-
Notifications
You must be signed in to change notification settings - Fork 9
/
train.py
281 lines (248 loc) · 9.63 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
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
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
import numpy as np
import torch
import torch.nn.functional as F
from tqdm import tqdm
from loss import *
from model_cl import *
from datasets import Dataset
from tensorboardX import SummaryWriter
def train_step(train_data, model, opt, loss_fn):
if opt.precision == 2:
data_type = torch.float64
elif opt.precision == 0:
data_type = torch.float16
else:
data_type = torch.float32
model.to(data_type)
# fetch the points, ground truth extrinsic and intrinsic matrices
correspondences, K1, K2 = train_data['correspondences'].to(opt.device, data_type), train_data['K1'].to(opt.device, data_type), \
train_data['K2'].to(opt.device, data_type)
gt_R, gt_t = train_data['gt_R'].to(opt.device, data_type), train_data['gt_t'].to(opt.device, data_type)
gt_E = train_data['gt_E'].to(opt.device, data_type)
gt_F = train_data['gt_F'].to(opt.device, data_type)
im_size1, im_size2 = train_data['im_size1'].to(opt.device, data_type), train_data['im_size2'].to(opt.device, data_type)
ground_truth = gt_F if opt.fmat else gt_E
if opt.tr:
# default the best choice of weight types
if opt.fmat and opt.sampler == 3: # 8PC
prob_type = 1
elif opt.fmat and opt.sampler == 2: # 7PC
prob_type = 1
else: # 5PC
prob_type = 0
else:
prob_type = opt.prob
# collect all the models
Es, weights, _ = model(
correspondences.to(data_type),
K1,
K2,
im_size1,
im_size2,
prob_type,
ground_truth
)
pts1 = correspondences.squeeze(-1)[:, 0:2].transpose(-1, -2)
pts2 = correspondences.squeeze(-1)[:, 2:4].transpose(-1, -2)
loss = []
for idx, f in enumerate(loss_fn):
if opt.w[idx] != 0:
if idx == 1:
# w1, binary classification loss
train_loss = f.forward(
ground_truth.cpu().detach().numpy(),
pts1,
pts2,
weights,
K1.cpu().detach().numpy(),
K2.cpu().detach().numpy(),
im_size1,
im_size2
)
elif idx == 2:
train_loss = f.forward(
Es,
gt_E.cpu().detach().numpy(),
pts1,
pts2,
K1,
K2,
im_size1,
im_size2
)
else:
# w0, pose error
train_loss = f.forward_average(
Es,
pts1,
pts2,
gt_R,
gt_t,
K1,
K2,
im_size1,
im_size2,
svd=False
)
loss.append(opt.w[idx] * train_loss)
return sum(loss), Es
def train(
model,
train_loader,
valid_loader,
opt
):
# the name of the folder we save models, logs
saved_file = create_session_string(
"train",
opt.sampler,
opt.epochs,
opt.fmat,
opt.nfeatures,
opt.snn,
opt.session,
opt.w0,
opt.w1,
opt.w2,
opt.threshold
)
writer = SummaryWriter('results/' + saved_file + '/vision', comment="model_vis")
optimizer = torch.optim.Adam(model.parameters(), lr=opt.learning_rate)
if opt.scheduler:
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=opt.epochs*len(train_loader), eta_min=opt.eta_min)
loss_function = [PoseLoss(opt.fmat), ClassificationLoss(opt.fmat), MatchLoss(opt.fmat)]
valid_loader_iter = iter(valid_loader)
# save the losses to npy file
train_losses = []
valid_losses = []
# start epoch
for epoch in range(opt.epochs):
train_loss_batch = []
# each step
for idx, train_data in enumerate(tqdm(train_loader)):
model.train()
# one step
optimizer.zero_grad()
train_loss, Es = train_step(train_data, model, opt, loss_function)
train_loss.retain_grad()
for i in Es:
i.retain_grad()
# gradient calculation, ready for back propagation
if torch.isnan(train_loss):
print("pls check, there is nan value in loss!", train_loss)
continue
try:
train_loss.backward()
# print("successfully back-propagation", train_loss)
except Exception as e:
print("we have trouble with back-propagation, pls check!", e)
continue
if torch.isnan(train_loss.grad):
print("pls check, there is nan value in the gradient of loss!", train_loss.grad)
continue
for E in Es:
if torch.isnan(E.grad).any():
print("pls check, there is nan value in the gradient of estimated models!", E.grad)
continue
# add gradient clipping after backward to avoid gradient exploding
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=5)
# check if the gradients of the training parameters contain nan values
nans = sum([torch.isnan(param.grad).any() for param in list(model.parameters()) if param.grad is not None])
if nans != 0:
print("parameter gradients includes {} nan values".format(nans))
continue
optimizer.step()
if opt.scheduler:
scheduler.step()
# check check if the training parameters contain nan values
nan_num = sum([torch.isnan(param).any() for param in optimizer.param_groups[0]['params']])
if nan_num != 0:
print("parameters includes {} nan values".format(nan_num))
continue
train_loss_batch.append(train_loss.item())
# store the network every so often
torch.save(model.state_dict(), 'results/' + saved_file + '/model' + str(epoch) + '.net')
writer.add_scalar('train_loss', np.mean(train_loss_batch), global_step=epoch)
train_losses.append(np.mean(train_loss_batch))
print("-------------------- Epoch-{} finished, do validation-----------------------------".format(epoch))
# validation
with torch.no_grad():
model.eval()
valid_loss_batch = []
for i in tqdm(range(len(valid_loader))):
try:
valid_data = next(valid_loader_iter)
except StopIteration:
pass
valid_loss, _ = train_step(valid_data, model, opt, loss_function)
valid_loss_batch.append(valid_loss.item())
writer.add_scalar('valid_loss', np.mean(valid_loss_batch), global_step=epoch)
valid_losses.append(np.mean(valid_loss_batch))
writer.flush()
print('Epoch: {:02d}| Train loss: {:.4f}| Validation loss: {:.4f}'.format(
epoch,
np.mean(train_loss_batch),
np.mean(valid_loss_batch)
), '\n')
np.save('results/' + saved_file + '/' + 'loss_record.npy', (train_losses, valid_losses))
if __name__ == '__main__':
OUT_DIR = 'results/'
# Parse the parameters
parser = create_parser(
description="Generalized Differentiable RANSAC.")
config = parser.parse_args()
# check if gpu device is available
config.device = torch.device('cuda:0' if torch.cuda.is_available() and config.device != 'cpu' else 'cpu')
print(f"Running on {config.device}")
train_model = DeepRansac_CLNet(config).to(config.device)
# use the pretrained model to initialize the weights if provided.
if len(config.model) > 0:
train_model.load_state_dict(torch.load(config.model, map_location = config.device))
else:
train_model.apply(init_weights)
train_model.train()
# normalize the weights for different losses
w0, w1, w2 = config.w0, config.w1, config.w2
w_sum = w0+w1+w2
if w_sum == 0:
# default loss, epipolar error if no weights are given.
config.w = [0., 0., 1.]
else:
config.w = [float(w0 / w_sum), float(w1 / w_sum), float(w2 / w_sum)]
# collect dataset list
if config.batch_mode:
scenes = test_datasets
print("\n=== BATCH MODE: Training on", len(scenes), "datasets. =================")
else:
scenes = [config.datasets]
print(f'Working on {scenes} with scoring {config.scoring}')
train_folders = [config.data_path + '/' + seq + '/train_data/' for seq in scenes]
valid_folders = [config.data_path + '/' + seq + '/valid_data/' for seq in scenes]
train_dataset = Dataset(
train_folders,
nfeatures=config.nfeatures,
fmat=config.fmat
)
valid_dataset = Dataset(
valid_folders,
nfeatures=config.nfeatures,
fmat=config.fmat
)
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=config.batch_size,
num_workers=config.num_workers,
pin_memory=True,
shuffle=True
)
print(f'Loading training data: {len(train_dataset)} image pairs.')
valid_data_loader = torch.utils.data.DataLoader(
valid_dataset,
batch_size=config.batch_size,
num_workers=0,
pin_memory=True,
shuffle=True
)
print(f'Loading validation data: {len(valid_dataset)} image pairs.')
#with torch.autograd.set_detect_anomaly(True):
train(train_model, train_data_loader, valid_data_loader, config)