-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathref_v2_self.py
182 lines (162 loc) · 6.23 KB
/
ref_v2_self.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
import os
import time
import scipy.io as scio
import matplotlib.pyplot as plt
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim.lr_scheduler import MultiStepLR
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm
from loss import TotalVariationLoss
from models import RefV2 as Model
from models import MagForward
from dataset import get_guide
from evaluate import psnr, ssim
start_epoch = 0
total_epoch = 3000
paddig = 3
in_channels = 20
size = (100, 62)
ano_size = (20, size[0] + paddig * 2, size[1] + paddig * 2)
zero_pad = nn.ZeroPad2d(3)
weight_forward = 1
weight_inversion = 0
weight_guide = 1
weight_tv = 1
weight_lambda = 1
tv_power = 2
lambda_type = 1
model_name = 'ref_v2_self_65'
work_dir = f'work_dirs/{model_name}'
load_from = None # 'work_dirs/ref_v1/weights/best.pth'
def train(
load_from=None,
in_path='data/field_data/field_data.mat',
out_path='data/field_data/field_ref_v2_65.mat'):
if torch.cuda.is_available():
device = torch.device('cuda:0')
else:
device = torch.device('cpu')
print(torch.cuda.is_available())
if not os.path.exists(f'./{work_dir}/weights/'):
os.makedirs(f'./{work_dir}/weights/')
inversion = Model(
mid_channels=64,
mid_channels3d=4,
target_size=ano_size,
num_encoder_blocks=(2, 4, 6, 8),
num_decoder_blocks=(8, 6, 4, 2),
num_3d_block=2,
res_scale=1.0).to(device)
forward = MagForward(in_channels, size, (65.66, 65.57, 100)).to(device)
forward.load_state_dict(torch.load(
'models/ckpt/mag_forward_100_62_20_real1.pth'))
forward.eval()
if load_from and os.path.exists(load_from):
inversion.load_state_dict(torch.load(load_from))
print('load inversion model')
def lambda_loss_function(x):
# print(tv_power, lambda_type)
if lambda_type == 1:
return torch.abs(x).mean()
elif lambda_type == 2:
return torch.pow(x, 2).mean()
inversion_loss_function = nn.L1Loss()
guide_loss_function = nn.L1Loss()
forward_loss_function = nn.L1Loss()
# lambda_loss_function = nn.MSELoss()
tv_loss_function = TotalVariationLoss(1, power=tv_power)
optimizer = optim.Adam(inversion.parameters(), lr=1e-4)
scheduler = MultiStepLR(optimizer, [1000, 2000, 2500], gamma=0.5)
ano = scio.loadmat(in_path)['ma'].astype(np.float32)
ano_std = ano.std()
# ano = ano / 400. # 来自于数值范围的估计
ano /= ano_std
ano = torch.from_numpy(ano[np.newaxis, np.newaxis, ...]).to(device)
pad_ano = zero_pad(ano)
g = torch.tensor([[0, 0.3, 0.8, 1, 0.9,
0.7, 0.6, 0.5, 0.4, 0.3,
0.2, 0.1, 0, 0, 0,
0, 0, 0, 0, 0
]]).to(device)
g = g/g.sum(dim=-1, keepdim=True)*g.shape[-1]/2
min_loss = 100
log_path = f'./{work_dir}/weights/log.txt'
writer = SummaryWriter(f"./{work_dir}/weights/log")
for _ in range(start_epoch):
optimizer.step()
scheduler.step()
t0 = time.time()
for i in range(start_epoch, total_epoch):
inversion.train()
loss_forward_sum = 0
loss_inversion_sum = 0
loss_guide_sum = 0
loss_tv_sum = 0
pad_body_ = inversion(pad_ano, g)
body_ = pad_body_[..., paddig:-paddig, paddig:-paddig]
g_ = get_guide(body_)
# ano_ = forward(body_) / 400 # 来自于数值范围的估计
ano_ = forward(body_) / ano_std # 来自于数值范围的估计
optimizer.zero_grad()
# print(ano.shape, ano.min(), ano.max())
# print(ano_gt.shape, ano_gt.min(), ano_gt.max())
# print(body_.shape, body_.min(), body_.max())
# print(ano_.shape, ano_.min(), ano_.max())
# print(g.shape, g.min(), g.max())
# input()
loss_forward = forward_loss_function(ano_, ano)
loss_guide = guide_loss_function(g_, g)
loss_lambda = lambda_loss_function(pad_body_)
loss_tv = tv_loss_function(pad_body_)
if i < 400:
loss = loss_guide*50 + loss_tv * weight_tv + loss_forward
elif i < 2500:
loss = (loss_forward * weight_forward +
loss_lambda * weight_lambda +
loss_guide * weight_guide +
loss_tv * weight_tv)
else:
loss = (loss_forward * weight_forward +
loss_lambda * weight_lambda * 5 +
loss_guide * weight_guide +
loss_tv * weight_tv * 5)
loss.backward()
optimizer.step()
scheduler.step()
if loss_forward < min_loss*0.99 and i > 1500:
min_loss = loss_forward
torch.save(inversion.state_dict(),
f'./{work_dir}/best.pth')
learn_rate = optimizer.state_dict()['param_groups'][0]['lr']
with open(log_path, 'a', encoding='utf-8') as f:
f.write(f'inference {i:04d}\tloss: {loss:.6f}'
f'\tforward {loss_forward:.6f}'
f'\tguide {loss_guide:.6f}'
f'\tlambda {loss_lambda:.6f}'
f'\ttv {loss_tv:.6f}'
f'\tlr {learn_rate:.6f}\n')
if i % 50 == 0:
print(time.time()-t0)
t0 = time.time()
print(f'inference {i:04d}\tloss: {loss:.6f}'
f'\tforward {loss_forward:.6f}'
f'\tguide {loss_guide:.6f}'
f'\tlambda {loss_lambda:.6f}'
f'\ttv {loss_tv:.6f}'
f'\tlr {learn_rate:.6f}')
if i % 200 == 0:
print('body_', body_.shape, body_.min(), body_.max())
print('ano', ano.shape, ano.min(), ano.max(), ano.mean())
print('ano_', ano_.shape, ano_.min(), ano_.max())
print('g_', g_.shape, g_.min(), g_.max())
body_ = body_[0, 0].permute(1, 2, 0).detach().cpu().numpy()
ano_ = ano_[0, 0].detach().cpu().numpy()
g_ = g_[0].detach().cpu().numpy()
scio.savemat(out_path, dict(mag=body_, g=g_, ma=ano_))
plt.imsave(out_path + '.png', ano_, cmap='jet')
if __name__ == '__main__':
train()