/
train_segmentation.py
executable file
·194 lines (161 loc) · 7.5 KB
/
train_segmentation.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
import argparse
import datetime
import os
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm
from hodgenet import HodgeNetModel
from meshdata import HodgenetMeshDataset
def main(args):
torch.set_default_dtype(torch.float64) # needed for eigenvalue problems
torch.manual_seed(args.seed)
np.random.seed(args.seed)
mesh_files_train = []
seg_files_train = []
mesh_files_val = []
seg_files_val = []
files = sorted([f.split('.')[0] for f in os.listdir(args.mesh_path)])
cutoff = round(0.85 * len(files) + 0.49)
for i in files[:cutoff]:
mesh_files_train.append(os.path.join(args.mesh_path, f'{i}.off'))
seg_files_train.append(os.path.join(args.seg_path, f'{i}.seg'))
for i in files[cutoff:]:
mesh_files_val.append(os.path.join(args.mesh_path, f'{i}.off'))
seg_files_val.append(os.path.join(args.seg_path, f'{i}.seg'))
features = ['vertices'] if args.no_normals else ['vertices', 'normals']
dataset = HodgenetMeshDataset(
mesh_files_train,
decimate_range=None if args.fine_tune is not None else (1000, 99999),
edge_features_from_vertex_features=features,
triangle_features_from_vertex_features=features,
max_stretch=0 if args.fine_tune is not None else 0.05,
random_rotation=False, segmentation_files=seg_files_train,
normalize_coords=True)
validation = HodgenetMeshDataset(
mesh_files_val, decimate_range=None,
edge_features_from_vertex_features=features,
triangle_features_from_vertex_features=features, max_stretch=0,
random_rotation=False, segmentation_files=seg_files_val,
normalize_coords=True)
def mycollate(b): return b
dataloader = DataLoader(dataset, batch_size=args.bs,
num_workers=args.num_workers, shuffle=True,
collate_fn=mycollate)
validationloader = DataLoader(validation, batch_size=args.bs,
num_workers=args.num_workers,
collate_fn=mycollate)
example = dataset[0]
hodgenet = HodgeNetModel(
example['int_edge_features'].shape[1],
example['triangle_features'].shape[1],
num_output_features=args.n_out_features, mesh_feature=False,
num_eigenvectors=args.n_eig, num_extra_eigenvectors=args.n_extra_eig,
resample_to_triangles=True,
num_vector_dimensions=args.num_vector_dimensions)
model = nn.Sequential(
hodgenet,
nn.Linear(args.n_out_features*args.num_vector_dimensions *
args.num_vector_dimensions, 32),
nn.BatchNorm1d(32),
nn.LeakyReLU(),
nn.Linear(32, 32),
nn.BatchNorm1d(32),
nn.LeakyReLU(),
nn.Linear(32, dataset.n_seg_categories))
# categorical variables
loss = nn.CrossEntropyLoss()
# optimization routine
print(sum(x.numel() for x in model.parameters()), 'parameters')
optimizer = optim.AdamW(model.parameters(), lr=args.lr)
if not os.path.exists(args.out):
os.makedirs(args.out)
if args.fine_tune is not None:
checkpoint = torch.load(os.path.join(args.fine_tune))
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['opt_state_dict'])
starting_epoch = checkpoint['epoch'] + 1
print(f'Fine tuning! Starting at epoch {starting_epoch}')
else:
starting_epoch = 0
train_writer = SummaryWriter(os.path.join(
args.out, datetime.datetime.now().strftime('train-%m%d%y-%H%M%S')),
flush_secs=1)
val_writer = SummaryWriter(os.path.join(
args.out, datetime.datetime.now().strftime('val-%m%d%y-%H%M%S')),
flush_secs=1)
def epoch_loop(dataloader, epochname, epochnum, writer, optimize=True):
epoch_loss, epoch_acc, epoch_acc_weighted, epoch_size = 0, 0, 0, 0
pbar = tqdm(total=len(dataloader), desc=f'{epochname} {epochnum}')
for batch in dataloader:
if optimize:
optimizer.zero_grad()
batch_loss, batch_acc, batch_acc_weighted = 0, 0, 0
seg_estimates = torch.split(model(batch), [m['triangles'].shape[0]
for m in batch], dim=0)
for mesh, seg_estimate in zip(batch, seg_estimates):
gt_segs = mesh['segmentation'].squeeze(-1)
areas = mesh['areas']
batch_loss += loss(seg_estimate, gt_segs)
batch_acc += (seg_estimate.argmax(1) == gt_segs).float().mean()
batch_acc_weighted += ((seg_estimate.argmax(1) == gt_segs)
* areas).sum() / areas.sum()
epoch_loss += batch_loss.item()
epoch_acc += batch_acc.item()
epoch_acc_weighted += batch_acc_weighted.item()
epoch_size += len(batch)
batch_loss /= len(batch)
batch_acc /= len(batch)
batch_acc_weighted /= len(batch)
pbar.set_postfix({
'loss': batch_loss.item(),
'accuracy': batch_acc.item(),
'accuracy_weighted': batch_acc_weighted.item(),
})
pbar.update(1)
if optimize:
batch_loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), 1)
optimizer.step()
writer.add_scalar('loss', epoch_loss / epoch_size, epochnum)
writer.add_scalar('accuracy', epoch_acc / epoch_size, epochnum)
writer.add_scalar('accuracy_weighted',
epoch_acc_weighted / epoch_size, epochnum)
pbar.close()
for epoch in range(starting_epoch, starting_epoch+args.n_epochs+1):
model.train()
epoch_loop(dataloader, 'Epoch', epoch, train_writer)
# compute validation score
if epoch % 5 == 0:
model.eval()
with torch.no_grad():
epoch_loop(validationloader, 'Validation',
epoch, val_writer, optimize=False)
torch.save({
'model_state_dict': model.state_dict(),
'opt_state_dict': optimizer.state_dict(),
'epoch': epoch
}, os.path.join(args.out,
f'{epoch}_finetune.pth'
if args.fine_tune is not None else f'{epoch}.pth'))
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--out', type=str, default='out/vase')
parser.add_argument('--mesh_path', type=str, default='data/coseg_vase')
parser.add_argument('--seg_path', type=str, default='data/coseg_vase_gt')
parser.add_argument('--bs', type=int, default=16)
parser.add_argument('--lr', type=float, default=1e-4)
parser.add_argument('--n_epochs', type=int, default=100)
parser.add_argument('--n_eig', type=int, default=32)
parser.add_argument('--n_extra_eig', type=int, default=32)
parser.add_argument('--n_out_features', type=int, default=32)
parser.add_argument('--fine_tune', type=str, default=None)
parser.add_argument('--num_workers', type=int, default=0)
parser.add_argument('--num_vector_dimensions', type=int, default=4)
parser.add_argument('--seed', type=int, default=123)
parser.add_argument('--no_normals', action='store_true', default=False)
args = parser.parse_args()
main(args)