Repository navigation
Expand file tree
/
Copy pathmlp_model.py
More file actions
484 lines (404 loc) · 16.6 KB
/
Copy pathmlp_model.py
File metadata and controls
484 lines (404 loc) · 16.6 KB
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
"""
Multi-layer perceptron model for alpha factor prediction.
"""
import copy
from collections import defaultdict
from collections.abc import Callable
from typing import Literal, cast
import numpy as np
import pandas as pd
import polars as pl
from sklearn.metrics import mean_squared_error # type: ignore
import torch
import torch.nn as nn
import torch.optim as optim
from vnpy.alpha import (
AlphaDataset,
AlphaModel,
Segment,
logger
)
class MlpModel(AlphaModel):
"""
Multi-Layer Perceptron Model
Alpha factor prediction model implemented using multi-layer perceptron, with main features including:
1. Building and training multi-layer perceptron neural networks
2. Predicting Alpha factor values
3. Model evaluation and feature importance analysis
4. Support for early stopping and overfitting prevention
5. Support for MSE loss function
6. Optional Adam or SGD optimizer
"""
def __init__(
self,
input_size: int,
hidden_sizes: tuple[int] = (256,),
lr: float = 0.001,
n_epochs: int = 300,
batch_size: int = 2000,
early_stop_rounds: int = 50,
eval_steps: int = 20,
optimizer: Literal["sgd", "adam"] = "adam",
weight_decay: float = 0.0,
device: str = "cpu",
seed: int | None = None
) -> None:
"""
Initialize the network on the given device, accepting only adam or sgd, and seed NumPy and PyTorch when a seed is provided.
"""
# Save model hyperparameters
self.input_size: int = input_size
self.hidden_sizes: tuple[int] = hidden_sizes
self.lr: float = lr
self.n_epochs: int = n_epochs
self.batch_size: int = batch_size
self.early_stop_rounds: int = early_stop_rounds
self.eval_steps: int = eval_steps
self.device: str = device
self.fitted: bool = False
self.feature_names: list[str] = []
self.best_step: int | None = None
# Set random seed for reproducibility
if seed is not None:
np.random.seed(seed)
torch.manual_seed(seed)
# Set loss function type
self._scorer: Callable[..., float] = mean_squared_error
# Initialize model
self.model: nn.Module = MlpNetwork(
input_size=input_size,
hidden_sizes=hidden_sizes,
)
# Move model to specified device
self.model = self.model.to(device)
# Set optimizer
optimizer_name: str = optimizer.lower()
if optimizer_name == "adam":
self.optimizer: optim.Optimizer = optim.Adam(
self.model.parameters(),
lr=lr,
weight_decay=weight_decay
)
elif optimizer_name == "sgd":
self.optimizer = optim.SGD(
self.model.parameters(),
lr=lr,
weight_decay=weight_decay
)
else:
raise NotImplementedError(f"optimizer {optimizer} is not supported!")
# Set learning rate scheduler
self.scheduler: optim.lr_scheduler.ReduceLROnPlateau = optim.lr_scheduler.ReduceLROnPlateau(
self.optimizer,
mode="min",
factor=0.5,
patience=10,
threshold=0.0001,
threshold_mode="rel",
cooldown=0,
min_lr=0.00001,
eps=1e-08,
)
def fit(
self,
dataset: AlphaDataset,
evaluation_results: dict | None = None,
) -> None:
"""
Train with periodic validation and early stopping, and append segment losses to evaluation_results when it is provided.
"""
# Initialize a new dictionary if evaluation_results is None
if evaluation_results is None:
evaluation_results = {}
# Dictionary to store training and validation data
train_valid_data: dict[str, dict] = defaultdict(dict)
# Process training and validation sets separately
segment: Segment
for segment in [Segment.TRAIN, Segment.VALID]:
# Get learning data and sort by time and trading code
df: pl.DataFrame = dataset.fetch_learn(segment)
df = df.sort(["datetime", "vt_symbol"])
# Extract features and labels
features: np.ndarray = df.select(df.columns[2: -1]).to_numpy()
labels: np.ndarray = np.array(df["label"])
# Store feature and label data
train_valid_data["x"][segment] = torch.from_numpy(features).float().to(self.device)
train_valid_data["y"][segment] = torch.from_numpy(labels).float().to(self.device)
# Initialize evaluation results list
evaluation_results[segment] = []
# Get feature names
df = dataset.fetch_learn(Segment.TRAIN)
self.feature_names = df.columns[2:-1]
# Initialize training state
early_stop_count: int = 0 # Number of steps without performance improvement
train_loss: float = 0 # Current training loss
best_valid_score: float = np.inf # Best validation loss
best_params: dict[str, torch.Tensor] | None = None # Best model parameters
train_samples: int = train_valid_data["y"][Segment.TRAIN].shape[0]
# Iterate through training steps
step: int
for step in range(1, self.n_epochs + 1):
# Check if early stopping condition is met
if early_stop_count >= self.early_stop_rounds:
logger.info("达到早停条件,训练结束")
break
# Train one batch
batch_loss: float = self._train_step(train_valid_data, train_samples)
train_loss += batch_loss
# Periodically evaluate the model
if step % self.eval_steps == 0 or step == self.n_epochs:
early_stop_count, best_valid_score, best_params = self._evaluate_step(
train_valid_data,
evaluation_results,
step,
train_loss,
early_stop_count,
best_valid_score
)
train_loss = 0
# Mark model as trained
self.fitted = True
# Load best model parameters
if best_params:
self.model.load_state_dict(best_params)
def _train_step(
self,
train_valid_data: dict[str, dict[Segment, torch.Tensor]],
train_samples: int
) -> float:
"""
Run one optimization step on a random training batch and return that batch loss.
"""
batch_loss: AverageMeter = AverageMeter()
self.model.train()
self.optimizer.zero_grad()
# Randomly select batch data
batch_indices: np.ndarray = np.random.choice(train_samples, self.batch_size)
batch_features: torch.Tensor = train_valid_data["x"][Segment.TRAIN][batch_indices]
batch_labels: torch.Tensor = train_valid_data["y"][Segment.TRAIN][batch_indices]
# Forward and backward propagation
predictions: torch.Tensor = self.model(batch_features)
cur_loss: torch.Tensor = self._loss_fn(predictions, batch_labels)
cur_loss.backward()
# Update model parameters
self.optimizer.step()
batch_loss.update(cur_loss.item())
return batch_loss.val
def _evaluate_step(
self,
train_valid_data: dict[str, dict[Segment, torch.Tensor]],
evaluation_results: dict[Segment, list[float]],
step: int,
train_loss: float,
early_stop_count: int,
best_valid_score: float
) -> tuple[int, float, dict[str, torch.Tensor] | None]:
"""
Score validation, reset the early-stop count when loss improves, and return that count, the best loss, and the new state dict if it improved.
"""
early_stop_count += 1
train_loss /= self.eval_steps
# Evaluate model on validation set
with torch.no_grad():
self.model.eval()
data: torch.Tensor = train_valid_data["x"][Segment.VALID]
pred: torch.Tensor = cast(torch.Tensor, self._predict_batch(data, return_cpu=False))
valid_loss: torch.Tensor = self._loss_fn(pred, train_valid_data["y"][Segment.VALID])
loss_val: float = valid_loss.item()
# Record evaluation results
logger.info(f"[Step {step}]: train_loss {train_loss:.6f}, valid_loss {loss_val:.6f}")
evaluation_results[Segment.TRAIN].append(train_loss)
evaluation_results[Segment.VALID].append(loss_val)
# Update best model if validation performance improves
best_params: dict[str, torch.Tensor] | None = None
if loss_val < best_valid_score:
logger.info(f"\t验证集损失从 {best_valid_score:.6f} 降低到 {loss_val:.6f}")
best_valid_score = loss_val
self.best_step = step
early_stop_count = 0
best_params = copy.deepcopy(self.model.state_dict())
# Update learning rate
if self.scheduler is not None:
self.scheduler.step(metrics=valid_loss, epoch=step)
return early_stop_count, best_valid_score, best_params
def _loss_fn(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
"""
Compute mean squared error after flattening predictions and targets.
"""
pred, target = pred.reshape(-1), target.reshape(-1)
loss: torch.Tensor = nn.MSELoss()(pred, target)
return loss
def _predict_batch(self, data: torch.Tensor, return_cpu: bool = True) -> np.ndarray | torch.Tensor:
"""
Predict in batches of 8096 and return a NumPy array, or a tensor when return_cpu is false.
"""
data = data.to(self.device)
predictions: list[torch.Tensor] = []
self.model.eval()
with torch.no_grad():
batch_size: int = 8096
i: int
for i in range(0, len(data), batch_size):
x: torch.Tensor = data[i: i + batch_size]
predictions.append(self.model(x.to(self.device)).detach().reshape(-1))
if return_cpu:
return np.concatenate([pr.cpu().numpy() for pr in predictions])
else:
return torch.cat(predictions, dim=0)
def predict(self, dataset: AlphaDataset, segment: Segment) -> np.ndarray:
"""
Predict one dataset segment, or raise ValueError if the model has not been trained.
"""
if not self.fitted:
raise ValueError("Model has not been trained yet!")
df: pl.DataFrame = dataset.fetch_infer(segment)
df = df.sort(["datetime", "vt_symbol"])
data: np.ndarray = df.select(df.columns[2: -1]).to_numpy()
return cast(np.ndarray, self._predict_batch(torch.Tensor(data)))
def _check_tensor_nan(self, tensor: torch.Tensor, name: str) -> None:
"""
Print a message when the tensor contains any NaN values.
"""
if torch.isnan(tensor).any():
print(f"NaN values detected: {name}")
def detail(self) -> pd.DataFrame | None:
"""
Log the model configuration and return feature importance, or None when the model is not trained.
"""
if not self.fitted:
logger.info("模型尚未训练，无法显示详细信息")
return None
# 显示模型基本信息
logger.info(f"输入特征维度: {self.input_size}")
logger.info(f"隐藏层大小: {self.hidden_sizes}")
# 计算模型总参数量
total_params: int = sum(p.numel() for p in self.model.parameters())
logger.info(f"模型总参数量: {total_params:,}")
# 显示训练状态信息
logger.info(f"训练设备: {self.device}")
logger.info(f"当前学习率: {self.lr}")
logger.info(f"批次大小: {self.batch_size}")
# Calculate feature importance
importance_df: pd.DataFrame = self._calculate_feature_importance()
return importance_df
def _calculate_feature_importance(self) -> pd.DataFrame:
"""
Estimate importance from how much predictions move when each input feature is perturbed.
"""
self.model.eval()
importance_dict: dict[str, float] = {}
test_data: torch.Tensor = torch.randn(1000, self.input_size).to(self.device)
base_pred: torch.Tensor = self.model(test_data).detach()
noise_level: float = 0.1
i: int
feature_name: str
for i, feature_name in enumerate(self.feature_names):
perturbed_data: torch.Tensor = test_data.clone()
perturbed_data[:, i] += torch.randn(1000).to(self.device) * noise_level
with torch.no_grad():
new_pred: torch.Tensor = self.model(perturbed_data)
importance: float = torch.std(torch.abs(new_pred - base_pred)).item()
importance_dict[feature_name] = importance
df: pd.DataFrame = pd.DataFrame({
"Feature": list(importance_dict.keys()),
"Importance": list(importance_dict.values())
})
df = df.sort_values("Importance", ascending=False)
df = df.set_index("Feature")
return df
class AverageMeter:
"""
Track the latest value together with its running sum, count, and average.
"""
def __init__(self) -> None:
"""
Start the meter with every statistic set to zero.
"""
self.reset()
def reset(self) -> None:
"""
Clear the current value, sum, count, and average.
"""
self.val: float = 0
self.avg: float = 0
self.sum: float = 0
self.count: int = 0
def update(self, val: float, n: int = 1) -> None:
"""
Record a value weighted by n and refresh the running average.
"""
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
class MlpNetwork(nn.Module):
"""
Multilayer perceptron of dropout, linear, batch-norm, and activation layers.
"""
def __init__(
self,
input_size: int,
output_size: int = 1,
hidden_sizes: tuple[int] = (256,),
activation: str = "LeakyReLU"
) -> None:
"""
Build hidden layers of linear, batch norm, and LeakyReLU or SiLU, with 0.05 dropout at the input and before the output. Any other activation name raises ValueError.
"""
super().__init__()
# Build network layers
layers: list[nn.Module] = []
layer_sizes: list[int] = [input_size] + list(hidden_sizes)
# Input layer Dropout
layers.append(nn.Dropout(0.05))
# Build hidden layers
in_size: int
out_size: int
for in_size, out_size in zip(layer_sizes[:-1], layer_sizes[1:], strict=False):
# Add a neural network block: linear layer + batch normalization + activation function
layers.extend([
nn.Linear(in_size, out_size),
nn.BatchNorm1d(out_size),
self._get_activation(activation)
])
# Output layer
layers.extend([
nn.Dropout(0.05),
nn.Linear(hidden_sizes[-1], output_size)
])
# Combine all layers into a sequence
self.network: nn.ModuleList = nn.ModuleList(layers)
# Initialize network weights
self._initialize_weights()
def _get_activation(self, name: str) -> nn.Module:
"""
Return LeakyReLU with negative slope 0.1 or SiLU, or raise ValueError for any other name.
"""
if name == "LeakyReLU":
return nn.LeakyReLU(negative_slope=0.1)
elif name == "SiLU":
return nn.SiLU()
else:
raise ValueError(f"Unsupported activation function type: {name}")
def _initialize_weights(self) -> None:
"""
Apply Kaiming normal initialization to every linear layer using leaky ReLU.
"""
module: nn.Module
for module in self.modules():
if isinstance(module, nn.Linear):
nn.init.kaiming_normal_(
module.weight,
a=0.1, # LeakyReLU negative slope
mode="fan_in", # Scale using input node count
nonlinearity="leaky_relu"
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Pass the input through each network layer in order.
"""
# Pass through all layers in the network sequentially
layer: nn.Module
for layer in self.network:
x = layer(x)
return x