-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
253 lines (186 loc) · 8.84 KB
/
model.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
import pandas as pd
import seaborn as sns
from tqdm import tqdm
import numpy as np
import time
import torch
import torch.nn as nn
import torch.optim as optim
from sklearn import metrics
from sklearn.metrics import accuracy_score
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.utils.class_weight import compute_class_weight
from torch.utils.data import TensorDataset
from torch.utils.data import DataLoader
from abc import ABC,abstractmethod
from sklearn.linear_model import LogisticRegression
torch.manual_seed(1)
class base_superhuman_model:
def __init__(self):
self.logi_params = {
'C': 100,
'penalty': 'l2',
'solver': 'newton-cg',
'max_iter': 1000
}
@abstractmethod
def fit(self,X,Y):
pass
@abstractmethod
def predict_proba(self,X):
pass
def predict(self,X):
return np.round(self.predict_proba(X))
def score(self,X,Y):
return 1 - np.mean(abs(self.predict(X) - Y))
def expected_error(self,X,Y):
proba = self.predict_proba(X)
return np.mean(np.where(Y == 1 , 1 - proba, proba))
class LR_base_superhuman_model(base_superhuman_model):
def __init__(self):
super().__init__()
self.model = LogisticRegression(**self.logi_params)
self.type = "LR"
def fit(self, X_train, Y_train):
return self.model.fit(X_train, Y_train)
def predict_proba(self, X):
return self.model.predict_proba(X)
def get_model_theta(self):
return self.model.coef_[0]
def update_model_theta(self, new_theta):
self.model.coef_ = np.asarray([new_theta]) # update the coefficient of our logistic regression model with the new theta
class NN_base_superhuman_model(base_superhuman_model):
def __init__(self, demo_list, num_of_demos, num_of_features, subdom_constant, alpha, sample_loss):
super().__init__()
self.type = "NN"
self.demo_list = demo_list
self.num_of_demos = num_of_demos
self.num_of_features = num_of_features
self.subdom_constant = subdom_constant
self.alpha = alpha
self.sample_loss = sample_loss
self.theta = 0
def pretrain_classifier(self, data_loader, optimizer):
for x, y in data_loader:
self.model.zero_grad()
#p_y = clf(x)
# self.model.cuda()
loss, pred = self.model(x, self.demo_list, self.num_of_demos, self.num_of_features, self.subdom_constant, self.alpha, self.sample_loss)
#loss = torch.tensor(loss)
loss.backward()
optimizer.step()
return self.model
def fit(self, X_train, Y_train):
train_data = PandasDataSet(X_train, Y_train)
n_features = X_train.shape[1]
train_loader = DataLoader(train_data, batch_size=32, shuffle=False, drop_last=True)
# train_data.train_data.to(torch.device('cuda:0'))
print('# training samples:', len(train_data))
print('# batches:', len(train_loader))
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.model = Classifier(n_features=n_features).to(device)
model_optimizer = optim.Adam(self.model.parameters())
N_CLF_EPOCHS = 2
for epoch in range(N_CLF_EPOCHS):
self.model = self.pretrain_classifier(train_loader, model_optimizer).to(device)
#Additional Info when using cuda
if device.type == 'cuda':
print(torch.cuda.get_device_name(0))
print('Memory Usage:')
print('Allocated:', round(torch.cuda.memory_allocated(0)/1024**3,1), 'GB')
print('Cached: ', round(torch.cuda.memory_reserved(0)/1024**3,1), 'GB')
return self.model
def predict_proba(self, X):
if type(X) is list: # used when X is only an item
data = torch.Tensor(X)
else: # when X is a dataframe of items
X = pd.DataFrame(X)
test_data = PandasDataSet(X)
data = test_data.tensors[0].to('cuda')
#print("test_data: ", test_data)
#print("test_data.tensors[0]: ", test_data.tensors[0])
with torch.no_grad():
loss, pred = self.model(data, self.demo_list, self.num_of_demos, self.num_of_features, self.subdom_constant, self.alpha, self.sample_loss) #clf(test_data.tensors[0])
#print(pred.numpy().ravel())
#acc = accuracy_score(Y, pred.numpy().ravel().round())
#print(acc)
#print("pred: ", pred)
#print("pred.numpy(): ", pred.numpy())
#print("pred.numpy().ravel(): ", pred.numpy().ravel())
return pred.cpu().numpy()#.ravel()
def get_model_theta(self):
return self.theta
def update_model_theta(self, new_theta):
self.theta = np.asarray([new_theta]) # update the coefficient of our logistic regression model with the new theta
class PandasDataSet(TensorDataset):
def __init__(self, *dataframes):
tensors = (self._df_to_tensor(df) for df in dataframes)
super(PandasDataSet, self).__init__(*tensors)
def _df_to_tensor(self, df):
if isinstance(df, pd.Series):
df = df.to_frame('dummy')
return torch.from_numpy(df.values).float().to(torch.device('cuda:0'))
class Classifier(nn.Module):
def __init__(self, n_features, n_hidden=32, p_dropout=0.2):
super(Classifier, self).__init__()
self.network = nn.Sequential(
nn.Linear(n_features, n_hidden),
nn.ReLU(),
nn.Dropout(p_dropout),
nn.Linear(n_hidden, n_hidden),
nn.ReLU(),
nn.Dropout(p_dropout),
nn.Linear(n_hidden, n_hidden),
nn.ReLU(),
nn.Dropout(p_dropout),
nn.Linear(n_hidden, 2),
)
torch.set_default_tensor_type(torch.cuda.FloatTensor)
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('Using device:', self.device)
def forward(self, x, demo_list, num_of_demos, num_of_features, subdom_constant, alpha, sample_loss):
loss = self.subdom_loss(demo_list, num_of_demos, num_of_features, subdom_constant, alpha, sample_loss)
return loss, torch.sigmoid(self.network(x))
def subdom_loss(self, demo_list, num_of_demos, num_of_features, subdom_constant, alpha, sample_loss):
start_time = time.time()
subdom_tensor = torch.empty([num_of_demos, num_of_features], device=self.device)
sample_loss_arr = torch.tensor(sample_loss, requires_grad=True, device=self.device)
for j, x in enumerate(demo_list):
for k in range(num_of_features):
sample_loss = sample_loss_arr[j, k]
demo_loss = torch.tensor(demo_list[j].metric[k], requires_grad=True, device=self.device)
subdom_val = max(torch.tensor(alpha[k], requires_grad=True, device=self.device)*(sample_loss - demo_loss) + 1, 0)
subdom_tensor[j, k] = subdom_val - subdom_constant # subtract constant c to optimize for useful demonstation instead of avoiding from noisy ones
#grad_theta += self.subdom_tensor[j, k] * self.feature_matching(j)
#print("--- %s end of compute_grad_theta ---" % (time.time() - start_time))
subdom_tensor_sum = torch.sum(subdom_tensor)
return subdom_tensor_sum
def pretrain_classifier(clf, data_loader, optimizer, sh_obj):
for x, y, _ in data_loader:
clf.zero_grad()
#p_y = clf(x)
loss, pred = clf(x, sh_obj.demo_list, sh_obj.num_of_demos, sh_obj.num_of_features, sh_obj.subdom_constant, sh_obj.alpha, sh_obj.sample_loss)
#loss = torch.tensor(loss)
loss.backward()
optimizer.step()
return clf
def predict_nn(X_train, y_train, Z_train, X_test, y_test, Z_test, sh_obj):
train_data = PandasDataSet(X_train, y_train, Z_train)
test_data = PandasDataSet(X_test, y_test, Z_test)
n_features = X_train.shape[1]
train_loader = DataLoader(train_data, batch_size=32, shuffle=True, drop_last=True)
print("predict_nn")
print('# training samples:', len(train_data))
print('# batches:', len(train_loader))
clf = Classifier(n_features=n_features, sh_obj = sh_obj)
clf_optimizer = optim.Adam(clf.parameters())
N_CLF_EPOCHS = 2
for epoch in range(N_CLF_EPOCHS):
clf = pretrain_classifier(clf, train_loader, clf_optimizer, clf_criterion, sh_obj)
with torch.no_grad():
loss, pre_clf_test = clf(test_data.tensors[0], sh_obj.demo_list, sh_obj.num_of_demos, sh_obj.num_of_features, sh_obj.subdom_constant, sh_obj.alpha, sh_obj.sample_loss) #clf(test_data.tensors[0])
print(pre_clf_test.numpy().ravel())
acc = accuracy_score(y_test, pre_clf_test.numpy().ravel().round())
print(acc)
return pre_clf_test.numpy().ravel().round()