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
| import logging
import os
from typing import Any, Dict, Optional, Union, Tuple
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from torch.utils.checkpoint import checkpoint # New
from tqdm import tqdm
import numpy as np # New
from sssd.core.model_specs import MASK_FN
from sssd.training.utils import training_loss
from sssd.data.utils import get_dataloader
from sssd.utils.logger import setup_logger
from sssd.utils.utils import find_max_epoch, std_normal # New
from autoFRK import AutoFRK, to_tensor, garbage_cleaner # New
LOGGER = setup_logger()
class DiffusionTrainer:
"""
Train Diffusion Models
Args:
dataloader (DataLoader): The training dataloader.
diffusion_hyperparams (Dict[str, Any]): Hyperparameters for the diffusion process.
net (nn.Module): The neural network model to be trained.
device (torch.device): The device to be used for training.
output_directory (str): Directory to save model checkpoints.
ckpt_iter (Optional[int, str]): The checkpoint iteration to be loaded; 'max' selects the maximum iteration.
n_iters (int): Number of iterations to train.
iters_per_ckpt (int): Number of iterations to save checkpoint.
iters_per_logging (int): Number of iterations to save training log and compute validation loss.
learning_rate (float): Learning rate for training.
only_generate_missing (int): Option to generate missing portions of the signal only.
masking (str): Type of masking strategy: 'mnr' for Missing Not at Random, 'bm' for Blackout Missing, 'rm' for Random Missing.
missing_k (int): K missing time steps for each feature across the sample length.
batch_size (int): Size of each training batch.
logger (Optional[logging.Logger]): Logger object for logging, defaults to None.
"""
def __init__(
self,
data_path: str,
diffusion_hyperparams: Dict[str, Any],
net: nn.Module,
device: Optional[Union[torch.device, str]],
output_directory: str,
ckpt_iter: Union[str, int],
n_iters: int,
iters_per_ckpt: int,
iters_per_logging: int,
learning_rate: float,
only_generate_missing: int,
masking: str,
missing_k: int,
batch_size: int,
enable_spatial_prediction: bool, # New
autoFRK_period: int, # New
location_path: str, # New
logger: Optional[logging.Logger] = None,
) -> None:
loader, ts_mean, ts_std = get_dataloader(
path=data_path,
batch_size=batch_size,
device=device,
)
self.dataloader = loader
self.ts_mean = ts_mean
self.ts_std = ts_std
self.diffusion_hyperparams = diffusion_hyperparams
self.net = nn.DataParallel(net).to(device)
self.device = device
self.output_directory = output_directory
self.ckpt_iter = ckpt_iter
self.n_iters = n_iters
self.iters_per_ckpt = iters_per_ckpt
self.iters_per_logging = iters_per_logging
self.learning_rate = learning_rate
self.only_generate_missing = only_generate_missing
self.masking = masking
self.missing_k = missing_k
self.writer = SummaryWriter(f"{output_directory}/log")
self.batch_size = batch_size
self.optimizer = torch.optim.Adam(self.net.parameters(), lr=self.learning_rate)
self.enable_spatial_prediction = enable_spatial_prediction # New
self.autoFRK_period = autoFRK_period # New
self.location_path = location_path # New
self.real_data = self.dataloader.dataset.tensors[0].to(self.device) * self.ts_std + self.ts_mean # New
self.real_data_shape = self.real_data.shape # New
self.logger = logger or LOGGER
if self.masking not in MASK_FN:
raise KeyError(f"Please enter a correct masking, but got {self.masking}")
def _load_checkpoint(self) -> None:
if self.ckpt_iter == "max":
self.ckpt_iter = find_max_epoch(self.output_directory)
if self.ckpt_iter >= 0:
try:
model_path = os.path.join(
self.output_directory, f"{self.ckpt_iter}.pkl"
)
checkpoint = torch.load(model_path, map_location="cpu")
self.net.load_state_dict(checkpoint["model_state_dict"])
if "optimizer_state_dict" in checkpoint:
self.optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
self.logger.info(
f"Successfully loaded model at iteration {self.ckpt_iter}"
)
except Exception as e:
self.ckpt_iter = -1
self.logger.error(f"No valid checkpoint model found. Error: {e}")
else:
self.ckpt_iter = -1
self.logger.info(
"No valid checkpoint model found, start training from initialization."
)
def _save_model(self, n_iter: int) -> None:
if n_iter > 0 and n_iter % self.iters_per_ckpt == 0:
torch.save(
{
"model_state_dict": self.net.state_dict(),
"optimizer_state_dict": self.optimizer.state_dict(),
},
os.path.join(self.output_directory, f"{n_iter}.pkl"),
)
def _update_mask(self, batch: torch.Tensor) -> torch.Tensor:
transposed_mask = MASK_FN[self.masking](batch[0], self.missing_k)
return (
transposed_mask.permute(1, 0)
.repeat(batch.size()[0], 1, 1)
.to(self.device, dtype=torch.float32)
)
def _sampling_differentiable(
self,
net: torch.nn.Module,
size: Tuple[int, int, int],
diffusion_hyperparams: Dict[str, torch.Tensor],
cond: torch.Tensor,
mask: torch.Tensor,
only_generate_missing: int = 0,
device: Union[torch.device, str] = "cpu",
sampling_mode: str = "ddpm", # ✅ "ddpm" 或 "ddim"
step_skip: Union[int, None] = None # ✅ 若為 None，則自動偵測
) -> torch.Tensor:
_dh = diffusion_hyperparams
T, Alpha, Alpha_bar, Sigma = _dh["T"], _dh["Alpha"], _dh["Alpha_bar"], _dh["Sigma"]
x = torch.randn(size, device=device)
mask = mask.to(device)
cond = cond.to(device)
# ✅ 自動偵測最適 step_skip
if step_skip is None:
if T <= 200:
step_skip = 1 # 小模型：全步精細推論
elif T <= 1000:
step_skip = max(1, T // 200) # 一般模型：約 200 步內完成
else:
step_skip = max(1, T // 400) # 大型模型：控制在 400 步以內
#LOGGER.info(f"T={T}, using step_skip={step_skip}")
# ✅ 建立跳步時間表
timesteps = list(range(T - 1, -1, -step_skip))
if timesteps[-1] != 0:
timesteps.append(0)
for t in timesteps:
def net_step(x_inner):
if only_generate_missing == 1:
x_inner = x_inner * (1 - mask).float() + cond * mask.float()
diffusion_steps = (t * torch.ones((size[0], 1), device=device))
epsilon_theta = net((x_inner, cond, mask, diffusion_steps))
# ✅ 根據 sampling_mode 決定取樣公式
if sampling_mode == "ddpm":
# DDPM：隨機性版本
x_next = (x_inner - (1 - Alpha[t]) / torch.sqrt(1 - Alpha_bar[t]) * epsilon_theta) / torch.sqrt(Alpha[t])
if t > 0:
x_next = x_next + Sigma[t] * torch.randn_like(x_next)
elif sampling_mode == "ddim":
# DDIM：確定性版本，支援跳步
if t == 0:
x_next = torch.sqrt(Alpha_bar[t]) * (
(x_inner - torch.sqrt(1 - Alpha_bar[t]) * epsilon_theta)
/ torch.sqrt(Alpha_bar[t])
)
else:
Alpha_bar_prev = Alpha_bar[max(t - step_skip, 0)]
x0_pred = (x_inner - torch.sqrt(1 - Alpha_bar[t]) * epsilon_theta) / torch.sqrt(Alpha_bar[t])
x_next = torch.sqrt(Alpha_bar_prev) * x0_pred + torch.sqrt(1 - Alpha_bar_prev) * epsilon_theta
else:
raise ValueError(f"Unknown sampling_mode: {sampling_mode}")
return x_next
# ✅ checkpoint 保留梯度可逆性
x = checkpoint(net_step, x, use_reentrant=False)
return x
def _sssd_prediction_step(self) -> torch.Tensor:
LOGGER.info(f"Start SSSD prediction step")
all_batches = []
for (batch,) in tqdm(self.dataloader, desc=f"{self.n_iter}-th predicting TS"):
mask = self._update_mask(batch)
batch = batch.permute(0, 2, 1)
# ✅ 新增：自動 step_skip 與 sampling_mode
batch_generated = self._sampling_differentiable(
net=self.net,
size=batch.shape,
diffusion_hyperparams=self.diffusion_hyperparams,
cond=batch,
mask=mask,
only_generate_missing=self.only_generate_missing,
device=self.device,
sampling_mode=getattr(self, "sampling_mode", "ddpm"),
step_skip=getattr(self, "step_skip", 10) # 若未指定則自動偵測
)
all_batches.append(batch_generated)
garbage_cleaner()
sssd_prediction = torch.cat(all_batches, dim=0).permute(1, 2, 0)
return sssd_prediction
def _autoFRK_step(self,
sssd_prediction,
) -> torch.Tensor:
# autoFRK
LOGGER.info(f"Start autoFRK inference step")
dtype = sssd_prediction.dtype
device = sssd_prediction.device
# unstandarize
ts_mean = self.ts_mean.permute(2, 1, 0).expand(-1, sssd_prediction.shape[1], -1)
ts_std = self.ts_std.permute(2, 1, 0).expand(-1, sssd_prediction.shape[1], -1)
sssd_prediction = sssd_prediction * ts_std + ts_mean
V, T, N = sssd_prediction.shape
total_steps = V * T
loc = to_tensor(np.load(self.location_path), dtype=dtype, device=device)
autoFRK_result = torch.empty_like(sssd_prediction, dtype=dtype, device=device)
frk = AutoFRK(
logger_level=30,
dtype=dtype,
device=device,
)
# 只在第一次初始化時編譯 AutoFRK
# if not hasattr(self, "_compiled_frk") or self._compiled_frk is None:
# frk = AutoFRK(
# logger_level=30,
# dtype=dtype,
# device=device,
# )
# self._compiled_frk = torch.compile(frk, mode="reduce-overhead", dynamic=True)
# LOGGER.info("Compiled AutoFRK for the first time with dynamic=True.")
# frk = self._compiled_frk
mrts = None
with tqdm(total=total_steps, desc=f"{self.n_iter}-th predicting autoFRK") as pbar:
for variable in range(V):
for time_index in range(T):
data_slice = sssd_prediction[variable, time_index, :]
try:
_ = frk.forward(
data=data_slice,
loc=loc,
G = mrts,
method="fast",
tps_method="rectangular",
requires_grad=True
)
pred = frk.predict()['pred.value']
mrts = frk.obj['G']['MRTS'] if mrts is None else mrts
except torch._C._LinAlgError:
LOGGER.warning(f"Skipped variable={variable}, time={time_index} due to ill-conditioned matrix")
pred = torch.zeros((N, 1), dtype=dtype, device=device, requires_grad=True)
autoFRK_result[variable, time_index, :] = pred.T
pbar.update(1)
autoFRK_result = autoFRK_result.permute(2, 1, 0)
# return
if autoFRK_result.shape != self.real_data_shape:
error_msg = f"Shape mismatch: autoFRK_result {autoFRK_result.shape} != real_data {self.real_data_shape}"
LOGGER.error(error_msg)
raise ValueError(error_msg)
return autoFRK_result
def _train_per_epoch(self) -> torch.Tensor:
# SSSD training
for (batch,) in tqdm(self.dataloader, desc=f"{self.n_iter}-th training TS"):
batch = batch.to(self.device)
mask = self._update_mask(batch)
loss_mask = ~mask.bool()
loss_function=nn.MSELoss()
batch = batch.permute(0, 2, 1)
assert batch.size() == mask.size() == loss_mask.size()
self.optimizer.zero_grad()
loss = training_loss(
model=self.net,
loss_function=loss_function,
training_data=(batch, batch, mask, loss_mask),
diffusion_parameters=self.diffusion_hyperparams,
generate_only_missing=self.only_generate_missing,
device=self.device,
)
loss.backward()
self.optimizer.step()
if self.enable_spatial_prediction and self.n_iter % self.autoFRK_period == 0:
LOGGER.info(f"Iteration {self.n_iter}: Start Spatial Prediction step")
sssd_prediction = self._sssd_prediction_step()
garbage_cleaner()
autoFRK_result = self._autoFRK_step(sssd_prediction=sssd_prediction)
# compute loss
self.optimizer.zero_grad()
loss = loss_function(
autoFRK_result,
self.real_data
)
# update model
loss.backward()
self.optimizer.step()
LOGGER.info(f"Iteration {self.n_iter}: Spatial Prediction step done, loss: {loss.item()}")
return loss
def train(self) -> None:
self._load_checkpoint()
n_iter_start = (
self.ckpt_iter + 2 if self.ckpt_iter == -1 else self.ckpt_iter + 1
)
self.logger.info(f"Start the {n_iter_start} iteration")
for n_iter in range(n_iter_start, self.n_iters + 1):
self.n_iter = n_iter
loss = self._train_per_epoch()
self.writer.add_scalar("Train/Loss", loss.item(), n_iter)
if n_iter % self.iters_per_logging == 0:
self.logger.info(f"Iteration: {n_iter} \tLoss: { loss.item()}")
self._save_model(n_iter)
|
