-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
185 lines (156 loc) · 7.26 KB
/
train.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
from model.VQAE import VQAE
from model.MARL import MARL
from utils import device, add_noise
from tqdm import tqdm
import torch
from data.FloorPlanLoader import *
import torch.nn.functional as F
import random
import json
USE_MULTISCALE = True
USE_MULTITASK = True
#Reproducability Checks:
random.seed(0) #Python
torch.manual_seed(0) #Torch
np.random.seed(0) #NumPy
#Hyperparameter
batch_size = 128
n_hiddens = 32
n_residual_hiddens = 32
n_residual_layers = 1
embedding_dim = 64
n_embeddings = 218
beta = .25
lr = 3e-3
epochs = 100
noise=False
noise_weight=0.05
img_channel=3 if USE_MULTISCALE else 1
def train_marl(train_loader=None, validation_loader=None,
data_variance=None, val_len=None, year_label_num=None, category_num=None,
get_pretrain=True, use_multi_task=USE_MULTITASK):
vqae = VQAE(n_hiddens, n_residual_hiddens, n_residual_layers,
n_embeddings, embedding_dim,
beta, img_channel).to(device)
if get_pretrain:
vqae.load_state_dict(torch.load("./best_checkpoint/final/55-vqae-0.04753296934928414.pt"))
marl = MARL(vqae, USE_MULTITASK, year_label_num, category_num)
optimizer = torch.optim.Adam(marl.parameters(), lr=lr, amsgrad=False)
train_recon_error = []
train_height_error = []
train_age_error = []
train_usage_error = []
test_recon_error = []
test_height_error = []
test_age_error = []
test_usage_error = []
best_loss = 1e10
for epoch in range(0, epochs):
with tqdm(train_loader, unit="batch") as tepoch:
marl.train()
for data_dict in tepoch:
data = data_dict['image_tensor']
bs = data.shape[0]
data_no_noise = data.to(device)
optimizer.zero_grad()
if noise:
data = add_noise(data_no_noise, noise_weight=noise_weight)
else:
data = data_no_noise
pred = marl(data)
# recon loss
vq_loss, data_recon, perplexity = pred['vqae']
recon_error = F.mse_loss(data_recon, data) / data_variance
train_recon_error.append(recon_error.item())
if USE_MULTITASK:
# height infer
height_pred = pred['height']
height_error = F.mse_loss(height_pred, data_dict['height'].to(device).view(bs,-1))
train_height_error.append(height_error.item())
# age infer
age_pred = pred['age']
labels = data_dict['age_label'].to(device).long()
age_error = F.cross_entropy(age_pred, labels)*0.3
train_age_error.append(age_error.item())
# category infer
category_pred = pred['category']
labels = data_dict['cate_onehot'].to(device)
criterion = torch.nn.BCEWithLogitsLoss()
category_error = criterion(category_pred, labels)*0.7
train_usage_error.append(category_error.item())
loss = (recon_error + vq_loss) + height_error + age_error + category_error
loss.backward()
optimizer.step()
tepoch.set_postfix(recon_error=float((recon_error+ vq_loss).detach().cpu()),
height_error=float(height_error.detach().cpu()),
age_error=float(age_error.detach().cpu()),
category_error=float(category_error.detach().cpu()))
avg_loss = 0
marl.eval()
with torch.no_grad():
for data_dict in validation_loader:
data = data_dict['image_tensor']
bs = data.shape[0]
data = data.to(device)
pred = marl(data)
# recon loss
vq_loss, data_recon, perplexity = pred['vqae']
recon_error = F.mse_loss(data_recon, data) / data_variance
test_recon_error.append(recon_error.item())
if USE_MULTITASK:
# height infer
height_pred = pred['height']
height_error = F.mse_loss(height_pred, data_dict['height'].to(device).view(bs,-1))
test_height_error.append(height_error.item())
# age infer
age_pred = pred['age']
labels = data_dict['age_label'].to(device).long()
age_error = F.cross_entropy(age_pred, labels)
test_age_error.append(age_error.item())
# category infer
category_pred = pred['category']
labels = data_dict['cate_onehot'].to(device)
criterion = torch.nn.BCEWithLogitsLoss()
category_error = criterion(category_pred, labels)
test_usage_error.append(category_error.item())
loss = (recon_error.item() \
+ height_error.item()\
+ age_error.item()\
+ category_error.item()\
) * batch_size
avg_loss += loss / val_len
if avg_loss<best_loss:
best_loss = avg_loss
best_epoch = epoch
torch.save(marl.state_dict(), f"./checkpoint/{best_epoch}-marl-{best_loss}.pt")
torch.save(optimizer.state_dict(), f"./checkpoint/{best_epoch}-adam-{best_loss}.pt")
if USE_MULTITASK:
error = {
'train_recon_error': train_recon_error,
'train_height_error': train_height_error,
'train_age_error': train_age_error,
'train_usage_error': train_usage_error,
'test_recon_error': test_recon_error,
'test_height_error': test_height_error,
'test_age_error': test_age_error,
'test_usage_error': test_usage_error
}
else:
error = {
'train_recon_error': train_recon_error,
'test_recon_error': test_recon_error
}
with open(f"./checkpoint/{best_epoch}-error-{best_loss}.json", 'w', encoding ='utf8') as json_file:
json.dump(error, json_file, ensure_ascii = False)
print(f'Validation Loss: {avg_loss}')
if __name__ == "__main__":
#Load Dataset
floor = FloorPlanDataset(multi_scale=True, root='./data/data_root/data00/', data_config='./data/data_config/', preprocess=True)
data_variance = floor.var
val_len = int(len(floor)/10)
train_set, val_set = torch.utils.data.random_split(floor, [len(floor)-val_len, val_len])
print(f"data shape: {floor[0]['image_tensor'].shape}, dataset size: {len(floor)}, data variance: {data_variance}")
train_loader = torch.utils.data.DataLoader(train_set, batch_size = batch_size, shuffle = True)
validation_loader = torch.utils.data.DataLoader(val_set, batch_size = batch_size, shuffle = False)
train_marl(train_loader, validation_loader, \
floor.var, int(len(floor)/10), floor.age_label_num, floor.category_num)