forked from mims-harvard/Raincoat
-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer_uni.py
302 lines (271 loc) · 14.6 KB
/
trainer_uni.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
import torch
import torch.nn.functional as F
import os
import pandas as pd
import numpy as np
import warnings
import sklearn.exceptions
warnings.filterwarnings("ignore", category=sklearn.exceptions.UndefinedMetricWarning)
warnings.simplefilter("ignore", category=RuntimeWarning)
np.seterr(all="ignore")
import copy
import diptest
from sklearn.cluster import KMeans
from algorithms.TFAC import TFAC
from sklearn.metrics import classification_report, accuracy_score
from dataloader.uni_dataloader import data_generator
from configs.data_model_configs import get_dataset_class
from configs.hparams import get_hparams_class
from utils import fix_randomness, copy_Files, starting_logs, save_checkpoint, _calc_metrics, calculate_risk
from algorithms.algorithms import get_algorithm_class
from models.models import get_backbone_class
from sklearn.metrics import f1_score
from sklearn.mixture import GaussianMixture
torch.backends.cudnn.benchmark = True
np.warnings.filterwarnings('ignore', category=np.VisibleDeprecationWarning)
class cross_domain_trainer(object):
"""
This class contain the main training functions for our AdAtime
"""
def __init__(self, args):
self.da_method = args.da_method # Selected DA Method
self.dataset = args.dataset # Selected Dataset
self.backbone = args.backbone
self.device = torch.device(args.device) # device
self.run_description = args.experiment_description
self.experiment_description = args.experiment_description
self.best_acc = 0
# paths
self.home_path = os.getcwd()
self.save_dir = args.save_dir
self.data_path = os.path.join(args.data_path, self.dataset)
self.create_save_dir()
# Specify runs
self.num_runs = args.num_runs
# get dataset and base model configs
self.dataset_configs, self.hparams_class = self.get_configs()
# to fix dimension of features in classifier and discriminator networks.
self.dataset_configs.final_out_channels = self.dataset_configs.tcn_final_out_channles if args.backbone == "TCN" else self.dataset_configs.final_out_channels
# Specify number of hparams
self.default_hparams = {**self.hparams_class.alg_hparams[self.da_method],
**self.hparams_class.train_params}
def train(self):
run_name = f"{self.run_description}"
self.hparams = self.default_hparams
# Logging
self.exp_log_dir = os.path.join(self.save_dir, self.experiment_description, run_name)
os.makedirs(self.exp_log_dir, exist_ok=True)
copy_Files(self.exp_log_dir) # save a copy of training files:
scenarios = self.dataset_configs.scenarios # return the scenarios given a specific dataset.
self.trg_acc_list = []
df_a = pd.DataFrame(columns=['scenario','run_id','accuracy','f1','H-score'])
df_c = pd.DataFrame(columns=['scenario','run_id','accuracy','f1','H-score'])
for i in scenarios:
src_id = i[0]
trg_id = i[1]
for run_id in range(self.num_runs): # specify number of consecutive runs
# fixing random seed
fix_randomness(run_id)
self.logger, self.scenario_log_dir = starting_logs(self.dataset, self.da_method, self.exp_log_dir,
src_id, trg_id, run_id)
self.fpath = os.path.join(self.home_path, self.scenario_log_dir, 'backbone.pth')
self.cpath = os.path.join(self.home_path, self.scenario_log_dir, 'classifier.pth')
self.best_acc = 0
# Load data
self.load_data(src_id, trg_id)
if self.da_method =='DANCE':
# get algorithm
algorithm_class = get_algorithm_class(self.da_method)
backbone_fe = get_backbone_class(self.backbone)
algorithm = algorithm_class(backbone_fe, self.dataset_configs, self.hparams, self.device, len(self.trg_train_dl.dataset))
else:
algorithm = TFAC(self.dataset_configs, self.hparams, self.device)
algorithm.to(self.device)
self.algorithm = algorithm
tar_uni_label_train, pri_class = self.preprocess_labels(self.src_train_dl, self.trg_train_dl)
tar_uni_label_test, pri_class = self.preprocess_labels(self.src_train_dl, self.trg_test_dl)
size_ltrain, size_ltest = len(tar_uni_label_train),len(tar_uni_label_test)
for epoch in range(1, self.hparams["num_epochs"] + 1):
joint_loaders = enumerate(zip(self.src_train_dl, self.trg_train_dl))
len_dataloader = min(len(self.src_train_dl), len(self.trg_train_dl))
algorithm.train()
for step, ((src_x, src_y, _), (trg_x, _, trg_index)) in joint_loaders:
src_x, src_y, trg_x, trg_index = src_x.float().to(self.device), src_y.long().to(self.device), \
trg_x.float().to(self.device), trg_index.to(self.device)
if self.da_method=='DANCE':
algorithm.update(src_x, src_y, trg_x, trg_index, step, epoch, len_dataloader)
else:
algorithm.update(src_x, src_y, trg_x)
if self.da_method=='DANCE':
acc, f1, H = self.evaluate_dance(size_ltest)
else:
acc, f1, H = self.evaluate_tfac(self.trg_test_dl.dataset.y_data)
log = {'scenario':i,'run_id':run_id,'accuracy':acc,'f1':f1,'H-score':H}
df_a = df_a.append(log, ignore_index=True)
# Step 2: correct
if self.da_method=='TFAC':
print("===== Correct ====")
dis2proto_a = self.calc_distance(size_ltrain, self.trg_train_dl)
dis2proto_a_test = self.calc_distance(size_ltest, self.trg_test_dl)
for epoch in range(1, self.hparams["num_epochs"] + 1):
joint_loaders = enumerate(zip(self.src_train_dl, self.trg_train_dl))
len_dataloader = min(len(self.src_train_dl), len(self.trg_train_dl))
algorithm.train()
for step, ((src_x, src_y, _), (trg_x, _, trg_index)) in joint_loaders:
src_x, src_y, trg_x, trg_index = src_x.float().to(self.device), src_y.long().to(self.device), \
trg_x.float().to(self.device), trg_index.to(self.device)
algorithm.correct(src_x, src_y, trg_x)
acc, f1, H = self.evaluate_tfac(self.trg_test_dl.dataset.y_data)
dis2proto_c = self.calc_distance(size_ltrain, self.trg_train_dl)
dis2proto_c_test = self.calc_distance(size_ltest, self.trg_test_dl)
c_list = self.learn_t(dis2proto_a, dis2proto_c)
print(c_list)
self.trg_true_labels = tar_uni_label_test
acc, f1, H = self.detect_private(dis2proto_a_test, dis2proto_c_test, tar_uni_label_test, c_list)
log = {'scenario':i,'run_id':run_id,'accuracy':acc,'f1':f1,'H-score':H}
df_c = df_c.append(log, ignore_index=True)
self.save_result(df_a,'average_align.csv')
self.save_result(df_c,'average_correct.csv')
def detect_private(self, d1, d2, tar_uni_label, c_list):
diff = np.abs(d2-d1)
for i in range(6):
cat = np.where(self.trg_pred_labels==i)
cc = diff[cat]
if cc.shape[0]>3:
dip, pval = diptest.diptest(diff[cat])
if dip < 0.05:
print("contain private")
# gm = GaussianMixture(n_components=2, random_state=0,max_iter=5000, n_init=50).fit(diff[cat].reshape(-1, 1))
# c = max(gm.means_)
# kmeans = KMeans(n_clusters=2, random_state=0,max_iter=5000, n_init=50, init="random").fit(diff[cat].reshape(-1, 1))
# c = max(kmeans.cluster_centers_)
c = c_list[i]
m1 = np.where(diff>c)
m2 = np.where(self.trg_pred_labels==i)
mask = np.intersect1d(m1, m2)
# print(m1, m2, mask)
self.trg_pred_labels[mask] = -1
accuracy = accuracy_score(tar_uni_label, self.trg_pred_labels)
f1 = f1_score(self.trg_pred_labels, tar_uni_label, pos_label=None, average="macro")
return accuracy*100, f1, self.H_score()
def preprocess_labels(self, source_loader, target_loader):
trg_y= copy.deepcopy(target_loader.dataset.y_data)
src_y = source_loader.dataset.y_data
pri_c = np.setdiff1d(trg_y, src_y)
mask = np.isin(trg_y, pri_c)
trg_y[mask] = -1
return trg_y, pri_c
def learn_t(self,d1,d2):
diff = np.abs(d2-d1)
c_list= []
for i in range(6):
cat = np.where(self.trg_train_dl.dataset.y_data==i)
cc = diff[cat]
if cc.shape[0]>3:
dip, pval = diptest.diptest(diff[cat])
print(i, dip)
if dip < 0.05:
kmeans = KMeans(n_clusters=2, random_state=0,max_iter=5000, n_init=50, init="random").fit(diff[cat].reshape(-1, 1))
c = max(kmeans.cluster_centers_)
else:
c = 1e10
else:
c = 1e10
c_list.append(c)
return c_list
def calc_distance(self, len_y, dataloader):
feature_extractor = self.algorithm.encoder.to(self.device)
classifier = self.algorithm.classifier.to(self.device)
feature_extractor.eval()
classifier.eval()
proto = classifier.logits.weight.data
# norm = proto.norm(p=2, dim=1, keepdim=True)
# proto = proto.div(norm.expand_as(norm))
trg_drift = np.zeros(len_y)
cos = torch.nn.CosineSimilarity(dim=1,eps=1e-6)
with torch.no_grad():
for data, labels, trg_index in dataloader:
data = data.float().to('cuda')
labels = labels.view((-1)).long().to('cuda')
features,_ = feature_extractor(data)
predictions = classifier(features.detach())
pred_label = torch.argmax(predictions, dim=1)
proto_M = torch.vstack([proto[l,:] for l in pred_label])
angle_c = cos(features,proto_M)**2
# dist = (torch.max(predictions,1).values).div(torch.log(angle_c))
trg_drift[trg_index] = angle_c.cpu().numpy()
return trg_drift
def evaluate_dance(self, labels, threshold=1.6):
feature_extractor = self.algorithm.feature_extractor.to(self.device)
classifier = self.algorithm.classifier.to(self.device)
feature_extractor.eval()
classifier.eval()
data = copy.deepcopy(self.trg_test_dl.dataset.x_data).float().to('cuda')
labels = labels.view((-1)).long().to(self.device)
features = feature_extractor(data)
out_t = classifier(features)
out_t = F.softmax(out_t,dim=-1)
entr = -torch.sum(out_t * torch.log(out_t), 1).data.cpu().numpy()
pred = out_t.argmax(dim=1)
# pred_cls = pred.cpu().numpy()
pred = pred.cpu().numpy()
pred_unk = np.where(entr > threshold)
pred[pred_unk[0]] = -1
accuracy = accuracy_score(labels.cpu().numpy(), pred)
f1 = f1_score(pred, labels.cpu().numpy(), pos_label=None, average="macro")
self.trg_pred_labels = pred
self.trg_true_labels = labels.cpu().numpy()
return accuracy*100, f1, self.H_score()
def evaluate_tfac(self, labels):
feature_extractor = self.algorithm.encoder.to(self.device)
classifier = self.algorithm.classifier.to(self.device)
feature_extractor.eval()
classifier.eval()
data = copy.deepcopy(self.trg_test_dl.dataset.x_data).float().to('cuda')
labels = labels.view((-1)).long().to(self.device)
features, _ = feature_extractor(data)
predictions = classifier(features)
pred = predictions.argmax(dim=1)
pred = pred.cpu().numpy()
accuracy = accuracy_score(labels.cpu().numpy(), pred)
f1 = f1_score(pred, labels.cpu().numpy(), pos_label=None, average="macro")
self.trg_pred_labels = pred
self.trg_true_labels = labels.cpu().numpy()
return accuracy*100, f1, self.H_score()
def H_score(self):
class_c = np.where(self.trg_true_labels!=-1)
class_p = np.where(self.trg_true_labels==-1)
label_c, pred_c = self.trg_true_labels[class_c], self.trg_pred_labels[class_c]
label_p, pred_p = self.trg_true_labels[class_p], self.trg_pred_labels[class_p]
acc_c = accuracy_score(label_c, pred_c)
acc_p = accuracy_score(label_p, pred_p)
# print(acc_c, acc_p)
if acc_c ==0 or acc_p==0:
H = 0
else:
H = 2*acc_c * acc_p/(acc_p+acc_c)
return H
def save_result(self, df, name):
mean_acc = df.groupby('scenario', as_index=False, sort=False)['accuracy'].mean()
mean_f1 = df.groupby('scenario', as_index=False, sort=False)['f1'].mean()
mean_H = df.groupby('scenario', as_index=False, sort=False)['H-score'].mean()
std_acc = df.groupby('scenario', as_index=False, sort=False)['accuracy'].std()
std_f1 = df.groupby('scenario', as_index=False, sort=False)['f1'].std()
std_H = df.groupby('scenario', as_index=False, sort=False)['H-score'].std()
result = pd.concat(objs=(iDF.set_index('scenario') for iDF in (mean_acc, mean_f1, mean_H,std_acc,std_f1,std_H)),
axis=1, join='inner').reset_index()
print(result)
path = os.path.join(self.exp_log_dir, name)
result.to_csv(path)
def get_configs(self):
dataset_class = get_dataset_class(self.dataset)
hparams_class = get_hparams_class(self.dataset)
return dataset_class(), hparams_class()
def load_data(self, src_id, trg_id):
self.src_train_dl, self.src_test_dl = data_generator(self.data_path, src_id, self.dataset_configs,
self.hparams)
self.trg_train_dl, self.trg_test_dl = data_generator(self.data_path, trg_id, self.dataset_configs,
self.hparams)
def create_save_dir(self):
if not os.path.exists(self.save_dir):
os.mkdir(self.save_dir)