-
Notifications
You must be signed in to change notification settings - Fork 29
/
run-training.py
288 lines (243 loc) · 12.8 KB
/
run-training.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
282
283
284
285
286
287
288
import datetime
import itertools
import os
import numpy as np
from path import Path
from tensorboardX import SummaryWriter
from torch.backends import cudnn
from torch.utils.data import DataLoader
from dvmvs.dataset_loader import MVSDataset
from dvmvs.fusionnet.model import *
from dvmvs.losses import LossMeter, update_losses
from dvmvs.train import train
from dvmvs.utils import zip_code, print_number_of_trainable_parameters, calculate_cost_volume_by_warping
class TrainingHyperparameters:
Config.train_subsequence_length = 8
batch_size = 4
learning_rate = 1e-4
momentum = 0.9
beta = 0.999
weight_decay = 0
# loss_type = "Huber"
# loss_type = "L1"
loss_type = "L1-inv"
# loss_type = "L1-rel"
finetune_epochs = 1
use_checkpoint = True
scaling = 0.5
x = np.linspace(0, Config.train_image_width * scaling - 1, num=int(Config.train_image_width * scaling))
y = np.linspace(0, Config.train_image_height * scaling - 1, num=int(Config.train_image_height * scaling))
ones = np.ones(shape=(int(Config.train_image_height * scaling), int(Config.train_image_width * scaling)))
x_grid, y_grid = np.meshgrid(x, y)
warp_grid = np.stack((x_grid, y_grid, ones), axis=-1)
warp_grid = torch.from_numpy(warp_grid).float()
warp_grid = warp_grid.view(-1, 3).t().cuda()
def main():
# set the manual seed for reproducibility
torch.manual_seed(Config.train_seed)
# create the directory for this run of the training
run_directory = os.path.join(Config.train_run_directory, datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
os.mkdir(run_directory)
# zip every code file
zip_code(run_directory)
summary_writer = SummaryWriter(run_directory)
print("=> fetching scenes in '{}'".format(Config.dataset))
train_set = MVSDataset(
root=Config.dataset,
seed=Config.train_seed,
split="TRAINING",
subsequence_length=Config.train_subsequence_length,
scale_rgb=255.0,
mean_rgb=[0.485, 0.456, 0.406],
std_rgb=[0.229, 0.224, 0.225],
geometric_scale_augmentation=True
)
val_set = MVSDataset(
root=Config.dataset,
seed=Config.train_seed,
split="VALIDATION",
subsequence_length=Config.train_subsequence_length,
scale_rgb=255.0,
mean_rgb=[0.485, 0.456, 0.406],
std_rgb=[0.229, 0.224, 0.225]
)
print('{} samples found in {} train scenes'.format(len(train_set), len(train_set.scenes)))
print('{} samples found in {} valid scenes'.format(len(val_set), len(val_set.scenes)))
train_loader = DataLoader(dataset=train_set,
batch_size=TrainingHyperparameters.batch_size,
shuffle=True,
num_workers=Config.train_data_pipeline_workers,
pin_memory=True,
drop_last=True)
val_loader = DataLoader(dataset=val_set,
batch_size=TrainingHyperparameters.batch_size,
shuffle=False,
num_workers=Config.train_data_pipeline_workers,
pin_memory=True,
drop_last=True)
feature_extractor = FeatureExtractor().cuda()
feature_shrinker = FeatureShrinker().cuda()
cost_volume_encoder = CostVolumeEncoder().cuda()
lstm_fusion = LSTMFusion().cuda()
cost_volume_decoder = CostVolumeDecoder().cuda()
model = [feature_extractor, feature_shrinker, cost_volume_encoder, lstm_fusion, cost_volume_decoder]
if TrainingHyperparameters.use_checkpoint:
checkpoints = sorted(Path("weights").files())
for i in range(len(model)):
try:
weights = torch.load(checkpoints[i])
model[i].load_state_dict(weights)
print("Loaded weights for", checkpoints[i])
except Exception as e:
print(e)
print("Skipping...")
cudnn.benchmark = True
best_loss = [np.inf, np.inf, np.inf, np.inf]
# TRAIN LSTM, DECODER
parameters = itertools.chain(lstm_fusion.parameters(),
cost_volume_decoder.parameters())
optimizer = torch.optim.Adam(parameters,
lr=TrainingHyperparameters.learning_rate,
betas=(TrainingHyperparameters.momentum, TrainingHyperparameters.beta),
weight_decay=TrainingHyperparameters.weight_decay)
print_number_of_trainable_parameters(optimizer)
for epoch in range(TrainingHyperparameters.finetune_epochs):
print("\n\nEPOCH:", epoch)
train(train_loader=train_loader,
val_loader=val_loader,
model=model,
optimizer=optimizer,
summary_writer=summary_writer,
epoch=epoch,
best_loss=best_loss,
run_directory=run_directory,
forward_pass_function=forward_pass)
# TRAIN MY PARTS
parameters = itertools.chain(feature_shrinker.parameters(),
cost_volume_encoder.parameters(),
lstm_fusion.parameters(),
cost_volume_decoder.parameters())
optimizer = torch.optim.Adam(parameters,
lr=TrainingHyperparameters.learning_rate,
betas=(TrainingHyperparameters.momentum, TrainingHyperparameters.beta),
weight_decay=TrainingHyperparameters.weight_decay)
print_number_of_trainable_parameters(optimizer)
for epoch in range(TrainingHyperparameters.finetune_epochs, 2 * TrainingHyperparameters.finetune_epochs):
print("\n\nEPOCH:", epoch)
train(train_loader=train_loader,
val_loader=val_loader,
model=model,
optimizer=optimizer,
summary_writer=summary_writer,
epoch=epoch,
best_loss=best_loss,
run_directory=run_directory,
forward_pass_function=forward_pass)
# TRAIN EVERYTHING
parameters = itertools.chain(feature_extractor.parameters(),
feature_shrinker.parameters(),
cost_volume_encoder.parameters(),
lstm_fusion.parameters(),
cost_volume_decoder.parameters())
optimizer = torch.optim.Adam(parameters,
lr=TrainingHyperparameters.learning_rate,
betas=(TrainingHyperparameters.momentum, TrainingHyperparameters.beta),
weight_decay=TrainingHyperparameters.weight_decay)
print_number_of_trainable_parameters(optimizer)
for epoch in range(2 * TrainingHyperparameters.finetune_epochs, Config.train_epochs):
print("\n\nEPOCH:", epoch)
train(train_loader=train_loader,
val_loader=val_loader,
model=model,
optimizer=optimizer,
summary_writer=summary_writer,
epoch=epoch,
best_loss=best_loss,
run_directory=run_directory,
forward_pass_function=forward_pass)
def forward_pass(images, depths, poses, K, model, is_training):
feature_extractor = model[0]
feature_shrinker = model[1]
cost_volume_encoder = model[2]
lstm_fusion = model[3]
cost_volume_decoder = model[4]
full_K = K.clone().cuda()
half_K = K.clone().cuda()
half_K[:, 0:2, :] = half_K[:, 0:2, :] * scaling
lstm_K = K.clone().cuda()
lstm_K[:, 0:2, :] = lstm_K[:, 0:2, :] / 32.0
images_cuda = []
depths_cuda = []
poses_cuda = []
feature_halfs = []
feature_quarters = []
feature_one_eights = []
feature_one_sixteens = []
# Extract image features
for i in range(0, len(images)):
images_cuda.append(images[i].cuda())
depths_cuda.append(depths[i].cuda())
poses_cuda.append(poses[i].cuda())
feature_half, feature_quarter, feature_one_eight, feature_one_sixteen = feature_shrinker(*feature_extractor(images_cuda[i]))
feature_halfs.append(feature_half)
feature_quarters.append(feature_quarter)
feature_one_eights.append(feature_one_eight)
feature_one_sixteens.append(feature_one_sixteen)
optimizer_loss = 0
predictions = None
l1_meter = LossMeter()
huber_meter = LossMeter()
l1_inv_meter = LossMeter()
l1_rel_meter = LossMeter()
batch_size, _, _ = full_K.size()
lstm_state_bottom = None
for i in range(1, len(images_cuda)):
reference_index = i
measurement_index = i - 1
initial_cost_volume = calculate_cost_volume_by_warping(image1=feature_halfs[reference_index],
image2=feature_halfs[measurement_index],
pose1=poses_cuda[reference_index],
pose2=poses_cuda[measurement_index],
K=half_K,
warp_grid=warp_grid,
min_depth=Config.train_min_depth,
max_depth=Config.train_max_depth,
n_depth_levels=Config.train_n_depth_levels,
device=torch.device('cuda'),
dot_product=True)
skip0, skip1, skip2, skip3, bottom = cost_volume_encoder(features_half=feature_halfs[reference_index],
features_quarter=feature_quarters[reference_index],
features_one_eight=feature_one_eights[reference_index],
features_one_sixteen=feature_one_sixteens[reference_index],
cost_volume=initial_cost_volume)
depth_estimation = depths_cuda[reference_index].view(batch_size, 1, Config.train_image_height, Config.train_image_width)
depth_estimation = torch.nn.functional.interpolate(input=depth_estimation,
scale_factor=(1.0 / 32.0),
mode="nearest")
lstm_state_bottom = lstm_fusion(current_encoding=bottom,
current_state=lstm_state_bottom,
previous_pose=poses_cuda[measurement_index],
current_pose=poses_cuda[reference_index],
estimated_current_depth=depth_estimation,
camera_matrix=lstm_K)
depth_full, depth_half, depth_quarter, depth_one_eight, depth_one_sixteen = cost_volume_decoder(images_cuda[reference_index],
skip0,
skip1,
skip2,
skip3,
lstm_state_bottom[0])
weights = [1, 1, 1, 1, 1]
optimizer_loss = optimizer_loss + update_losses(predictions=[depth_one_sixteen, depth_one_eight, depth_quarter, depth_half, depth_full],
weights=weights,
groundtruth=depths_cuda[reference_index],
is_training=is_training,
l1_meter=l1_meter,
l1_inv_meter=l1_inv_meter,
l1_rel_meter=l1_rel_meter,
huber_meter=huber_meter,
loss_type=TrainingHyperparameters.loss_type)
predictions = [depth_quarter, depth_half, depth_full]
predictions_names = ["prediction_quarter", "prediction_half", "prediction_full"]
return l1_meter, huber_meter, l1_inv_meter, l1_rel_meter, optimizer_loss, predictions, predictions_names
if __name__ == '__main__':
main()