-
Notifications
You must be signed in to change notification settings - Fork 0
/
predict.py
429 lines (362 loc) · 17.8 KB
/
predict.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
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
import pandas as pd
import numpy as np
from tqdm import tqdm
# ML相关库
from sklearn.linear_model import LinearRegression
from sklearn.svm import SVR
from sklearn.preprocessing import PolynomialFeatures,StandardScaler
from sklearn.model_selection import KFold,GridSearchCV,train_test_split,RandomizedSearchCV
from sklearn.metrics import mean_squared_error,mean_absolute_error,mean_absolute_percentage_error
# DL相关库
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import TensorDataset,DataLoader
from torch.optim import AdamW
from transformers import TimeSeriesTransformerConfig,TimeSeriesTransformerForPrediction
from transformers import PretrainedConfig
# 时间序列相关库
from gluonts.dataset.common import ListDataset
from gluonts.dataset.multivariate_grouper import MultivariateGrouper
from gluonts.dataset.field_names import FieldName
from gluonts.time_feature import time_features_from_frequency_str,get_lags_for_frequency,TimeFeature
from gluonts.transform import (
AddAgeFeature,AddTimeFeatures,VstackFeatures,
AddObservedValuesIndicator,
Chain,
RenameFields,RemoveFields,SelectFields,
InstanceSplitter,
)
from gluonts.transform.sampler import ExpectedNumInstanceSampler,ValidationSplitSampler,TestSplitSampler
from gluonts.itertools import Cyclic, IterableSlice, PseudoShuffled
from gluonts.torch.util import IterableDataset
import streamlit as st
from pylab import mpl,plt
plt.style.use('seaborn')
class ML_predict:
def __init__(self,df:pd.DataFrame,train_win,tgt_win) -> None:
self.df=df
self.train_win=train_win
self.tgt_win=tgt_win
def gen_src_tgt(self):
'''生成滞后项,np.ndarray'''
src,tgt=[],[]
L=self.df.shape[0]
for i in range(L-self.train_win-self.tgt_win):
src.append(self.df.iloc[i:i+self.train_win].values)
tgt.append(self.df.iloc[i+self.train_win:i+self.train_win+self.tgt_win].values)
# 用于后续绘图
if np.array(src).shape[-1]!=1:
st.error('本网页中线性回归只适用于单变量预测!请删减变量!')
else:
self.src=np.array(src).squeeze(-1)
self.tgt=np.array(tgt).squeeze(-1)
def predict(self,model,test_size,batch_size=None):
self.model_name=str(model).split('(')[0]
xtrain,xtest,ytrain,ytest=train_test_split(self.src,self.tgt,test_size=test_size,shuffle=False)
model.fit(xtrain,ytrain)
pred=model.predict(xtest)
mse,mae,mape=mean_squared_error(ytest,pred),mean_absolute_error(ytest,pred),mean_absolute_percentage_error(ytest,pred)
score={'MSE':round(mse,3),'MAE':round(mae,3),'MAPE':round(mape,3)}
return pred,score
def plot(self,pred):
fig=plt.figure(figsize=(10,5))
plt.plot(self.df,label='true values')
length_test=pred.shape[0]
pred_idx=self.df.index[-length_test:]
plt.plot(pred_idx,pred,label='predictions')
plt.vlines(pred_idx[0],ymin=self.src[:,0].min(),ymax=self.src[:,0].max(),colors='k',linestyles='dashed',alpha=0.3)
plt.vlines(pred_idx[-1],ymin=self.src[:,0].min(),ymax=self.src[:,0].max(),colors='k',linestyles='dashed',alpha=0.3)
# plt.text(pred_idx[0],pred[0],s=f'{pred_idx[0]}')
plt.legend(fontsize=16)
plt.title(f"{self.model_name}",fontsize=20)
# plt.title(f"{self.model_name} MSE:{score['MSE']:.3f} MAE:{score['MAE']:.3f} MAPE:{score['MAPE']:.3f}",fontsize=20)
st.pyplot(fig)
def __call__(self,model,test_size,batch_size=None):
self.gen_src_tgt()
pred,score=self.predict(model,test_size,batch_size)
self.plot(pred)
return score
class DL_predict(ML_predict):
def __init__(self,df: pd.DataFrame,train_win, tgt_win) -> None:
super().__init__(df,train_win, tgt_win)
def gen_src_tgt(self):
'''生成滞后项,np.ndarray'''
src,tgt,tgt_y=[],[],[]
L=self.df.shape[0]
for i in range(L-self.train_win-self.tgt_win):
src.append(self.df.iloc[i : i + self.train_win].values)
tgt.append(self.df.iloc[i + self.train_win - 1 : i + self.train_win + self.tgt_win -1].values)
tgt_y.append(self.df.iloc[i + self.train_win : i + self.train_win + self.tgt_win].values)
self.src=torch.Tensor(np.array(src))
self.tgt=torch.Tensor(np.array(tgt))
self.tgt_y=torch.Tensor(np.array(tgt_y))
def predict(self, model,test_size,batch_size):
self.model_name=str(model).split('(')[0]
# 划分数据集并构造dataset与dataloader
xtrain,xtest,tgttrain,tgttest,ytrain,ytest=train_test_split(self.src,self.tgt,self.tgt_y,test_size=test_size,shuffle=False)
train_ds=TensorDataset(torch.Tensor(xtrain),torch.Tensor(tgttrain),torch.Tensor(ytrain))
train_dl=DataLoader(train_ds,batch_size=batch_size)
# 训练
criterion=nn.MSELoss()
optimizer=torch.optim.Adam(model.parameters(),lr=1)
tgt_mask=self.gen_square_subsequent_mask(self.tgt_win,self.tgt_win)
memory_mask=self.gen_square_subsequent_mask(self.tgt_win,self.train_win)
model.train()
for epoch in range(1):
print(f"epoch: {epoch+1}")
print('-'*100)
pbar=tqdm(train_dl)
for src,tgt,tgt_y in pbar:
tgt_pred=model(src,tgt,tgt_mask=tgt_mask,memory_mask=memory_mask)
loss=criterion(tgt_pred,tgt_y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
pbar.set_postfix_str(f'当前batch的loss为:{loss.item():.3f}')
# 推理并计算metric
model.eval()
with torch.no_grad():
pred=model(xtest,tgttest,tgt_mask=tgt_mask,memory_mask=memory_mask).detach().numpy()
# 多变量时间预测
mse_lis,mae_lis,mape_lis=[],[],[]
for nvar in range(pred.shape[-1]):
mse_lis.append(mean_squared_error(ytest[:,:,nvar],pred[:,:,nvar]))
mae_lis.append(mean_absolute_error(ytest[:,:,nvar],pred[:,:,nvar]))
mape_lis.append(mean_absolute_percentage_error(ytest[:,:,nvar],pred[:,:,nvar]))
score={'MSE':round(np.mean(mse_lis),3),'MAE':round(np.mean(mae_lis),3),'MAPE':round(np.mean(mape_lis),3)}
return pred.squeeze(1),score
def gen_square_subsequent_mask(self,dim1,dim2):
'''
Args:
dim1: int, for both src and tgt masking, this must be target sequence
length
dim2: int, for src masking this must be encoder sequence length (i.e.
the length of the input sequence to the model),
and for tgt masking, this must be target sequence length
Return:
A Tensor of shape [dim1, dim2]
'''
return torch.triu(torch.ones(dim1, dim2) * float('-inf'), diagonal=1)
# return torch.triu(torch.ones(dim1,dim2),diagonal=1).to(torch.bool)#这种计算mask的方式不行
class TST_predict(ML_predict):
def __init__(self, series: pd.Series, train_win, tgt_win,) -> None:
super().__init__(series, train_win, tgt_win)
self.series=series
self.context_length=train_win
self.prediction_length=tgt_win
def gen_src_tgt(self,freq,config,train_dataloader_batch_size=64,test_dataloader_batch_size=16):
'''在🤗框架下生成dataset,并经transformation后返回dataloader'''
target=self.series.to_numpy().T # shape=[num_series,num_steps],Convention: time axis is always the last axis.
start_series=self.series.index
# define the dataset
self.train_ds=ListDataset(
[
{
FieldName.START:start_series[0],
FieldName.TARGET:target
}
for target in target[:,:-self.prediction_length]
],
freq=freq
)
self.test_ds=ListDataset(
[
{
FieldName.START:start_series[0],
FieldName.TARGET:target
}
for target in target
],
freq=freq
)
# group a univariate dataset into a single multivariate time series: [n_vars,time_steps]
n_vars=len(self.train_ds)
train_grouper=MultivariateGrouper(max_target_dim=n_vars)
test_grouper=MultivariateGrouper(max_target_dim=n_vars)
self.multivariate_train_ds=train_grouper(self.train_ds)
self.multivariate_test_ds=test_grouper(self.test_ds)
# define the dataloader
self.train_dataloader=createTrainDataLoader(config,freq,self.multivariate_train_ds,batch_size=16,num_batches_per_epoch=100)
self.test_dataloader=createTestDataLoader(config,freq,self.multivariate_test_ds,batch_size=4)
def predict(self, model, epochs=5,logging_steps=10):
# train
device='cuda' if torch.cuda.is_available() else 'cpu'
model.to(device)
optimizer=AdamW(model.parameters(),lr=6e-4,betas=(0.9,0.95),weight_decay=1e-1)
total_steps=0
loss_history=[]
model.train()
for epoch in range(epochs):
for batch in self.train_dataloader:
if batch['past_values'].shape[-1]==1:#只有一个变量
batch['past_values']=batch['past_values'].squeeze(2)
batch['past_observed_mask']=batch['past_observed_mask'].squeeze(2)
batch['future_values']=batch['future_values'].squeeze(2)
batch['future_observed_mask']=batch['future_observed_mask'].squeeze(2)
output=model(
past_values=batch['past_values'].to(device),
past_time_features=batch['past_time_features'].to(device),
past_observed_mask=batch['past_observed_mask'].to(device),
future_time_features=batch['future_time_features'].to(device),
future_values=batch['future_values'].to(device),
future_observed_mask=batch['future_observed_mask'].to(device),
)
loss=output.loss
loss_history.append(loss.item())
loss.backward()
optimizer.step()
total_steps+=1
if total_steps%logging_steps==0:
st.write(loss.item())
# inference
model.eval()
forecast_lis=[]# (batch_size,1,number of sample paths,prediction length)
for batch in self.test_dataloader:
if batch['past_values'].shape[-1]==1:#只有一个变量
batch['past_values']=batch['past_values'].squeeze(2)
batch['past_observed_mask']=batch['past_observed_mask'].squeeze(2)
output=model.generate(
past_values=batch['past_values'].to(device),
past_time_features=batch['past_time_features'].to(device),
past_observed_mask=batch['past_observed_mask'].to(device),
future_time_features=batch['future_time_features'].to(device),
).sequences
if len(output.shape)==3:#表明是变量个数是1
output=output.unsqueeze(-1)
forecast_lis.append(output.cpu().detach().numpy())
forecast_array=np.vstack(forecast_lis)# after np.vstack, forecast_array'shape becomes (batch_size,number of sample paths,prediction length,n_vars))
return forecast_array,loss_history
def evaluate(self,forecast_array):
mse_lis,mae_lis,mape_lis=[],[],[]
for idx,forecast in enumerate(forecast_array):#forecast: (number of sample paths,prediction length,n_vars))
forecast_mean=np.mean(forecast,axis=0).T#forecast_mean: (n_vars, prediction length)
st.write('forecast_mean',forecast_mean.shape)
gold=self.multivariate_test_ds[idx]['target'][:,-self.prediction_length:]
st.write('gold',gold.shape)
mse_lis.append(mean_squared_error(gold,forecast_mean))
mae_lis.append(mean_absolute_error(gold,forecast_mean))
mape_lis.append(mean_absolute_percentage_error(gold,forecast_mean))
score={'MSE':round(np.mean(mse_lis),3),'MAE':round(np.mean(mae_lis),3),'MAPE':round(np.mean(mape_lis),3)}
return score
def plot(self, freq,mv_index,forecast_array):
fig,axes=plt.subplots()
index=pd.period_range(
start=self.multivariate_test_ds[0]['start'],
periods=len(self.multivariate_test_ds[0]['target'][mv_index]),
freq=freq,
).to_timestamp()
# true values
axes.plot(index[-self.context_length:],self.multivariate_test_ds[0]['target'][mv_index,-self.context_length:],label='true values')
# preditions
forecast_mean=np.mean(forecast_array[0],axis=0)[:,mv_index]
axes.plot(index[-self.prediction_length:],forecast_mean[-self.prediction_length:],label='preditions')
# shadow area
axes.fill_between(
index[-self.prediction_length:],
y1=forecast_mean-forecast_mean.std(),
y2=forecast_mean+forecast_mean.std(),
)
plt.legend()
st.pyplot(fig)
class Informer_predict(TST_predict):
def __init__(self, series: pd.Series, train_win, tgt_win) -> None:
super().__init__(series, train_win, tgt_win)
def createTransformation(config:PretrainedConfig,freq):
return Chain(
[
AddObservedValuesIndicator(
target_field=FieldName.TARGET,
output_field=FieldName.OBSERVED_VALUES,
),
# temporal features serve as positional encodings
AddTimeFeatures(
start_field=FieldName.START,
target_field=FieldName.TARGET,
output_field=FieldName.FEAT_TIME,
time_features=time_features_from_frequency_str(freq),# 返回len(time_features_from_frequency_str(freq))*len(train_ds)的array
pred_length=config.prediction_length,
),
# another temporal feature
AddAgeFeature(
target_field=FieldName.TARGET,
output_field=FieldName.FEAT_AGE,
pred_length=config.prediction_length,
log_scale=True,#age feature grows logarithmically otherwise linearly overtime.
),
# vertically stack all the temporal features into FieldName.FEAT_TIME
VstackFeatures(
output_field=FieldName.FEAT_TIME,
input_fields=[FieldName.FEAT_TIME,FieldName.FEAT_AGE],
h_stack=False,#dim=0 if h_stack=False else dim=1
),
RenameFields(
mapping={
FieldName.OBSERVED_VALUES:'observed_mask',
FieldName.FEAT_TIME:'time_features',
FieldName.TARGET:'values'
}
)
]
)
def createInstanceSplitter(config:PretrainedConfig,mode,train_sampler=None,validation_sampler=None):
assert mode in ['train','validation','test']
instance_sampler={
'train':train_sampler or ExpectedNumInstanceSampler(num_instances=1,min_future=config.prediction_length),
'validation':validation_sampler or ValidationSplitSampler(min_future=config.prediction_length),
'test':TestSplitSampler(),
}[mode]
return InstanceSplitter(
target_field='values',
is_pad_field=FieldName.IS_PAD,
start_field=FieldName.START,
forecast_start_field=FieldName.FORECAST_START,
instance_sampler=instance_sampler,
past_length=config.context_length+max(config.lags_sequence),
future_length=config.prediction_length,
time_series_fields=['time_features','observed_mask']
)
def createTrainDataLoader(config:PretrainedConfig,freq,data,batch_size,num_batches_per_epoch,shuffle_buffer_length=None):
prediction_input_names=[
'past_values',
'past_time_features',
'past_observed_mask',
'future_time_features',
]
if config.num_static_categorical_features>0:
prediction_input_names.append('static_categorical_features')
if config.num_static_real_features>0:
prediction_input_names.append('static_real_features')
training_input_names=prediction_input_names+[
'future_values',
'future_observed_mask',
]
transformation=createTransformation(freq=freq,config=config)
transformed_data=transformation.apply(data,is_train=True)
instance_splitter=createInstanceSplitter(config,'train')+SelectFields(training_input_names)#在所选择的fields上create InstanceSplitter
train_instances=instance_splitter.apply(
Cyclic(transformed_data) if shuffle_buffer_length is None
else PseudoShuffled(Cyclic(transformed_data),shuffle_buffer_length=shuffle_buffer_length)
)
return IterableSlice(
iter(DataLoader(
IterableDataset(train_instances),
batch_size=batch_size,
)),num_batches_per_epoch,#表示一次取num_batches_per_epoch个batches
)
def createTestDataLoader(config,freq,data,batch_size):
prediction_input_names=[
'past_values',
'past_time_features',
'past_observed_mask',
'future_time_features',
]
if config.num_static_categorical_features>0:
prediction_input_names.append('static_categorical_features')
if config.num_static_real_features>0:
prediction_input_names.append('static_real_features')
transformation=createTransformation(freq=freq,config=config)
transformed_data=transformation.apply(data,is_train=False)
instance_splitter=createInstanceSplitter(config,'test')+SelectFields(prediction_input_names)#在所选择的fields上create InstanceSplitter
test_instances=instance_splitter.apply(transformed_data,is_train=False)
return DataLoader(IterableDataset(test_instances),batch_size=batch_size)