-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_kyu_20M.py
411 lines (346 loc) · 13.9 KB
/
train_kyu_20M.py
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
from bisect import bisect
from collections import defaultdict
from datetime import datetime, timedelta
from timeit import default_timer as timer
import numpy as np
import torch
import torch.nn as nn
from lion_pytorch import Lion
from torch.utils.data import DataLoader, random_split
from tqdm.auto import tqdm
class BatchRenorm2d(nn.Module):
def __init__(
self,
num_features: int,
eps: float = 1e-3,
momentum: float = 0.01,
affine: bool = True,
):
super().__init__()
self.register_buffer("running_mean", torch.zeros(num_features, dtype=torch.float))
self.register_buffer("running_std", torch.ones(num_features, dtype=torch.float))
self.register_buffer("num_batches_tracked", torch.tensor(0, dtype=torch.long))
self.weight = torch.nn.Parameter(torch.ones(num_features, dtype=torch.float))
self.bias = torch.nn.Parameter(torch.zeros(num_features, dtype=torch.float))
self.affine = affine
self.eps = eps
self.step = 0
self.momentum = momentum
@property
def rmax(self) -> torch.Tensor:
return (2 / 35000 * self.num_batches_tracked + 25 / 35).clamp_(1.0, 3.0)
@property
def dmax(self) -> torch.Tensor:
return (5 / 20000 * self.num_batches_tracked - 25 / 20).clamp_(0.0, 5.0)
def forward(self, x: torch.Tensor, mask=None) -> torch.Tensor:
"""
Mask is a boolean tensor used for indexing, where True values are padded
i.e for 3D input, mask should be of shape (batch_size, seq_len)
mask is used to prevent padded values from affecting the batch statistics
"""
if x.dim() > 2:
x = x.transpose(1, -1)
if self.training:
dims = [i for i in range(x.dim() - 1)]
if mask is not None:
z = x[~mask]
batch_mean = z.mean(0)
batch_std = z.std(0, unbiased=False) + self.eps
else:
batch_mean = x.mean(dims)
batch_std = x.std(dims, unbiased=False) + self.eps
r = (batch_std.detach() / self.running_std.view_as(batch_std)).clamp_(1 / self.rmax, self.rmax)
d = (
(batch_mean.detach() - self.running_mean.view_as(batch_mean)) / self.running_std.view_as(batch_std)
).clamp_(-self.dmax, self.dmax)
x = (x - batch_mean) / batch_std * r + d
self.running_mean += self.momentum * (batch_mean.detach() - self.running_mean)
self.running_std += self.momentum * (batch_std.detach() - self.running_std)
self.num_batches_tracked += 1
else:
x = (x - self.running_mean) / self.running_std
if self.affine:
x = self.weight * x + self.bias
if x.dim() > 2:
x = x.transpose(1, -1)
return x
eps = 1e-3
momentum = 1e-2
class GlobalPool(nn.Module):
def __init__(self):
super().__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.max_pool = nn.AdaptiveMaxPool2d(1)
def forward(self, x):
avg_out = self.avg_pool(x)
max_out = self.max_pool(x)
out = torch.cat([avg_out, max_out], dim=1)
return out
class SEBlock(nn.Module):
def __init__(self, channels, reduction=3):
super().__init__()
self.channels = channels
self.pool = GlobalPool()
self.conv = nn.Sequential(
nn.Conv2d(channels * 2, channels // reduction, kernel_size=1, padding=0, bias=True),
nn.Mish(inplace=True),
nn.Conv2d(channels // reduction, channels * 2, kernel_size=1, padding=0, bias=True),
)
self.sigmoid = nn.Sigmoid()
def forward(self, x):
b, c, _, _ = x.size()
out = self.pool(x)
out = self.conv(out)
gammas, betas = torch.split(out, self.channels, dim=1)
gammas = torch.reshape(gammas, (b, c, 1, 1))
betas = torch.reshape(betas, (b, c, 1, 1))
out = self.sigmoid(gammas) * x + betas
return out
class NormActConv(nn.Module):
def __init__(self, in_channels, out_channels, use_se=False):
super().__init__()
self.norm = BatchRenorm2d(in_channels, eps=eps, momentum=momentum)
self.act = nn.Mish(inplace=True)
self.conv_3x3 = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1, bias=False)
self.conv_1x1 = nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0, bias=False)
self.se = SEBlock(out_channels) if use_se else None
def forward(self, x):
out = x
out = self.norm(out)
out = self.act(out)
if self.se is None:
return self.conv_3x3(out) + self.conv_1x1(out)
else:
return self.se(out)
class InnerResidualBlock(nn.Module):
def __init__(self, channels, use_se=False):
super().__init__()
self.conv1 = NormActConv(channels, channels, use_se=use_se)
self.conv2 = NormActConv(channels, channels)
def forward(self, x):
out = x
out = self.conv1(out)
out = self.conv2(out)
return out + x
class NestedResidualBlock(nn.Module):
def __init__(self, channels, use_se=False):
super().__init__()
c = channels
c2 = c // 2
self.conv_in = NormActConv(c, c2)
self.inner_block1 = InnerResidualBlock(c2, use_se=use_se)
self.inner_block2 = InnerResidualBlock(c2)
self.conv_out = NormActConv(c2, c)
def forward(self, x):
out = x
out = self.conv_in(out)
out = self.inner_block1(out)
out = self.inner_block2(out)
out = self.conv_out(out)
return out + x
class PolicyHead(nn.Module):
def __init__(self, channels):
super().__init__()
self.conv = nn.Conv2d(channels, 2, kernel_size=1, padding=0, bias=False)
self.norm = BatchRenorm2d(2, eps=eps, momentum=momentum)
self.act = nn.Mish(inplace=True)
self.fc = nn.Linear(2 * 19 * 19, 19 * 19)
def forward(self, x):
out = self.act(self.norm(self.conv(x)))
out = out.view(out.size(0), -1)
out = self.fc(out)
return out
class Model(nn.Module):
def __init__(self):
super().__init__()
channels = 192
self.conv_in = nn.Conv2d(17, channels, kernel_size=3, padding=1, bias=False)
self.blocks = []
for _ in range(2):
self.blocks += [
NestedResidualBlock(channels),
NestedResidualBlock(channels, use_se=True),
]
self.blocks.append(NestedResidualBlock(channels))
self.blocks = nn.Sequential(*self.blocks)
self.policy_head = PolicyHead(channels)
def forward(self, x):
out = self.conv_in(x)
out = self.blocks(out)
out = self.policy_head(out)
return out
# Define the training step:
def train_step(
model: torch.nn.Module,
dataloader: torch.utils.data.DataLoader,
loss_func: torch.nn.Module,
optimizer: torch.optim.Optimizer,
device: torch.device,
scaler: torch.cuda.amp.GradScaler,
) -> tuple[float, float]:
# Set the model to train mode
model.train()
# Set the training loss to 0
train_loss = train_accuracy = 0
# Iterate over the DataLoader batches
for batch, (X, y) in enumerate(dataloader):
# Move the batch to the device
X, y = X.to(device, non_blocking=True), y.to(device, non_blocking=True)
# 1. Zero the gradients
optimizer.zero_grad()
with torch.cuda.amp.autocast(enabled=False):
# 2. Forward pass
y_pred = model(X)
# 3. Calculate and accumulate loss
loss = loss_func(y_pred, y)
train_loss += loss.item()
with torch.autograd.detect_anomaly():
# 4. Backward pass
scaler.scale(loss).backward()
# 5. Update the optimizer
scaler.step(optimizer)
scaler.update()
# 6. Calculate the accuracy
topk_preds = torch.topk(y_pred, 1, dim=1)[1]
correct = topk_preds.eq(y.view(-1, 1).expand_as(topk_preds)).sum().item()
train_accuracy += correct / len(y)
# Return average loss
return train_loss / len(dataloader), train_accuracy / len(dataloader)
# Define the validating step:
# Turn on inference context manager
@torch.no_grad()
def valid_step(
model: torch.nn.Module,
dataloader: torch.utils.data.DataLoader,
loss_func: torch.nn.Module,
device: torch.device,
) -> tuple[float, float]:
# Set the model to eval mode
model.eval()
# Set the validating loss to 0
valid_loss = valid_accuracy = 0
# Iterate over the DataLoader batches
for batch, (X, y) in enumerate(dataloader):
# Move the batch to the device
X, y = X.to(device, non_blocking=True), y.to(device, non_blocking=True)
# 1. Forward pass
y_pred = model(X)
# 2. Calculate and accumulate loss
valid_loss += loss_func(y_pred, y).item()
# 3. Calculate the accuracy
topk_preds = torch.topk(y_pred, 1, dim=1)[1]
correct = topk_preds.eq(y.view(-1, 1).expand_as(topk_preds)).sum().item()
valid_accuracy += correct / len(y)
# Return average loss
return valid_loss / len(dataloader), valid_accuracy / len(dataloader)
# Define the training and validating loops:
def train(
model: torch.nn.Module,
train_dataloader: torch.utils.data.DataLoader,
valid_dataloader: torch.utils.data.DataLoader,
loss_func: torch.nn.Module,
optimizer: torch.optim.Optimizer,
scheduler: torch.optim.lr_scheduler,
epochs: int,
device: torch.device,
) -> defaultdict:
# Init the results
result = defaultdict(list)
# Set the model to the device
model.to(device)
best_accuracy = 0
best_loss = float("inf")
patience = 2
early_stopping_counter = 0
scaler = torch.cuda.amp.GradScaler()
torch.backends.cudnn.benchmark = True
elapsed_time = 0
# Iterate over the epochs
for epoch in tqdm(range(1, epochs + 1)):
start_time = datetime.now()
print(f"\n{start_time.strftime('%Y/%m/%d %H:%M:%S')} ({elapsed_time})")
# Train the model
train_loss, train_accuracy = train_step(model, train_dataloader, loss_func, optimizer, device, scaler)
scheduler.step()
# Validate the model
valid_loss, valid_accuracy = valid_step(model, valid_dataloader, loss_func, device)
if valid_accuracy > best_accuracy:
best_accuracy = valid_accuracy
torch.save(model.state_dict(), "./checkpoints/kyu_20M_best_model.pth")
# Record the loss
result["train_loss"].append(train_loss)
result["valid_loss"].append(valid_loss)
# Record the accuracy
result["train_accuracy"].append(train_accuracy)
result["valid_accuracy"].append(valid_accuracy)
# Print the results for this epoch:
if epoch:
print(
f"Epoch: {epoch}, train_loss: {train_loss}, train_accuracy: {train_accuracy}, "
f"valid_loss: {valid_loss}, valid_accuracy: {valid_accuracy}"
)
elapsed_time = timedelta(seconds=(datetime.now() - start_time).seconds)
if valid_loss < best_loss:
best_loss = valid_loss
early_stopping_counter = 0
else:
early_stopping_counter += 1
if early_stopping_counter >= patience:
print(f"Early stopping at epoch {epoch}...")
break
# Return the results
return result
class BoardDataSet(torch.utils.data.Dataset):
def __init__(self, data_paths, target_paths):
self.data_memmaps = [np.load(path, mmap_mode="r") for path in data_paths]
self.target_memmaps = [np.load(path, mmap_mode="r") for path in target_paths]
self.start_indices = [0] * len(data_paths)
self.data_count = 0
for index, memmap in enumerate(self.data_memmaps):
self.start_indices[index] = self.data_count
self.data_count += memmap.shape[0]
def __len__(self):
return self.data_count
def __getitem__(self, index):
memmap_index = bisect(self.start_indices, index) - 1
index_in_memmap = index - self.start_indices[memmap_index]
data = self.data_memmaps[memmap_index][index_in_memmap]
target = self.target_memmaps[memmap_index][index_in_memmap]
return torch.tensor(data, dtype=torch.float32), torch.tensor(target, dtype=torch.long)
def load_data(train_size=0.8, batch_size=512):
data_paths = []
target_paths = []
for i in range(1, 8):
data_paths.append(f"./data/kyu/20M/train_x_{i}.npz")
target_paths.append(f"./data/kyu/20M/train_y_{i}.npz")
traindataset = BoardDataSet(data_paths, target_paths)
print(traindataset.__len__())
train_size = int(train_size * len(traindataset))
valid_size = len(traindataset) - train_size
train_dataset, valid_dataset = random_split(traindataset, [train_size, valid_size])
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, pin_memory=True, num_workers=20)
valid_loader = DataLoader(valid_dataset, batch_size=batch_size, shuffle=True, pin_memory=True, num_workers=20)
return train_loader, valid_loader
def train_model(epochs=15):
train_loader, valid_loader = load_data()
model = Model().to(device)
model.load_state_dict(torch.load("./checkpoints/kyu_10M_best_model.pth"))
loss_func = torch.nn.CrossEntropyLoss(label_smoothing=0.1)
optimizer = Lion(model.parameters(), lr=1e-4, weight_decay=1e-2)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=epochs)
start_time = timer()
train(
model,
train_loader,
valid_loader,
loss_func,
optimizer,
scheduler,
epochs,
device,
)
end_time = timer()
print(f"Total training time: {timedelta(seconds=end_time-start_time)}")
if __name__ == "__main__":
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
train_model()