-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
180 lines (130 loc) · 7.99 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
# Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
from inference import TracksTo4DNet,CustomDataset,get_losses
import argparse
import torch
import torch.optim as optim
import datetime
import json
import argparse
from pathlib import Path
from torch.utils.data import DataLoader
import json
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--sparsity_dynamic_coeff', type=float, default=0.001, help="Sparsity loss coefficient")
parser.add_argument('--reprojection_dynamic_coeff', type=float, default=50.0, help="Reprojection error coefficient")
parser.add_argument('--dataset_folder', default="data/pet_test_set/our_data_format_4_validation_rgbd", help="Training data path")
parser.add_argument('--dataset_folder_validation', default="data/pet_test_set/our_data_format_4_validation_rgbd", help="Validation data path")
# Architechture hyper parameters
parser.add_argument('--conv_kernel_size', type=int, default=15)
parser.add_argument('--network_width1', type=int, default=256)
parser.add_argument('--positional_dim', type=int, default=12, help="if only_static>0, only optimize the static part")
parser.add_argument('--K_basis', type=int, default=12, help="if only_static>0, only optimize the static part")
parser.add_argument('--max_cameras', type=int, default=50, help="Maximum number of frames")
parser.add_argument('--continue_trainining_checkpoint', default="pretrained_checkpoints/TracksTo4D_pretrained_cats.pt", help="Path of the checkpoint to continue training from. (Used if continue_trainining>0)")
parser.add_argument('--continue_trainining', type=int, default=0, help="if >0, continue training from a checkpoint")
parser.add_argument('--predict_focal_length', type=int, default=0, help="If>0, the network predicts focal length correction.")
opt = parser.parse_args()
predict_focal_length=opt.predict_focal_length>0
positional_dim=opt.positional_dim
continue_trainining=opt.continue_trainining>0
continue_trainining_checkpoint=opt.continue_trainining_checkpoint
dataset_folder=opt.dataset_folder
dataset_folder_validation=opt.dataset_folder_validation
conv_kernel_size=opt.conv_kernel_size
network_width1=opt.network_width1
K_basis=opt.K_basis
max_cameras=opt.max_cameras
training_data_init=CustomDataset(22,dataset_folder,max_points=100,sample_cameras=False)
training_data=CustomDataset(max_cameras,dataset_folder,max_points=100,sample_cameras=True)
validation_data=CustomDataset(max_cameras,dataset_folder_validation,max_points=100)
reprojection_dynamic_coeff=opt.reprojection_dynamic_coeff
sparsity_dynamic_coeff=opt.sparsity_dynamic_coeff
if continue_trainining:
logs_folder=continue_trainining_checkpoint[:-22]
else:
logs_folder= "runs/"+datetime.datetime.utcnow().strftime("%m_%d_%Y__%H_%M_%S_%f")+"_"+dataset_folder
checkpoints_folder="%s/checkpoints"%logs_folder
Path(logs_folder).mkdir(parents=True, exist_ok=True)
Path(checkpoints_folder).mkdir(parents=True, exist_ok=True)
opt.logs_folder=logs_folder
input_dim=3
net=TracksTo4DNet(inputdim=input_dim,conv_kernel_size=conv_kernel_size,width1=network_width1,positional_dim=positional_dim,K=K_basis,predict_focal_length=predict_focal_length)
net=net.to("cuda")
optimizer = optim.Adam(net.parameters(), lr=0.0001)
num_pre_train_epochs=100
with open("%s/conf.json"%logs_folder, "w") as out_file:
json.dump(vars(opt), out_file)
train_dataloader = DataLoader(training_data_init, batch_size=1, shuffle=True)
validation_dataloader = DataLoader(validation_data, batch_size=1, shuffle=True)
start_epoch=0
if continue_trainining:
checkpoint=torch.load(continue_trainining_checkpoint)
aa=0
start_epoch=checkpoint["epoch"]+1
net.load_state_dict(checkpoint["model_state_dict"])
optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
if True:
if start_epoch>=50:
train_dataloader = DataLoader(training_data, batch_size=1, shuffle=True)
for epoch in range(start_epoch,1000000,1):
if epoch==50:
train_dataloader = DataLoader(training_data, batch_size=1, shuffle=True)
net.train()
sum_epoch_loss=0
sum_epoch_reprojection_loss=0
num_epoch_tot=0
for gt_mask,gt_depth,tracks,_,GT_poses,tracks_vis,camera_num,GT_Ks,dataset_name,start_ind,trecklets_num in train_dataloader:
optimizer.zero_grad()
if trecklets_num.min()<100:
continue
focal,projections_,projections_static_,rotation_params_,translation_params_,B_,points3D_,points3D_static_,depths_,depths_static_,projection_before_devide_,basis_params,_,_,points3D_camera,NR=net(tracks)
if epoch<num_pre_train_epochs:
if torch.isnan(translation_params_.sum()):
aaa=0
loss=((translation_params_-torch.tensor([[[0.0,0.0,-15.0]]]).cuda())**2).sum(dim=1).mean()
loss+=((rotation_params_-torch.eye(3).unsqueeze(0).unsqueeze(0).cuda())**2).mean()*100
if torch.isnan(loss):
aaa=0
loss/=100
if loss<0.0001:
num_pre_train_epochs=0
print("epoch %d"%epoch)
print(loss)
else:
loss,sparsity_loss,negative_depth_loss,reprojection_error,reprojection_error_static=get_losses(projections_,tracks[:,:,:2,:],tracks_vis,projections_static_,NR,camera_num,depths_,reprojection_dynamic_coeff,B_,sparsity_dynamic_coeff)
if epoch>3:
loss=torch.clamp(loss,max=100)
sum_epoch_reprojection_loss += reprojection_error.item()
loss.backward()
optimizer.step()
sum_epoch_loss+=loss.item()
num_epoch_tot+=1
mean_epoch_loss=sum_epoch_loss/num_epoch_tot
if epoch%10==0:
print("---epoch---%d"%epoch)
print(mean_epoch_loss)
if (epoch%100==0 ) and epoch>0:
torch.save({
'epoch': epoch,
'model_state_dict': net.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': mean_epoch_loss
}, "%s/%06d.pt"%(checkpoints_folder,epoch))
net.eval()
with torch.no_grad():
sum_validation_loss=0
sum_validation_reprojection_loss=0
num_validation_loss=0
for gt_mask,gt_depth,tracks,_,GT_poses,tracks_vis,camera_num,GT_Ks,dataset_name,start_ind,trecklets_num in validation_dataloader:
focal,projections_,projections_static_,rotation_params_,translation_params_,B_,points3D_,points3D_static_,depths_,depths_static_,projection_before_devide_,basis_params,_,_,points3D_camera,NR=net(tracks)
loss,sparsity_loss,negative_depth_loss,reprojection_error,reprojection_error_static=get_losses(projections_,tracks[:,:,:2,:],tracks_vis,projections_static_,NR,camera_num,depths_,reprojection_dynamic_coeff,B_,sparsity_dynamic_coeff)
sum_validation_loss+=loss.item()
sum_validation_reprojection_loss+=reprojection_error.item()
num_validation_loss+=1.0
print("-------------------------")
print("validation_loss:")
print(sum_validation_loss/num_validation_loss)
print("-------------------------")
if __name__ == '__main__':
main()