self, vocab):
|
66 |
+
self.vocab = vocab
|
67 |
+
self.stoi = {ch: idx for idx, ch in enumerate(vocab)}
|
68 |
+
self.itos = {idx: ch for idx, ch in enumerate(vocab)}
|
69 |
+
|
70 |
+
def encode(self, s):
|
71 |
+
return [self.stoi[c] for c in s]
|
72 |
|
73 |
+
def decode(self, i):
|
74 |
+
return ''.join([self.itos[x] for x in i])
|
75 |
|
76 |
+
@classmethod
|
77 |
+
def from_pretrained(cls, path):
|
78 |
+
with open(path, 'r') as f:
|
79 |
+
vocab = json.load(f)
|
80 |
+
return cls(vocab)
|
81 |
|
82 |
+
def save_pretrained(self, path):
|
83 |
+
with open(path, 'w') as f:
|
84 |
+
json.dump(self.vocab, f)
|
85 |
+
|
86 |
+
|
87 |
+
tokenizer = Tokenizer(characters)
|
88 |
+
encode = tokenizer.encode
|
89 |
+
decode = tokenizer.decode
|
90 |
+
|
91 |
+
if mask_all_data:
|
92 |
+
data = torch.tensor(encode(book), dtype=torch.long)
|
93 |
+
else:
|
94 |
+
data = [torch.tensor(encode(s), dtype=torch.long) for s in book]
|
95 |
+
max_len = max(len(x) for x in og_samples)
|
96 |
+
context_size = min(context_size, max_len)
|
97 |
+
|
98 |
+
|
99 |
+
n = int(0.8 * len(data))
|
100 |
train_data = data[:n]
|
101 |
val_data = data[n:]
|
102 |
|
103 |
|
104 |
+
|
105 |
+
# Constants for piece movement validation
|
106 |
+
PIECE_VALUES = {
|
107 |
+
'P': 1, 'N': 3, 'B': 3, 'R': 5, 'Q': 9, 'K': 0, # White pieces
|
108 |
+
'p': 1, 'n': 3, 'b': 3, 'r': 5, 'q': 9, 'k': 0 # Black pieces
|
109 |
+
}
|
110 |
+
|
111 |
+
def initialize_board():
|
112 |
+
"""Initializes the standard chessboard setup."""
|
113 |
+
return [
|
114 |
+
['r', 'n', 'b', 'q', 'k', 'b', 'n', 'r'], # 8th rank (Black)
|
115 |
+
['p', 'p', 'p', 'p', 'p', 'p', 'p', 'p'], # 7th rank (Black)
|
116 |
+
['.', '.', '.', '.', '.', '.', '.', '.'], # 6th rank
|
117 |
+
['.', '.', '.', '.', '.', '.', '.', '.'], # 5th rank
|
118 |
+
['.', '.', '.', '.', '.', '.', '.', '.'], # 4th rank
|
119 |
+
['.', '.', '.', '.', '.', '.', '.', '.'], # 3rd rank
|
120 |
+
['P', 'P', 'P', 'P', 'P', 'P', 'P', 'P'], # 2nd rank (White)
|
121 |
+
['R', 'N', 'B', 'Q', 'K', 'B', 'N', 'R'] # 1st rank (White)
|
122 |
+
]
|
123 |
+
|
124 |
+
def get_piece(board, position):
|
125 |
+
"""Returns the piece at a given board position (e.g., e4 -> 'P' or '.')."""
|
126 |
+
col = ord(position[0]) - ord('a')
|
127 |
+
row = 8 - int(position[1])
|
128 |
+
return board[row][col]
|
129 |
+
|
130 |
+
def set_piece(board, position, piece):
|
131 |
+
"""Sets a piece on the board at a given position."""
|
132 |
+
col = ord(position[0]) - ord('a')
|
133 |
+
row = 8 - int(position[1])
|
134 |
+
board[row][col] = piece
|
135 |
+
|
136 |
+
def validate_pawn_move(board, start, end, is_white_turn):
|
137 |
+
"""Validates pawn movement including capturing, advancing, and promotion."""
|
138 |
+
start_col, start_row = ord(start[0]) - ord('a'), 8 - int(start[1])
|
139 |
+
end_col, end_row = ord(end[0]) - ord('a'), 8 - int(end[1])
|
140 |
+
|
141 |
+
pawn_direction = -1 if is_white_turn else 1 # White moves up, black moves down
|
142 |
+
|
143 |
+
# Regular forward move
|
144 |
+
if start_col == end_col and board[end_row][end_col] == '.':
|
145 |
+
if start_row + pawn_direction == end_row: # 1 square move
|
146 |
+
return True
|
147 |
+
if (is_white_turn and start_row == 6 or not is_white_turn and start_row == 1) and start_row + 2 * pawn_direction == end_row:
|
148 |
+
return True
|
149 |
+
|
150 |
+
# Capture
|
151 |
+
if abs(start_col - end_col) == 1 and start_row + pawn_direction == end_row:
|
152 |
+
target_piece = board[end_row][end_col]
|
153 |
+
if (is_white_turn and target_piece.islower()) or (not is_white_turn and target_piece.isupper()):
|
154 |
+
return True
|
155 |
+
|
156 |
+
return False
|
157 |
+
|
158 |
+
def validate_knight_move(start, end):
|
159 |
+
"""Validates knight movement (L-shape)."""
|
160 |
+
start_col, start_row = ord(start[0]) - ord('a'), 8 - int(start[1])
|
161 |
+
end_col, end_row = ord(end[0]) - ord('a'), 8 - int(end[1])
|
162 |
+
|
163 |
+
col_diff = abs(start_col - end_col)
|
164 |
+
row_diff = abs(start_row - end_row)
|
165 |
+
|
166 |
+
return (col_diff == 2 and row_diff == 1) or (col_diff == 1 and row_diff == 2)
|
167 |
+
|
168 |
+
def validate_rook_move(board, start, end):
|
169 |
+
"""Validates rook movement (straight lines along rank or file)."""
|
170 |
+
start_col, start_row = ord(start[0]) - ord('a'), 8 - int(start[1])
|
171 |
+
end_col, end_row = ord(end[0]) - ord('a'), 8 - int(end[1])
|
172 |
+
|
173 |
+
if start_col != end_col and start_row != end_row:
|
174 |
+
return False # Must be either same column or row
|
175 |
+
|
176 |
+
# Check if path is clear
|
177 |
+
if start_col == end_col:
|
178 |
+
step = 1 if end_row > start_row else -1
|
179 |
+
for row in range(start_row + step, end_row, step):
|
180 |
+
if board[row][start_col] != '.':
|
181 |
+
return False
|
182 |
+
else:
|
183 |
+
step = 1 if end_col > start_col else -1
|
184 |
+
for col in range(start_col + step, end_col, step):
|
185 |
+
if board[start_row][col] != '.':
|
186 |
+
return False
|
187 |
+
|
188 |
+
return True
|
189 |
+
|
190 |
+
def validate_bishop_move(board, start, end):
|
191 |
+
"""Validates bishop movement (diagonals)."""
|
192 |
+
start_col, start_row = ord(start[0]) - ord('a'), 8 - int(start[1])
|
193 |
+
end_col, end_row = ord(end[0]) - ord('a'), 8 - int(end[1])
|
194 |
+
|
195 |
+
if abs(start_col - end_col) != abs(start_row - end_row):
|
196 |
+
return False # Must move diagonally
|
197 |
+
|
198 |
+
# Check if path is clear
|
199 |
+
col_step = 1 if end_col > start_col else -1
|
200 |
+
row_step = 1 if end_row > start_row else -1
|
201 |
+
col, row = start_col + col_step, start_row + row_step
|
202 |
+
while col != end_col and row != end_row:
|
203 |
+
if board[row][col] != '.':
|
204 |
+
return False
|
205 |
+
col += col_step
|
206 |
+
row += row_step
|
207 |
+
|
208 |
+
return True
|
209 |
+
|
210 |
+
def validate_move(board, move, is_white_turn):
|
211 |
+
"""Validates a move based on the current board state."""
|
212 |
+
if move == "O-O" or move == "O-O-O":
|
213 |
+
return True # Castling placeholder
|
214 |
+
|
215 |
+
piece_type = 'P' if move[0].islower() else move[0]
|
216 |
+
start = move[-2:] # Simplification; would need to parse actual source square
|
217 |
+
end = move[-2:] # Actual end position is the destination
|
218 |
+
|
219 |
+
if piece_type == 'P':
|
220 |
+
return validate_pawn_move(board, start, end, is_white_turn)
|
221 |
+
elif piece_type == 'N':
|
222 |
+
return validate_knight_move(start, end)
|
223 |
+
elif piece_type == 'R':
|
224 |
+
return validate_rook_move(board, start, end)
|
225 |
+
elif piece_type == 'B':
|
226 |
+
return validate_bishop_move(board, start, end)
|
227 |
+
|
228 |
+
# Other pieces can be added similarly
|
229 |
+
return True # Placeholder for other pieces
|
230 |
+
|
231 |
+
def update_board(board, move, is_white_turn):
|
232 |
+
"""Updates the board according to the move."""
|
233 |
+
start = move[-2:]
|
234 |
+
end = move[-2:]
|
235 |
+
piece = get_piece(board, start)
|
236 |
+
|
237 |
+
# Move the piece
|
238 |
+
set_piece(board, end, piece)
|
239 |
+
set_piece(board, start, '.')
|
240 |
+
|
241 |
+
return board # Placeholder for now
|
242 |
+
|
243 |
+
def validate_pgn(pgn_string):
|
244 |
+
"""
|
245 |
+
Validates the PGN string format and chess move legality.
|
246 |
+
"""
|
247 |
+
|
248 |
+
move_pattern = r'([PNBRQK]?[a-h]?[1-8]?[x]?[a-h][1-8](=[QRNB])?|O-O(-O)?)[+#]?' # Chess move
|
249 |
+
result_pattern = r'(1-0|0-1|1/2-1/2)' # Game results
|
250 |
+
tag_pattern = r'\[([A-Za-z0-9_]+)\s+"([^"]+)"\]' # PGN tags
|
251 |
+
|
252 |
+
pgn_lines = pgn_string.strip().splitlines()
|
253 |
+
|
254 |
+
tags = [line for line in pgn_lines if line.startswith('[')]
|
255 |
+
for tag in tags:
|
256 |
+
if not re.match(tag_pattern, tag):
|
257 |
+
return False # Invalid tag format
|
258 |
+
|
259 |
+
moves_section = ' '.join([line for line in pgn_lines if not line.startswith('[')]).strip()
|
260 |
+
|
261 |
+
if not re.search(result_pattern, moves_section):
|
262 |
+
return False # No valid result found
|
263 |
+
|
264 |
+
moves_section = re.sub(result_pattern, '', moves_section).strip()
|
265 |
+
|
266 |
+
board = initialize_board()
|
267 |
+
is_white_turn = True
|
268 |
+
|
269 |
+
move_tokens = re.split(r'\s|\d+\.', moves_section)
|
270 |
+
for token in move_tokens:
|
271 |
+
if token:
|
272 |
+
if not re.match(move_pattern, token):
|
273 |
+
return False # Invalid move format
|
274 |
+
|
275 |
+
if not validate_move(board, token, is_white_turn):
|
276 |
+
return False # Invalid chess move
|
277 |
+
|
278 |
+
board = update_board(board, token, is_white_turn)
|
279 |
+
is_white_turn = not is_white_turn
|
280 |
+
|
281 |
+
return True
|
282 |
+
|
283 |
+
# Test case
|
284 |
+
pgn_string = """
|
285 |
+
[Event "World Championship"]
|
286 |
+
[Site "Moscow URS"]
|
287 |
+
[Date "1985.11.09"]
|
288 |
+
[Round "16"]
|
289 |
+
[White "Kasparov, Garry"]
|
290 |
+
[Black "Karpov, Anatoly"]
|
291 |
+
[Result "1-0"]
|
292 |
+
|
293 |
+
1. e4 e5 2. Nf3 Nc6 3. Bb5 a6 4. Ba4 Nf6 5. O-O Be7 6. Re1 b5 7. Bb3 d6
|
294 |
+
8. c3 O-O 9. h3 Nb8 10. d4 Nbd7 11. c4 Bb7 12. Nbd2 c6 13. Bc2 Re8 14. b3 Bf8
|
295 |
+
15. Bb2 Qc7 16. Rc1 Rad8 17. a3 Qb8 18. Bd3 g6 19. Qc2 Nh5 20. g3 Ng7 21. Qb1
|
296 |
+
exd4 22. Nxd4 c5 23. N4f3 Ne6 24. Bf1 Ne5 25. Qa1 Nxf3+ 26. Nxf3 Qa8 27. b4
|
297 |
+
Rc8 28. Bd3 Bh6 29. Rc2 Bc6 30. h4 f5 31. exf5 Bxf3 32. fxe6 Bh1 33. Bf1 Qf3
|
298 |
+
34. Re2 Bg7 35. Kh2 Rc7 36. Bxg7 Rxg7 37. Qf6 bxc4 38. e7 Qxf6 39. exf6 1-0
|
299 |
+
"""
|
300 |
+
|
301 |
+
|
302 |
+
|
303 |
+
def get_batch_from_samples(split):
|
304 |
+
data = train_data if split == 'train' else val_data
|
305 |
+
sample_idx = torch.randint(len(data), (batch_size,))
|
306 |
+
inputs = []
|
307 |
+
outputs = []
|
308 |
+
space = encode(' ')[0]
|
309 |
+
for idx in sample_idx:
|
310 |
+
sample_size = len(data[idx])
|
311 |
+
start = torch.randint(max(sample_size - 2, sample_size - context_size), (1,))
|
312 |
+
end = start + context_size
|
313 |
+
i1 = data[idx][start:end].tolist()
|
314 |
+
i2 = [space] * (context_size - len(i1))
|
315 |
+
input_sample = torch.tensor(i1 + i2)
|
316 |
+
o1 = data[idx][start+1:end+1].tolist()
|
317 |
+
o2 = [space] * (context_size - len(o1))
|
318 |
+
output_sample = torch.tensor(o1 + o2)
|
319 |
+
|
320 |
+
inputs.append(input_sample)
|
321 |
+
outputs.append(output_sample)
|
322 |
+
|
323 |
+
x = torch.stack(inputs)
|
324 |
+
y = torch.stack(outputs)
|
325 |
+
return x.to(device), y.to(device)
|
326 |
+
|
327 |
+
|
328 |
def get_batch(split):
|
329 |
data = train_data if split == 'train' else val_data
|
330 |
idx = torch.randint(len(data) - context_size, (batch_size,))
|
|
|
332 |
y = torch.stack([data[i+1:i+context_size+1] for i in idx])
|
333 |
return x.to(device), y.to(device)
|
334 |
|
335 |
+
if not mask_all_data:
|
336 |
+
get_batch = get_batch_from_samples
|
337 |
+
|
338 |
## END BUILD DATA SET ##
|
339 |
## MODEL DEFINITION ##
|
340 |
|
|
|
357 |
for k in range(eval_iters):
|
358 |
X, Y = get_batch(split)
|
359 |
logits, loss = model(X, Y)
|
360 |
+
"""
|
361 |
+
input_string = X[0].tolist()
|
362 |
+
gen = model.generate(X[0].view(1, -1), max_new_tokens=5, context_size=context_size)
|
363 |
+
o = tokenizer.decode(gen[0].tolist())
|
364 |
+
try:
|
365 |
+
valid = int(not validate_pgn(o))
|
366 |
+
except Exception:
|
367 |
+
valid = 2
|
368 |
+
"""
|
369 |
losses[k] = loss.item()
|
370 |
out[split] = losses.mean()
|
371 |
|
372 |
+
input_string = '1. e4 g6 2.'
|
373 |
print_sample(torch.tensor(encode(input_string), dtype=torch.long, device=device).view((1, len(input_string))))
|
374 |
model.train()
|
375 |
return out
|
376 |
|
377 |
|
378 |
+
class CosineAnnealingScheduler(_LRScheduler):
|
379 |
+
def __init__(self, optimizer, T_max, eta_min=0, last_epoch=-1):
|
380 |
+
"""
|
381 |
+
Args:
|
382 |
+
optimizer (Optimizer): Wrapped optimizer.
|
383 |
+
T_max (int): Maximum number of iterations.
|
384 |
+
eta_min (float): Minimum learning rate. Default: 0.
|
385 |
+
last_epoch (int): The index of last epoch. Default: -1.
|
386 |
+
"""
|
387 |
+
self.T_max = T_max
|
388 |
+
self.eta_min = eta_min
|
389 |
+
super().__init__(optimizer, last_epoch)
|
390 |
+
|
391 |
+
def get_lr(self):
|
392 |
+
if not self._get_lr_called_within_step:
|
393 |
+
warnings.warn("To get the last learning rate computed by the scheduler, "
|
394 |
+
"please use `get_last_lr()`.", UserWarning)
|
395 |
+
|
396 |
+
if self.last_epoch == 0:
|
397 |
+
return [group['lr'] for group in self.optimizer.param_groups]
|
398 |
+
elif self._step_count == 1 and self.last_epoch > 0:
|
399 |
+
return [self.eta_min + (base_lr - self.eta_min) *
|
400 |
+
(1 + math.cos((self.last_epoch) * math.pi / self.T_max)) / 2
|
401 |
+
for base_lr in self.base_lrs]
|
402 |
+
elif (self.last_epoch - 1 - self.T_max) % (2 * self.T_max) == 0:
|
403 |
+
return [group['lr'] + (base_lr - self.eta_min) *
|
404 |
+
(1 - math.cos(math.pi / self.T_max)) / 2
|
405 |
+
for base_lr, group in
|
406 |
+
zip(self.base_lrs, self.optimizer.param_groups)]
|
407 |
+
return [(1 + math.cos(math.pi * self.last_epoch / self.T_max)) /
|
408 |
+
(1 + math.cos(math.pi * (self.last_epoch - 1) / self.T_max)) *
|
409 |
+
(group['lr'] - self.eta_min) + self.eta_min
|
410 |
+
for group in self.optimizer.param_groups]
|
411 |
+
|
412 |
if __name__ == "__main__":
|
413 |
args = argparse.ArgumentParser()
|
414 |
args.add_argument('--load', '-l', action='store_true', default=False, help='Load model state.')
|
|
|
419 |
params = {'vocab_size': vocab_size, 'n_embed': n_embed, 'context_size': context_size, 'n_layer': n_layer, 'n_head': n_head, 'dropout': dropout}
|
420 |
if args.load:
|
421 |
m = DecoderTransformer(vocab_size, n_embed, context_size, n_layer, n_head, dropout)
|
422 |
+
m.load_state_dict(torch.load(f'./models/{base_name}'))# + ''.join(f'{key}={v}' for key, v in params.items())))
|
423 |
else:
|
424 |
m = DecoderTransformer(vocab_size, n_embed, context_size, n_layer, n_head, dropout)
|
425 |
model = m.to(device)
|
426 |
|
427 |
if args.inference:
|
428 |
+
input_string = input('Enter a PGN string: ')
|
429 |
+
print_sample(torch.tensor(encode(input_string), dtype=torch.long, device=device).view((1, len(input_string))))
|
430 |
+
with open(f'./models/{base_name}_params.json', 'w') as f:
|
431 |
+
json.dump(params, f)
|
432 |
+
|
433 |
+
tokenizer.save_pretrained(f'./models/{base_name}_vocab.json')
|
434 |
exit()
|
435 |
## END MODEL ##
|
436 |
## START TRAINING ##
|
437 |
+
wandb.init(project='chessPT')
|
438 |
+
|
439 |
+
wandb.watch(model)
|
440 |
optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
|
441 |
+
if use_scheduler:
|
442 |
+
scheduler = CosineAnnealingScheduler(optimizer, max_iters, eta_min=learning_rate//1e6)
|
443 |
|
444 |
+
for step in tqdm(range(max_iters), total=max_iters, desc='Training'):
|
445 |
if step % eval_interval == 0:
|
446 |
losses = estimate_loss()
|
447 |
+
if use_scheduler:
|
448 |
+
print(f'step {step:4d}: train loss {losses["train"]:.4f}, val loss: {losses["val"]:.4f}, lr: {scheduler.get_last_lr()[0]}')
|
449 |
+
else:
|
450 |
+
print(f'step {step:4d}: train loss {losses["train"]:.4f}, val loss: {losses["val"]:.4f}')
|
451 |
+
wandb.log({'train_loss': losses['train'], 'val_loss': losses['val']})
|
452 |
|
453 |
xb, yb = get_batch('train')
|
454 |
|
455 |
logits, loss = model(xb, yb)
|
456 |
+
"""
|
457 |
+
|
458 |
+
input_string = xb[0].tolist()
|
459 |
+
gen = model.generate(xb[0].view(1, -1), max_new_tokens=5, context_size=context_size)
|
460 |
+
out = tokenizer.decode(gen[0].tolist())
|
461 |
+
try:
|
462 |
+
valid = int(not validate_pgn(out))
|
463 |
+
except Exception:
|
464 |
+
valid = 2
|
465 |
+
loss += valid
|
466 |
+
"""
|
467 |
+
|
468 |
+
if use_scheduler:
|
469 |
+
wandb.log({'running_train_loss': loss.item(), 'lr': scheduler.get_last_lr()[0]})
|
470 |
+
else:
|
471 |
+
wandb.log({'running_train_loss': loss.item()})
|
472 |
+
|
473 |
optimizer.zero_grad(set_to_none=True)
|
474 |
loss.backward()
|
475 |
optimizer.step()
|
476 |
+
if use_scheduler:
|
477 |
+
scheduler.step()
|
478 |
|
479 |
print()
|
480 |
print('Loss:')
|
|
|
486 |
## END VALIDATION ##
|
487 |
|
488 |
# save model weights
|
489 |
+
torch.save(model.state_dict(), f'./models/{base_name}')
|
490 |
+
with open(f'./models/{base_name}_params.json', 'w') as f:
|
491 |
+
json.dump(params, f)
|
492 |
with open('train.log', 'a') as f:
|
493 |
f.write(f'{max_iters},{learning_rate}\n')
|