-
Notifications
You must be signed in to change notification settings - Fork 19
/
Copy pathmain_tabular_data.py
262 lines (236 loc) · 12.3 KB
/
main_tabular_data.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
'''
'''
import os, sys
import time
sys.path.insert(0, './python-package/')
isMORT = len(sys.argv)>1 and sys.argv[1] == "mort"
if isMORT:
sys.path.insert(1, 'E:/LiteMORT/python-package/')
import litemort
from litemort import *
print(f"litemort={litemort.__version__}")
import numpy as np
import matplotlib.pyplot as plt
#import node_lib
import quantum_forest
import pandas as pd
import pickle
import torch, torch.nn as nn
import torch.nn.functional as F
import argparse
from sklearn.model_selection import KFold
from qhoptim.pyt import QHAdam
from GBDT_test import *
#You should set the path of each dataset!!!
# data_root = "F:/Datasets/"
def cascade_LR(): #意义不大
if config.cascade_LR:
LinearRgressor = quantum_forest.Linear_Regressor({'cascade':"ridge"})
y_New = LinearRgressor.BeforeFit((data.X_train, data.y_train),[(data.X_valid, data.y_valid),(data.X_test, data.y_test)])
YY_train = y_New[0]
YY_valid,YY_test = y_New[1],y_New[2]
else:
YY_train,YY_valid,YY_test = data.y_train, data.y_valid, data.y_test
return YY_train,YY_valid,YY_test
def VisualAfterEpoch(epoch,visual,config,mse):
if visual is None:
if config.plot_train:
clear_output(True)
plt.figure(figsize=[18, 6])
plt.subplot(1, 2, 1)
plt.plot(loss_history)
plt.title('Loss')
plt.grid()
plt.subplot(1, 2, 2)
plt.plot(mse_history)
plt.title('MSE')
plt.grid()
plt.show()
else:
visual.UpdateLoss(title=f"Accuracy on \"{config.dataset}\"",legend=f"{config.experiment}", loss=mse,yLabel="Accuracy")
def QF_test(data,fold_n,config,visual=None,feat_info=None):
YY_train,YY_valid,YY_test = data.y_train, data.y_valid, data.y_test
data.Y_mean,data.Y_std = YY_train.mean(), YY_train.std()
#config.mean,config.std = mean,std
print(f"====== QF_test \ttrain={data.X_train.shape} valid={data.X_valid.shape} YY_train_mean={data.Y_mean:.3f} YY_train_std={data.Y_std:.3f}\n")
in_features = data.X_train.shape[1]
config.in_features = in_features
#config.tree_module = ODST
config.tree_module = quantum_forest.DeTree
if config.QF_fit>0: #sklearn-like style
learner = quantum_forest.QuantumForest(config,data,feat_info=feat_info,visual=visual).fit(data.X_train, YY_train, eval_set=[(data.X_valid,YY_valid)])
trainer,best_mse = learner.trainer,learner.best_score
epoch = config.nMostEpochs
else:
Learners,last_train_prediction=[],0
qForest = quantum_forest.QF_Net(in_features,config, feat_info=feat_info,visual=visual).to(config.device)
Learners.append(qForest)
if False: # trigger data-aware init,作用不明显
with torch.no_grad():
res = qForest(torch.as_tensor(data.X_train[:1000], device=config.device))
#if torch.cuda.device_count() > 1: model = nn.DataParallel(model)
#weight_decay的值需要反复适配 如取1.0e-6 还可以 0.61142-0.58948
optimizer=QHAdam;
optimizer_params = { 'nus':(0.7, 1.0), 'betas':(0.95, 0.998),'lr':config.lr_base,'weight_decay':1.0e-8 }
#一开始收敛快，后面要慢一些
#optimizer = torch.optim.Adam; optimizer_params = {'lr':config.lr_base }
from IPython.display import clear_output
loss_history, mse_history = [], []
best_mse = float('inf')
best_step_mse = 0
early_stopping_rounds = 3000
report_frequency = 1000
#report_frequency = 10
config.eval_batch_size = 512 if config.leaf_output=="distri2CNN" else \
512 if config.path_way=="TREE_map" else 1024
wLearner=Learners[-1]
trainer = quantum_forest.Experiment(
config,data,
model=wLearner, loss_function=F.mse_loss,
experiment_name=config.experiment,
warm_start=False,
Optimizer=optimizer, optimizer_params=optimizer_params,
verbose=True, #True
n_last_checkpoints=5
)
config.trainer = trainer
trainer.SetLearner(wLearner)
print(f"====== trainer.learner={trainer.model}\ntrainer.opt={trainer.opt}"\
f"\n====== config={config.__dict__}")
print(f"====== X_train={data.X_train.shape},YY_train={YY_train.shape}")
print(f"====== |YY_train|={np.linalg.norm(YY_train):.3f},mean={data.Y_mean:.3f} std={data.Y_std:.3f}")
wLearner.AfterEpoch(isBetter=True, epoch=0)
epoch,t0=0,time.time()
for batch in quantum_forest.experiment.iterate_minibatch(data.X_train, YY_train, batch_size=config.batch_size,shuffle=True, epochs=float('inf')):
metrics = trainer.train_on_batch(*batch, device=config.device)
loss_history.append(metrics['loss'])
if trainer.step%10==0:
symbol = "^" if config.cascade_LR else ""
print(f"\r============ {trainer.step}{symbol}\t{metrics['loss']:.5f}\tL1=[{wLearner.reg_L1:.4g}*{config.reg_L1}]"
f"\tL2=[{wLearner.L_gate:.4g}*{config.reg_Gate}]\ttime={time.time()-t0:.2f}\t"
,end="")
if trainer.step % report_frequency == 0:
epoch=epoch+1
if torch.cuda.is_available(): torch.cuda.empty_cache()
mse = trainer.AfterEpoch(epoch,data.X_valid,YY_valid,best_mse)
if mse < best_mse:
best_mse = mse
best_step_mse = trainer.step
trainer.save_checkpoint(tag='best_mse')
mse_history.append(mse)
if config.average_training:
trainer.load_checkpoint() # last
trainer.remove_old_temp_checkpoints()
VisualAfterEpoch(epoch,visual,config,mse)
if False and epoch%10==9: #有bug啊
#YY_valid = YY_valid- prediction
dict_info,train_pred = trainer.evaluate_mse(data.X_train, YY_train, device=config.device, batch_size=config.eval_batch_size)
#last_train_prediction = last_train_prediction+train_pred
mse_train = dict_info["mse"]
YY_train = YY_train-train_pred
mean,std = YY_train.mean(), YY_train.std()
qForest = quantum_forest.QF_Net(in_features,config, feat_info=feat_info,visual=visual).to(config.device)
#Learners.append(qForest)
wLearner=qForest#Learners[-1]
print(f"QF_test::Expand@{epoch} eval_train={mse_train:.2f} YY_train={np.linalg.norm(YY_train)}")
trainer.SetModel(wLearner)
if trainer.step>50000:
break
if trainer.step > best_step_mse + early_stopping_rounds:
print('BREAK. There is no improvment for {} steps'.format(early_stopping_rounds))
print("Best step: ", best_step_mse)
print(f"Best Val MSE: {best_mse:.5f}")
break
if data.X_test is not None:
mse = trainer.AfterEpoch(epoch,data.X_test, YY_test,best_mse,isTest=True)
if False:
if torch.cuda.is_available(): torch.cuda.empty_cache()
trainer.load_checkpoint(tag='best_mse')
t0=time.time()
dict_info,prediction = trainer.evaluate_mse(data.X_test, YY_test, device=config.device, batch_size=config.eval_batch_size)
if config.cascade_LR:
prediction=LinearRgressor.AfterPredict(data.X_test,prediction)
#prediction = prediction*data.accu_scale+data.Y_mu_0
prediction = data.Y_trans(prediction)
mse = ((data.y_test - prediction) ** 2).mean()
#mse = dict_info["mse"]
reg_Gate = dict_info["reg_Gate"]
print(f'====== Best step: {trainer.step} test={data.X_test.shape} ACCU@Test={mse:.5f} \treg_Gate:{reg_Gate:.4g}time={time.time()-t0:.2f}' )
best_mse = mse
trainer.save_checkpoint(tag=f'last_{mse:.6f}')
return best_mse,mse
def Fold_learning(fold_n,data,config,visual):
t0 = time.time()
if config.model=="QForest":
if config.feat_info == "importance":
feat_info = get_feature_info(data,fold_n)
else:
feat_info = None
accu,_ = QF_test(data,fold_n,config,visual,feat_info)
elif config.model=="GBDT" or config.model=="Catboost" or config.model=="XGBoost" or config.model=="LightGBM":
accu,_ = GBDT_test(config,data,fold_n,num_rounds=config.nMostEpochs)
else: #"LinearRegressor"
model = quantum_forest.Linear_Regressor({'cascade':"ridge"})
accu,_ = model.fit((data.X_train, data.y_train),[(data.X_test, data.y_test)])
print(f"\n======\n====== Fold_{fold_n}@{data.name}\t{data.problem()}"\
f"\tACCURACY={accu:.5f},time={time.time() - t0:.2f} ====== \n======\n")
return
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# parser.add_argument('--use-gpu', action='store_true')
parser.add_argument('--data_root',required=True)
parser.add_argument('--dataset', default='CLICK',help="MICROSOFT,YAHOO,YEAR,CLICK,HIGGS,EPSILON")
parser.add_argument('--iterations', default=100000, type=int)
parser.add_argument('--model', default="QForest", help='QForest,GBDT,LinearRegressor')
parser.add_argument('--learning_rate', default="0.001", type=float)
parser.add_argument('--subsample', default="1", type=float)
parser.add_argument('--QF_fit', default="1", type=int)
parser.add_argument('--attention', default="eca_response", type=str)
parser.add_argument('--scale', default="medium",help='small，medium，large', type=str)
args = parser.parse_args()
print(f"===== {args.__dict__}")
dataset = args.dataset
data = quantum_forest.TabularDataset(dataset,data_path=args.data_root, random_state=1337, quantile_transform=True, quantile_noise=1e-3)
#data = quantum_forest.TabularDataset(dataset,data_path=data_root, random_state=1337, quantile_transform=True)
config = quantum_forest.QForest_config(data,0.002) #,feat_info="importance","attention"
random_state = 42
config.device = quantum_forest.OnInitInstance(random_state)
config.model=args.model #"QForest" "GBDT" "LinearRegressor"
if config.model[0]=="Q": config.model="QForest"
config.lr_base = args.learning_rate
config.dataset = args.dataset
config.bagging_fraction = args.subsample
config.nMostEpochs = args.iterations
config.QF_fit = args.QF_fit
config.attention_alg = args.attention
if args.scale == "small":
config.depth, config.batch_size, config.nTree = 4, 256, 256
elif args.scale == "medium":
config.depth, config.batch_size, config.nTree = 5, 512, 1024
elif args.scale == "large":
config.depth, config.batch_size, config.nTree = 5, 512, 2048
if dataset=="YAHOO" or dataset=="MICROSOFT" or dataset=="CLICK" or dataset=="HIGGS" or dataset=="EPSILON":
config,visual = quantum_forest.InitExperiment(config, 0)
data.onFold(0,config,pkl_path=f"{args.data_root}{dataset}/FOLD_Quantile_{config.model}.pickle")
Fold_learning(0,data, config,visual)
else: #"YEAR"
nFold = 5 if dataset != "HIGGS" else 20
folds = KFold(n_splits=nFold, shuffle=True)
index_sets=[]
for fold_n, (train_index, valid_index) in enumerate(folds.split(data.X)):
index_sets.append(valid_index)
for fold_n in range(len(index_sets)):
config, visual = quantum_forest.InitExperiment(config, fold_n)
train_list=[]
for i in range(nFold):
if i==fold_n: #test
continue
elif i==fold_n+1: #valid
valid_index=index_sets[i]
else:
train_list.append(index_sets[i])
train_index=np.concatenate(train_list)
print(f"train={len(train_index)} valid={len(valid_index)} test={len(index_sets[fold_n])}")
data.onFold(fold_n,config,train_index=train_index, valid_index=valid_index,test_index=index_sets[fold_n],pkl_path=f"{args.data_root}{dataset}/FOLD_{fold_n}_{config.model}.pickle")
Fold_learning(fold_n,data,config,visual)
break