-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathexperiment.py
210 lines (187 loc) · 9.41 KB
/
experiment.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
import torch
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.optim.lr_scheduler import ExponentialLR
from model import *
from data import *
from utils import *
from timeit import default_timer as timer
from datetime import datetime
import pandas as pd
import numpy as np
import shutil
from torchgan.losses import LeastSquaresDiscriminatorLoss, LeastSquaresGeneratorLoss
class Experiment(object):
def __init__(self, option):
self.device = torch.device('cuda' if option.cuda else 'cpu')
self.scale = SCALE_FACTOR
self.image_size = option.image_size
self.save_dir = option.save_dir
self.save_dir.mkdir(parents=True, exist_ok=True)
self.train_dir = self.save_dir / 'train'
self.train_dir.mkdir(exist_ok=True)
self.test_dir = self.save_dir / 'test'
self.test_dir.mkdir(exist_ok=True)
self.history = self.train_dir / 'history.csv'
self.best = self.train_dir / 'best.pth'
self.last_g = self.train_dir / 'generator.pth'
self.last_d = self.train_dir / 'discriminator.pth'
self.logger = get_logger()
self.logger.info('Model initialization')
pretrained = 'assets/autoencoder.pth'
self.pretrained = AutoEncoder().to(self.device)
load_pretrained(self.pretrained, pretrained)
self.generator = Generator().to(self.device)
self.discriminator = Discriminator().to(self.device)
device_ids = [i for i in range(option.ngpu)]
if option.cuda and option.ngpu > 1:
self.generator = nn.DataParallel(self.generator, device_ids)
self.discriminator = nn.DataParallel(self.discriminator, device_ids)
self.criterion = ReconstructionLoss(self.pretrained)
self.g_loss = LeastSquaresGeneratorLoss()
self.d_loss = LeastSquaresDiscriminatorLoss()
self.g_optimizer = optim.Adam(self.generator.parameters(), lr=option.lr)
self.d_optimizer = optim.Adam(self.discriminator.parameters(), lr=option.lr)
self.logger.info(str(self.generator))
self.logger.info(str(self.discriminator))
def train_on_epoch(self, n_epoch, data_loader):
self.generator.train()
self.discriminator.train()
epg_loss = AverageMeter()
epd_loss = AverageMeter()
epg_error = AverageMeter()
batches = len(data_loader)
self.logger.info(f'Epoch[{n_epoch}] - {datetime.now()}')
for idx, data in enumerate(data_loader):
t_start = timer()
data = [im.to(self.device) for im in data]
inputs, target = data[:-1], data[-1]
prediction = self.generator(inputs)
############################
# (1) Update D network
###########################
self.discriminator.zero_grad()
self.generator.zero_grad()
d_loss = self.d_loss(self.discriminator(torch.cat((target, inputs[-1]), 1)),
self.discriminator(torch.cat((prediction.detach(), inputs[-1]), 1)))
d_loss.backward()
self.d_optimizer.step()
epd_loss.update(d_loss.item())
############################
# (2) Update G network
###########################
a_loss = (self.criterion(prediction, target) + 5e-3 *
self.g_loss(self.discriminator(torch.cat((prediction, inputs[-1]), 1))))
a_loss.backward()
self.g_optimizer.step()
epg_loss.update(a_loss.item())
mse = F.mse_loss(prediction.detach(), target).item()
epg_error.update(mse)
t_end = timer()
self.logger.info(f'Epoch[{n_epoch} {idx}/{batches}] - '
f'A-Loss: {a_loss.item():.6f} - '
f'D-Loss: {d_loss.item():.6f} - '
f'MSE: {mse:.6f} - '
f'Time: {t_end - t_start}s')
self.logger.info(f'Epoch[{n_epoch}] - {datetime.now()}')
# 记录Checkpoint
save_checkpoint(self.generator, self.g_optimizer, self.last_g)
save_checkpoint(self.discriminator, self.d_optimizer, self.last_d)
return epg_loss.avg, epd_loss.avg, epg_error.avg
@torch.no_grad()
def test_on_epoch(self, data_loader):
self.generator.eval()
self.discriminator.eval()
epoch_error = AverageMeter()
for data in data_loader:
data = [im.to(self.device) for im in data]
inputs, target = data[:-1], data[-1]
prediction = F.relu(self.generator(inputs), True)
error = F.mse_loss(prediction, target).item()
epoch_error.update(error)
return epoch_error.avg
def train(self, train_dir, val_dir, patch_stride, batch_size,
epochs=30, num_workers=0, resume=True):
last_epoch = -1
if resume and self.history.exists():
df = pd.read_csv(self.history)
last_epoch = int(df.iloc[-1]['epoch'])
load_checkpoint(self.last_g, self.generator, optimizer=self.g_optimizer)
load_checkpoint(self.last_d, self.discriminator, optimizer=self.d_optimizer)
start_epoch = last_epoch + 1
least_error = float('inf')
# 加载数据
self.logger.info('Loading data...')
train_set = PatchSet(train_dir, self.image_size, PATCH_SIZE, patch_stride, mode=Mode.TRAINING)
val_set = PatchSet(val_dir, self.image_size, PATCH_SIZE, patch_stride, mode=Mode.VALIDATION)
train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True,
num_workers=num_workers, drop_last=True)
val_loader = DataLoader(val_set, batch_size=batch_size,
num_workers=num_workers)
self.logger.info('Training...')
g_scheduler = ExponentialLR(self.g_optimizer, 0.99)
d_scheduler = ExponentialLR(self.d_optimizer, 0.99)
for epoch in range(start_epoch, epochs + start_epoch):
self.logger.info(f"Learning rate for Generator: {self.g_optimizer.param_groups[0]['lr']}")
self.logger.info(f"Learning rate for Discriminator: {self.d_optimizer.param_groups[0]['lr']}")
train_g_loss, train_d_loss, train_g_error = self.train_on_epoch(epoch, train_loader)
val_error = self.test_on_epoch(val_loader)
csv_header = ['epoch', 'train_g_loss', 'train_d_loss', 'train_error', 'val_error']
csv_values = [epoch, train_g_loss, train_d_loss, train_g_error, val_error]
log_csv(self.history, csv_values, header=csv_header)
g_scheduler.step()
d_scheduler.step()
if val_error < least_error:
least_error = val_error
shutil.copy(str(self.last_g), str(self.best))
@torch.no_grad()
def test(self, test_dir, patch_size, num_workers=0):
load_checkpoint(self.best, self.generator)
self.generator.eval()
patch_size = make_tuple(patch_size)
self.logger.info('Predicting...')
# 记录测试文件夹中的文件路径，用于最后投影信息的匹配
image_dirs = [p for p in test_dir.glob('*') if p.is_dir()]
image_paths = [get_pair_path(d, Mode.PREDICTION) for d in image_dirs]
# 在预测阶段，对图像进行切块的时候必须刚好裁切完全，这样才能在预测结束后进行完整的拼接
assert self.image_size[0] % patch_size[0] == 0
assert self.image_size[1] % patch_size[1] == 0
rows = int(self.image_size[1] / patch_size[1])
cols = int(self.image_size[0] / patch_size[0])
n_blocks = rows * cols # 一张图像中的分块数目
test_set = PatchSet(test_dir, self.image_size, patch_size, mode=Mode.PREDICTION)
test_loader = DataLoader(test_set, batch_size=1, num_workers=num_workers)
scale_factor = 10000
im_count = 0
patches = []
for data in test_loader:
inputs = [im.to(self.device) for im in data]
name = image_paths[im_count][-1].name
if len(patches) == 0:
t_start = timer()
self.logger.info(f'Predict on image {name}')
# 分块进行预测（每次进入深度网络的都是影像中的一块）
prediction = F.relu(self.generator(inputs), True)
prediction = prediction.squeeze_().cpu().numpy()
prediction = (prediction * scale_factor).astype(np.int16)
patches.append(prediction)
# 完成一张影像以后进行拼接
if len(patches) == n_blocks:
result = np.empty((NUM_BANDS, *self.image_size), dtype=np.int16)
block_count = 0
for i in range(rows):
row_start = i * patch_size[1]
for j in range(cols):
col_start = j * patch_size[0]
result[:,
col_start: col_start + patch_size[0],
row_start: row_start + patch_size[1],
] = patches[block_count]
block_count += 1
patches.clear()
# 存储预测影像结果
prototype = str(image_paths[im_count][0])
save_array_as_tif(result, self.test_dir / name, prototype=prototype)
im_count += 1
t_end = timer()
self.logger.info(f'Time cost: {t_end - t_start}s')