-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
200 lines (165 loc) · 8.86 KB
/
main.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
import os
import gin
import numpy as np
import copy
import torch
import torch.nn as nn
import torchvision.models as models
from trainer import KDTrainer, MultiTaskTrainer, SmallTrainer
from models import AutoEncoder, MultiTaskModel
from models import AbnormalNet
from dataset import CustomDatasetFromImages
from dataset import GradedDatasetFromImages
from torch.optim.lr_scheduler import ReduceLROnPlateau
# Hacks for Reproducibility
seed = 3
torch.manual_seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
np.random.seed(seed)
# from cnn_model import MnistCNNModel
@gin.configurable
def run(batch_size, epochs, val_split, num_workers, print_every,
trainval_csv_path, test_csv_path, model_type, tasks, lr, weight_decay,
momentum, dataset_dir):
train_dataset = CustomDatasetFromImages(trainval_csv_path, data_dir = dataset_dir)
# test_dataset = CustomDatasetFromImages(test_csv_path, data_dir = dataset_dir)
val_from_images = GradedDatasetFromImages(test_csv_path, data_dir = dataset_dir)
dset_len = len(val_from_images)
val_size = int(val_split * dset_len)
test_size = int(0.15 * dset_len)
train_size = dset_len - val_size - test_size
train_dataset_small, val_dataset, test_dataset = torch.utils.data.random_split(val_from_images,
[train_size,
val_size,
test_size])
# Load opth labelled data
train_loader_small = torch.utils.data.DataLoader(dataset=train_dataset_small,
batch_size=batch_size,
pin_memory=False,
drop_last=True,
shuffle=True,
num_workers=num_workers)
val_loader = torch.utils.data.DataLoader(dataset=val_dataset,
batch_size=batch_size,
pin_memory=False,
drop_last=True,
shuffle=True,
num_workers=num_workers)
test_loader = torch.utils.data.DataLoader(dataset=test_dataset,
batch_size=batch_size,
pin_memory=False,
drop_last=True,
shuffle=True,
num_workers=num_workers)
# Load unlabelled data
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=2 * batch_size,
pin_memory=False,
drop_last=True,
shuffle=True,
num_workers=num_workers)
if model_type == 'densenet121':
model = models.densenet121(pretrained=False)
elif model_type == 'resnet101':
model = models.resnet101(pretrained=False)
elif model_type == 'resnet50':
model = models.resnet50(pretrained=False)
elif model_type == 'resnet34':
model = models.resnet34(pretrained=False)
elif model_type == 'vgg19':
model = models.vgg19(pretrained=False)
# model = models.googlenet(pretrained=True)
# transfer_model = AutoEncoder(model_type, model = model)
# transfer_model = nn.DataParallel(transfer_model)
# transfer_model.load_state_dict(torch.load('auto_models/0.7/best_model.pt'))
kd_model = AbnormalNet(model_type, model = model)
kd_model = nn.DataParallel(kd_model)
# kd_model.load_state_dict(torch.load('small_models/{:.2f}-resnet/best_model.pt'.format(round(1 - val_split - 0.15, 2))))
model = AbnormalNet(model_type, model=model)
model = nn.DataParallel(model)
# kd_model.module.conv.load_state_dict(transfer_model.module.conv.state_dict())
# model.module.conv = copy.deepcopy(transfer_model.module.conv)
# model.load_state_dict(torch.load('models/2019-09-27 07:17:11.148440/best_model.pt'))
# All Device training
# model.load_state_dict(torch.load('models/0_1_2-2019-09-3007:56:52.568630/best_model.pt'))
# BEST WITH NORMAL MODEL
# model.load_state_dict(torch.load('models/2019-09-26 06:39:37.468635/best_model.pt'))
# model.load_state_dict(torch.load('models/2019-09-2609:17:12.0260540_1_2/best_model.pt'))
# OISCapture
print(kd_model)
print(model)
kd_model = kd_model.to('cuda')
model = model.to('cuda')
criterion = nn.CrossEntropyLoss()
# =============================== PRE-TRAIN KD MODEL on small labelled data ========================
optimizer = torch.optim.SGD(kd_model.parameters(),
weight_decay=weight_decay,
momentum=momentum,
lr=lr,
nesterov=True)
scheduler = ReduceLROnPlateau(optimizer,
factor=0.5,
patience=3,
min_lr=1e-7,
verbose=True)
trainset_percent = (1 - val_split - 0.15)
trainer = SmallTrainer(kd_model, optimizer, scheduler, criterion, epochs,
print_every = print_every, trainset_split = trainset_percent)
trainer.train(train_loader_small, val_loader)
# Load best KD model into model
model.load_state_dict(torch.load(os.path.join(trainer.save_location_dir,'best_model.pt')))
val_loss, total_d_acc, total_f1, total_recall, total_precision, total_cm = trainer.validate(test_loader)
with open(trainer.output_log, 'a+') as out:
print('Test Loss',val_loss,'total_d_acc',total_d_acc, 'F1', total_f1, 'R', total_recall,'P', total_precision, file=out)
print(total_cm, file=out)
# =============================== TRAIN MODEL WITH KD on small labelled data ========================
optimizer = torch.optim.SGD(model.parameters(),
weight_decay=weight_decay,
momentum=momentum,
lr=lr,
nesterov=True)
scheduler = ReduceLROnPlateau(optimizer,
factor=0.5,
patience=3,
min_lr=1e-7,
verbose=True)
trainset_percent = (1 - val_split - 0.15)
model = copy.deepcopy(kd_model)
trainer = KDTrainer(kd_model, model, optimizer, scheduler, criterion, epochs, print_every = print_every, trainset_split = trainset_percent, kd_type='kd_only')
trainer.train(train_loader, val_loader)
# Load best KD model into model
model.load_state_dict(torch.load(os.path.join(trainer.save_location_dir,'best_model.pt')))
val_loss, total_d_acc, total_f1, total_recall, total_precision, total_cm = trainer.validate(test_loader)
with open(trainer.output_log, 'a+') as out:
print('Test Loss',val_loss,'total_d_acc',total_d_acc, 'F1', total_f1, 'R', total_recall,'P', total_precision, file=out)
print(total_cm, file=out)
# =============================== TRAIN WITH KD MODEL on unlabelled ========================
optimizer = torch.optim.SGD(model.parameters(),
weight_decay=weight_decay,
momentum=momentum,
lr = lr,
nesterov=True)
scheduler = ReduceLROnPlateau(optimizer,
factor=0.5,
patience=3,
min_lr=1e-7,
verbose=True)
epochs = 20
print_every = 100
kd_model = copy.deepcopy(model)
trainer = KDTrainer(kd_model, model, optimizer, scheduler, criterion, epochs, print_every=print_every, trainset_split = trainset_percent, kd_type = 'kd_only')
trainer.train(train_loader, val_loader)
val_loss, total_d_acc, total_f1, total_recall, total_precision, total_cm = trainer.validate(test_loader)
with open(trainer.output_log, 'a+') as out:
print('Test Loss',val_loss,'total_d_acc',total_d_acc, 'F1', total_f1, 'R', total_recall,'P', total_precision, file=out)
print(total_cm, file=out)
trainer.test(test_loader)
if __name__ == "__main__":
task_configs = [0.15, 0.25, 0.4, 0.55, 0.7, 0.85]
for i, t in enumerate(task_configs):
print("Running", (1 -t - 0.15))
gin.parse_config_file('config_small.gin')
gin.bind_parameter('run.val_split', t)
run()
gin.clear_config()