-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathexperiment.py
199 lines (176 loc) · 8.94 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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.optim.lr_scheduler import ReduceLROnPlateau
from model import *
from data import PatchSet, get_pair_path, SCALE_FACTOR
import utils
from timeit import default_timer as timer
from datetime import datetime
import numpy as np
import pandas as pd
import shutil
import sys
class Experiment(object):
def __init__(self, option):
self.device = torch.device('cuda' if option.cuda else 'cpu')
self.resolution_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.history = self.train_dir / 'history.csv'
self.test_dir = self.save_dir / 'test'
self.test_dir.mkdir(exist_ok=True)
self.checkpoint = self.train_dir / 'last.pth'
self.best = self.train_dir / 'best.pth'
self.logger = utils.get_logger()
self.logger.info('Model initialization')
self.model = FusionNet().to(self.device)
self.pretrained = Pretrained().to(self.device)
utils.load_pretrained(self.pretrained, option.pretrained)
if option.cuda and option.ngpu > 1:
device_ids = [i for i in range(option.ngpu)]
self.model = nn.DataParallel(self.model, device_ids=device_ids)
self.pretrained = nn.DataParallel(self.pretrained, device_ids=device_ids)
self.criterion = CompoundLoss(self.pretrained)
self.optimizer = optim.Adam(self.model.parameters(), lr=option.lr, weight_decay=1e-6)
self.logger.info(str(self.model))
n_params = sum(p.numel() for p in self.model.parameters() if p.requires_grad)
self.logger.info(f'There are {n_params} trainable parameters.')
def train_on_epoch(self, n_epoch, data_loader):
self.model.train()
epoch_loss = utils.AverageMeter()
epoch_error = utils.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]
self.optimizer.zero_grad()
predictions = self.model(inputs)
loss = (0.5 * (self.criterion(predictions[0], target) +
self.criterion(predictions[1], target))
if len(predictions) == 2 else self.criterion(predictions, target))
epoch_loss.update(loss.item())
loss.backward()
self.optimizer.step()
with torch.no_grad():
score = (0.5 * (F.mse_loss(predictions[0], target) +
F.mse_loss(predictions[1], target))
if len(predictions) == 2 else F.mse_loss(predictions, target))
epoch_error.update(score.item())
t_end = timer()
self.logger.info(f'Epoch[{n_epoch} {idx}/{batches}] - '
f'Loss: {loss.item():.10f} - '
f'MSE: {score.item():.5f} - '
f'Time: {t_end - t_start}s')
self.logger.info(f'Epoch[{n_epoch}] - {datetime.now()}')
return epoch_loss.avg, epoch_error.avg
@torch.no_grad()
def test_on_epoch(self, data_loader):
self.model.eval()
epoch_loss = utils.AverageMeter()
epoch_error = utils.AverageMeter()
for data in data_loader:
data = [im.to(self.device) for im in data]
inputs, target = data[:-1], data[-1]
prediction = self.model(inputs)
loss = self.criterion(prediction, target)
epoch_loss.update(loss.item())
score = F.mse_loss(prediction, target)
epoch_error.update(score.item())
utils.save_checkpoint(self.model, self.optimizer, self.checkpoint)
return epoch_loss.avg, epoch_error.avg
def train(self, train_dir, val_dir, patch_size, patch_stride, batch_size,
train_refs, num_workers=0, epochs=30, resume=True):
self.logger.info('Loading data...')
train_set = PatchSet(train_dir, self.image_size, patch_size, patch_stride,
n_refs=train_refs)
val_set = PatchSet(val_dir, self.image_size, patch_size, n_refs=train_refs)
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)
least_error = sys.maxsize
start_epoch = 0
if resume and self.checkpoint.exists():
utils.load_checkpoint(self.checkpoint, self.model, self.optimizer)
if self.history.exists():
df = pd.read_csv(self.history)
least_error = df['val_error'].min()
start_epoch = int(df.iloc[-1]['epoch']) + 1
self.logger.info('Training...')
scheduler = ReduceLROnPlateau(self.optimizer, mode='min', factor=0.1, patience=5)
for epoch in range(start_epoch, epochs + start_epoch):
for param_group in self.optimizer.param_groups:
self.logger.info(f"Current learning rate: {param_group['lr']}")
train_loss, train_error = self.train_on_epoch(epoch, train_loader)
val_loss, val_error = self.test_on_epoch(val_loader)
csv_header = ['epoch', 'train_loss', 'train_error', 'val_loss', 'val_error']
csv_values = [epoch, train_loss, train_error, val_loss, val_error]
utils.log_csv(self.history, csv_values, header=csv_header)
scheduler.step(val_loss)
if val_error < least_error:
shutil.copy(str(self.checkpoint), str(self.best))
least_error = val_error
@torch.no_grad()
def test(self, test_dir, patch_size, test_refs, num_workers=0):
self.model.eval()
patch_size = utils.make_tuple(patch_size)
utils.load_checkpoint(self.best, model=self.model)
self.logger.info('Testing...')
# 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, test_refs) 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
test_set = PatchSet(test_dir, self.image_size, patch_size, n_refs=test_refs)
test_loader = DataLoader(test_set, batch_size=1, num_workers=num_workers)
scaled_patch_size = tuple(i * self.resolution_scale for i in patch_size)
scaled_image_size = tuple(i * self.resolution_scale for i in self.image_size)
pixel_value_scale = 10000
im_count = 0
patches = []
t_start = datetime.now()
for inputs in test_loader:
# 如果包含了target数据，则去掉最后的target
if len(inputs) % 2 == 0:
del inputs[-1]
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）
inputs = [im.to(self.device) for im in inputs]
prediction = self.model(inputs)
prediction = prediction.cpu().numpy()
patches.append(prediction * pixel_value_scale)
# 完成一张影像以后进行拼接
if len(patches) == n_blocks:
result = np.empty((NUM_BANDS, *scaled_image_size), dtype=np.float32)
block_count = 0
for i in range(rows):
row_start = i * scaled_patch_size[1]
for j in range(cols):
col_start = j * scaled_patch_size[0]
result[:,
col_start: col_start + scaled_patch_size[0],
row_start: row_start + scaled_patch_size[1]
] = patches[block_count]
block_count += 1
patches.clear()
# 存储预测影像结果
result = result.astype(np.int16)
prototype = str(image_paths[im_count][1])
utils.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')