/
eval_reliability.py
403 lines (357 loc) · 19.3 KB
/
eval_reliability.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
PROJECT_ROOT = "./"
import sys
sys.path.append(PROJECT_ROOT)
import argparse
import os.path as osp
import os
from tqdm import tqdm
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import numpy as np
import torch
import torch.nn as nn
import torch.nn.parallel
import torch.nn.functional as F
import torch.optim
import torch.utils.data
import torchvision.transforms as transforms
import torchvision.datasets as datasets
from tools import save_obj, load_obj, update_lr, AverageValueMeter
from net.resnet import resnetdisentangler
from net.resnetbig import resnetdisentanglerbig
DATA_DICT_PATH = None
DATA_DICT = None
def create_data_dict(dataset, model_name, data_size):
if not osp.exists(f"./data/reliability_{dataset}"):
os.makedirs(f"./data/reliability_{dataset}")
depths = [1, 2, 4, 8, 16, 32]
train_data = ["var_teacher", "var_normal_current", "var_normal_new", "var_normal_remain",
"var_f_current", "var_f_new",
"var_af_current", "var_af_new",
"var_gaf_current", "var_gaf_new",
"var_agaf_current", "var_agaf_new",
"var_normal-agaf_current", "var_normal-agaf_new"]
test_data = ["teacher_acc", "teacher_lbl_loss", "student_acc", "student_lbl_loss"]
data_dict = {
"train": {data: {depth: None for depth in depths} for data in train_data},
"test": {data: {depth: None for depth in depths} for data in test_data}
}
save_obj(data_dict, f"./data/reliability_{dataset}/{model_name}_{data_size}_data.bin")
class Flatten(nn.Module):
def __init__(self):
super(Flatten, self).__init__()
def forward(self, x):
return x.view(x.shape[0], -1)
class Evaluator(object):
def __init__(self):
self.GPU_ID = None
self.block_num = None
self.BATCH_SIZE = None
self.PREVIOUS_LIST = None
self.PATH = {}
def prepare(self):
self._prepare_model()
self._prepare_dataset()
def _prepare_dataset(self):
self.train_loader = None
self.test_loader = None
raise NotImplementedError
def _prepare_model(self):
raise NotImplementedError
def set_eval_mode(self):
self.strong1.eval()
self.strong2.eval()
self.teacher.eval()
self.strong1_outconv.eval()
self.strong2_outconv.eval()
self.teacher_outconv.eval()
self.student.eval()
self.common.eval()
self.conv1.eval()
self.conv2.eval()
self.conv3.eval()
self.g1.eval()
self.g2.eval()
self.g3.eval()
if len(self.PREVIOUS_LIST) != 0:
self.p_student.eval()
self.p_common.eval()
self.p_conv1.eval()
self.p_conv2.eval()
self.p_conv3.eval()
self.p_g1.eval()
self.p_g2.eval()
self.p_g3.eval()
def add_data(self, mode, data_type, value):
DATA_DICT[mode][data_type][self.block_num] = value
print(f" - {mode} {data_type}: {value}")
def eval_train(self):
self.set_eval_mode()
meter_teacher = AverageValueMeter(gpu_id=self.GPU_ID, dim=self.FEATURE_DIM)
meter_normal_current = AverageValueMeter(gpu_id=self.GPU_ID, dim=self.FEATURE_DIM)
meter_normal_new = AverageValueMeter(gpu_id=self.GPU_ID, dim=self.FEATURE_DIM)
meter_normal_remain = AverageValueMeter(gpu_id=self.GPU_ID, dim=self.FEATURE_DIM)
meter_f_current = AverageValueMeter(gpu_id=self.GPU_ID, dim=self.FEATURE_DIM)
meter_f_new = AverageValueMeter(gpu_id=self.GPU_ID, dim=self.FEATURE_DIM)
meter_af_current = AverageValueMeter(gpu_id=self.GPU_ID, dim=self.FEATURE_DIM)
meter_af_new = AverageValueMeter(gpu_id=self.GPU_ID, dim=self.FEATURE_DIM)
meter_gaf_current = AverageValueMeter(gpu_id=self.GPU_ID, dim=self.FEATURE_DIM)
meter_gaf_new = AverageValueMeter(gpu_id=self.GPU_ID, dim=self.FEATURE_DIM)
meter_agaf_current = AverageValueMeter(gpu_id=self.GPU_ID, dim=self.FEATURE_DIM)
meter_agaf_new = AverageValueMeter(gpu_id=self.GPU_ID, dim=self.FEATURE_DIM)
meter_normal_agaf_current = AverageValueMeter(gpu_id=self.GPU_ID, dim=self.FEATURE_DIM)
meter_normal_agaf_new = AverageValueMeter(gpu_id=self.GPU_ID, dim=self.FEATURE_DIM)
with torch.no_grad():
for images, labels in tqdm(self.train_loader, desc="Train set", mininterval=1):
images, labels = images.to(self.GPU_ID), labels.to(self.GPU_ID)
f_teacher = self.teacher(images)
f_student = self.student(images)
f = self.common(images)
af = self.conv1(f)
if len(self.PREVIOUS_LIST) != 0:
p_f_student = self.p_student(images)
p_f = self.p_common(images)
p_af = self.p_conv1(p_f)
else:
p_f_student = 0
p_f = 0
p_af = 0
gaf = gbf = gcf = None
f_reliable = self.common(images)
for _ in range(self.G_USE_LOOP):
gaf = self.g1(self.conv1(f_reliable))
gbf = self.g2(self.conv2(f_reliable))
gcf = self.g3(self.conv3(f_reliable))
f_reliable = (gaf + gbf + gcf) / 3
agaf = self.conv1(f_reliable)
if len(self.PREVIOUS_LIST) != 0:
p_gaf = p_gbf = p_gcf = None
p_f_reliable = self.p_common(images)
for _ in range(self.G_USE_LOOP):
p_gaf = self.p_g1(self.p_conv1(p_f_reliable))
p_gbf = self.p_g2(self.p_conv2(p_f_reliable))
p_gcf = self.p_g3(self.p_conv3(p_f_reliable))
p_f_reliable = (p_gaf + p_gbf + p_gcf) / 3
p_agaf = self.p_conv1(p_f_reliable)
else:
p_gaf = 0
p_agaf = 0
meter_teacher.add(f_teacher)
meter_normal_current.add(f_student)
meter_normal_new.add(f_student - p_f_student)
meter_normal_remain.add(f_teacher - f_student)
meter_f_current.add(f)
meter_f_new.add(f - p_f)
meter_af_current.add(af)
meter_af_new.add(af - p_af)
meter_gaf_current.add(gaf)
meter_gaf_new.add(gaf - p_gaf)
meter_agaf_current.add(agaf)
meter_agaf_new.add(agaf - p_agaf)
meter_normal_agaf_current.add(f_student - agaf)
meter_normal_agaf_new.add((f_student - p_f_student) - (agaf - p_agaf))
self.add_data("train", "var_teacher", meter_teacher.get_var().item())
self.add_data("train", "var_normal_current", meter_normal_current.get_var().item())
self.add_data("train", "var_normal_new", meter_normal_new.get_var().item())
self.add_data("train", "var_normal_remain", meter_normal_remain.get_var().item())
self.add_data("train", "var_f_current", meter_f_current.get_var().item())
self.add_data("train", "var_f_new", meter_f_new.get_var().item())
self.add_data("train", "var_af_current", meter_af_current.get_var().item())
self.add_data("train", "var_af_new", meter_af_new.get_var().item())
self.add_data("train", "var_gaf_current", meter_gaf_current.get_var().item())
self.add_data("train", "var_gaf_new", meter_gaf_new.get_var().item())
self.add_data("train", "var_agaf_current", meter_agaf_current.get_var().item())
self.add_data("train", "var_agaf_new", meter_agaf_new.get_var().item())
self.add_data("train", "var_normal-agaf_current", meter_normal_agaf_current.get_var().item())
self.add_data("train", "var_normal-agaf_new", meter_normal_agaf_new.get_var().item())
def eval_test(self):
self.set_eval_mode()
correct_student = 0
correct_teacher = 0
lbl_loss_student = 0
lbl_loss_teacher = 0
with torch.no_grad():
for images, labels in tqdm(self.test_loader, desc="Test set", mininterval=1):
images, labels = images.to(self.GPU_ID), labels.to(self.GPU_ID)
pred_student = self.teacher_outconv(self.student(images))
correct_student += pred_student.data.max(1)[1].eq(labels.data).sum()
for i in range(labels.shape[0]):
lbl_loss_student += self.label_criterion(pred_student[i].unsqueeze(0), labels[i].unsqueeze(0)).data
pred_teacher = self.teacher_outconv(self.teacher(images))
correct_teacher += pred_teacher.data.max(1)[1].eq(labels.data).sum()
for i in range(labels.shape[0]):
lbl_loss_teacher += self.label_criterion(pred_teacher[i].unsqueeze(0), labels[i].unsqueeze(0)).data
acc_student = 100. * float(correct_student) / len(self.test_loader.dataset)
acc_teacher = 100. * float(correct_teacher) / len(self.test_loader.dataset)
lbl_loss_student = lbl_loss_student.item() / len(self.test_loader.dataset)
lbl_loss_teacher = lbl_loss_teacher.item() / len(self.test_loader.dataset)
self.add_data("test", "teacher_acc", acc_teacher)
self.add_data("test", "teacher_lbl_loss", lbl_loss_teacher)
self.add_data("test", "student_acc", acc_student)
self.add_data("test", "student_lbl_loss", lbl_loss_student)
def run(self):
self.eval_train()
self.eval_test()
class Eval_CIFAR10(Evaluator):
def __init__(self, args):
super(Eval_CIFAR10, self).__init__()
self.GPU_ID = args.gpu_id
self.block_num = args.block_num
self.BATCH_SIZE = args.batch_size
self.PREVIOUS_LIST = args.previous_list
self.MODEL_NAME = args.model_name
self.DATA_SIZE = args.data_size
self.FEATURE_DIM = args.feature_dim
self.MODEL_SIZE = f"{args.model_name}_{args.data_size}"
self.DATA_ROOT = f"./data/CIFAR10"
self.LOOP = args.loop
self.G_USE_LOOP = args.use_loop
self.NORMAL_MODEL_ROOT = osp.join(PROJECT_ROOT, "models/models_cifar10", self.MODEL_NAME, self.MODEL_SIZE, "disentangle")
self.RELIABLE_MODEL_ROOT = osp.join(PROJECT_ROOT, "models/models_cifar10", self.MODEL_NAME, self.MODEL_SIZE, "reliability")
self.G_MODEL_ROOT = osp.join(PROJECT_ROOT, "models/models_cifar10", self.MODEL_NAME, self.MODEL_SIZE, "reliability/g/loop{}".format(args.loop))
self.PATH = {}
self.PATH["strong_1"] = osp.join(PROJECT_ROOT, "pretrained_models/resnet44_cifar10/resnet44_strong_1_model.pth")
self.PATH["strong_2"] = osp.join(PROJECT_ROOT, "pretrained_models/resnet44_cifar10/resnet44_strong_2_model.pth")
self.PATH['teacher'] = osp.join(PROJECT_ROOT, f"pretrained_models/{self.MODEL_NAME}_cifar10/{self.MODEL_SIZE}_model.pth")
self.PATH['student'] = osp.join(self.NORMAL_MODEL_ROOT, "{}_model.pth".format(self.block_num))
self.PATH["reliable_common"] = osp.join(self.RELIABLE_MODEL_ROOT, "{}/current.pth".format(self.block_num))
self.PATH['conv_1'] = osp.join(self.RELIABLE_MODEL_ROOT, "{}/conv1.pth".format(self.block_num))
self.PATH['conv_2'] = osp.join(self.RELIABLE_MODEL_ROOT, "{}/conv2.pth".format(self.block_num))
self.PATH['conv_3'] = osp.join(self.RELIABLE_MODEL_ROOT, "{}/conv3.pth".format(self.block_num))
self.PATH['g_1'] = osp.join(self.G_MODEL_ROOT, "{}/ga.pth".format(self.block_num))
self.PATH['g_2'] = osp.join(self.G_MODEL_ROOT, "{}/gb.pth".format(self.block_num))
self.PATH['g_3'] = osp.join(self.G_MODEL_ROOT, "{}/gc.pth".format(self.block_num))
if len(self.PREVIOUS_LIST) != 0:
self.PATH['p_student'] = osp.join(self.NORMAL_MODEL_ROOT, "{}_model.pth".format(self.PREVIOUS_LIST[-1]))
self.PATH['p_reliable_common'] = osp.join(self.RELIABLE_MODEL_ROOT, "{}/current.pth".format(self.PREVIOUS_LIST[-1]))
self.PATH['p_conv_1'] = osp.join(self.RELIABLE_MODEL_ROOT, "{}/conv1.pth".format(self.PREVIOUS_LIST[-1]))
self.PATH['p_conv_2'] = osp.join(self.RELIABLE_MODEL_ROOT, "{}/conv2.pth".format(self.PREVIOUS_LIST[-1]))
self.PATH['p_conv_3'] = osp.join(self.RELIABLE_MODEL_ROOT, "{}/conv3.pth".format(self.PREVIOUS_LIST[-1]))
self.PATH['p_g_1'] = osp.join(self.G_MODEL_ROOT, "{}/ga.pth".format(self.PREVIOUS_LIST[-1]))
self.PATH['p_g_2'] = osp.join(self.G_MODEL_ROOT, "{}/gb.pth".format(self.PREVIOUS_LIST[-1]))
self.PATH['p_g_3'] = osp.join(self.G_MODEL_ROOT, "{}/gc.pth".format(self.PREVIOUS_LIST[-1]))
def _prepare_dataset(self):
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
self.train_loader = torch.utils.data.DataLoader(
datasets.CIFAR10(self.DATA_ROOT, train=True, download=True, transform=transform),
batch_size=self.BATCH_SIZE,
shuffle=False,
num_workers=1
)
self.test_loader = torch.utils.data.DataLoader(
datasets.CIFAR10(self.DATA_ROOT, train=False, transform=transform),
batch_size=self.BATCH_SIZE,
shuffle=False,
num_workers=1
)
def _prepare_model(self):
from net.resnet import resnet44
self.strong1 = resnet44(disentangle=True).to(self.GPU_ID)
self.strong1.load_state_dict(torch.load(self.PATH["strong_1"]))
self.strong1_outconv = nn.Sequential(
nn.ReLU(),
nn.AvgPool2d(8, 8),
Flatten(),
self.strong1.linear
)
self.strong2 = resnet44(disentangle=True).to(self.GPU_ID)
self.strong2.load_state_dict(torch.load(self.PATH["strong_2"]))
self.strong2_outconv = nn.Sequential(
nn.ReLU(),
nn.AvgPool2d(8, 8),
Flatten(),
self.strong2.linear
)
if self.MODEL_NAME == "resnet8":
from net.resnet import resnet8
self.teacher = resnet8(disentangle=True).to(self.GPU_ID)
elif self.MODEL_NAME == "resnet14":
from net.resnet import resnet14
self.teacher = resnet14(disentangle=True).to(self.GPU_ID)
elif self.MODEL_NAME == "resnet20":
from net.resnet import resnet20
self.teacher = resnet20(disentangle=True).to(self.GPU_ID)
elif self.MODEL_NAME == "resnet32":
from net.resnet import resnet32
self.teacher = resnet32(disentangle=True).to(self.GPU_ID)
elif self.MODEL_NAME == "resnet44":
from net.resnet import resnet44
self.teacher = resnet44(disentangle=True).to(self.GPU_ID)
self.teacher.load_state_dict(torch.load(self.PATH['teacher']))
self.teacher_outconv = nn.Sequential(
nn.ReLU(),
nn.AvgPool2d(8, 8),
Flatten(),
self.teacher.linear
)
self.student = resnetdisentangler(self.block_num).to(self.GPU_ID)
self.student.load_state_dict(torch.load(self.PATH['student']))
self.common = resnetdisentangler(self.block_num).to(self.GPU_ID)
self.common.load_state_dict(torch.load(self.PATH["reliable_common"]))
self.conv1 = nn.Sequential(nn.Conv2d(64, 64, stride=1, kernel_size=1, bias=False)).to(self.GPU_ID)
self.conv1.load_state_dict(torch.load(self.PATH['conv_1']))
self.conv2 = nn.Sequential(nn.Conv2d(64, 64, stride=1, kernel_size=1, bias=False)).to(self.GPU_ID)
self.conv2.load_state_dict(torch.load(self.PATH['conv_2']))
self.conv3 = nn.Sequential(nn.Conv2d(64, 64, stride=1, kernel_size=1, bias=False)).to(self.GPU_ID)
self.conv3.load_state_dict(torch.load(self.PATH['conv_3']))
self.g1 = nn.Sequential(nn.Conv2d(64, 64, stride=1, kernel_size=1, bias=False)).to(self.GPU_ID)
self.g1.load_state_dict(torch.load(self.PATH['g_1']))
self.g2 = nn.Sequential(nn.Conv2d(64, 64, stride=1, kernel_size=1, bias=False)).to(self.GPU_ID)
self.g2.load_state_dict(torch.load(self.PATH['g_2']))
self.g3 = nn.Sequential(nn.Conv2d(64, 64, stride=1, kernel_size=1, bias=False)).to(self.GPU_ID)
self.g3.load_state_dict(torch.load(self.PATH['g_3']))
if len(self.PREVIOUS_LIST) != 0:
self.p_student = resnetdisentangler(self.PREVIOUS_LIST[-1]).to(self.GPU_ID)
self.p_student.load_state_dict(torch.load(self.PATH['p_student']))
self.p_common = resnetdisentangler(self.PREVIOUS_LIST[-1]).to(self.GPU_ID)
self.p_common.load_state_dict(torch.load(self.PATH['p_reliable_common']))
self.p_conv1 = nn.Sequential(nn.Conv2d(64, 64, stride=1, kernel_size=1, bias=False)).to(self.GPU_ID)
self.p_conv1.load_state_dict(torch.load(self.PATH['p_conv_1']))
self.p_conv2 = nn.Sequential(nn.Conv2d(64, 64, stride=1, kernel_size=1, bias=False)).to(self.GPU_ID)
self.p_conv2.load_state_dict(torch.load(self.PATH['p_conv_2']))
self.p_conv3 = nn.Sequential(nn.Conv2d(64, 64, stride=1, kernel_size=1, bias=False)).to(self.GPU_ID)
self.p_conv3.load_state_dict(torch.load(self.PATH['p_conv_3']))
self.p_g1 = nn.Sequential(nn.Conv2d(64, 64, stride=1, kernel_size=1, bias=False)).to(self.GPU_ID)
self.p_g1.load_state_dict(torch.load(self.PATH['p_g_1']))
self.p_g2 = nn.Sequential(nn.Conv2d(64, 64, stride=1, kernel_size=1, bias=False)).to(self.GPU_ID)
self.p_g2.load_state_dict(torch.load(self.PATH['p_g_2']))
self.p_g3 = nn.Sequential(nn.Conv2d(64, 64, stride=1, kernel_size=1, bias=False)).to(self.GPU_ID)
self.p_g3.load_state_dict(torch.load(self.PATH['p_g_3']))
self.distill_criterion = nn.MSELoss()
self.label_criterion = nn.CrossEntropyLoss().to(self.GPU_ID)
if __name__ == '__main__':
train_list = [1, 2, 4, 8, 16, 32]
bs_list = [128, 128, 128, 128, 128, 128]
parser = argparse.ArgumentParser('parameters')
parser.add_argument('--batch-size', type=int, default=128)
parser.add_argument('--gpu-id', type=int, default=1)
parser.add_argument('--block-num', type=int, default=None, help='train with n blocks')
parser.add_argument('--previous-list', type=list, default=None, help='previous models')
parser.add_argument('--loop', type=int, default=10)
parser.add_argument('--use-loop', type=int, default=5)
parser.add_argument('--model-name', type=str, default=None)
parser.add_argument('--data-size', type=int, default=None)
parser.add_argument('--dataset', type=str, default=None)
parser.add_argument('--feature-dim', type=int, default=None)
args = parser.parse_args()
print(args)
create_data_dict(args.dataset, args.model_name, args.data_size)
DATA_DICT_PATH = f"./data/reliability_{args.dataset}/{args.model_name}_{args.data_size}_data.bin"
DATA_DICT = load_obj(DATA_DICT_PATH)
if args.dataset == "cifar10":
args.feature_dim = 8 * 8 * 64
for i in range(0, len(train_list)):
args.block_num = train_list[i]
args.previous_list = train_list[:i]
args.batch_size = bs_list[i]
print()
print(args)
evaluator = Eval_CIFAR10(args)
evaluator.prepare()
evaluator.run()
save_obj(DATA_DICT, DATA_DICT_PATH)
save_obj(DATA_DICT, DATA_DICT_PATH)