-
Notifications
You must be signed in to change notification settings - Fork 8
/
train_sdf_space.py
180 lines (148 loc) · 5.8 KB
/
train_sdf_space.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
import os
import torch
import numpy as np
from tqdm import tqdm
from config import parse_args
from models import MLPSpace
from losses import sdf_loss
from utils import write_sdf_summary, create_mesh
from datasets import DFaustDataset
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from functools import partial
# configs
FLAGS = parse_args()
# dataset
flags_data = FLAGS.DATA.train
dfaust_dataset = DFaustDataset(**flags_data)
dataloader = DataLoader(dfaust_dataset, batch_size=flags_data.batch_size,
num_workers=24, shuffle=True, pin_memory=True,
drop_last=True)
# model
model = MLPSpace(**FLAGS.MODEL)
print(model)
model.cuda()
# load checkpoints
flags_solver = FLAGS.SOLVER
if flags_solver.ckpt:
print('loading checkpoint %s' % flags_solver.ckpt)
model.load_state_dict(torch.load(flags_solver.ckpt))
# init from sphere
if FLAGS.MODEL.name == 'optpos' and flags_solver.sphere_init:
print('Init from sphere, load: %s' % flags_solver.sphere_init)
trained_dict = torch.load(flags_solver.sphere_init)
shape_num = FLAGS.MODEL.shape_num
shape_code = trained_dict.pop('pos_enc.shape_code')
trained_dict['pos_enc.shape_code'] = shape_code.repeat(1, shape_num)
model_dict = model.state_dict()
model_dict.update(trained_dict)
model.load_state_dict(model_dict)
if FLAGS.MODEL.name == 'mlp' and flags_solver.sphere_init:
net = model.net.net
for i in range(len(net)-1):
weight, bias = net[i].linear.weight, net[i].linear.bias
torch.nn.init.normal_(weight, 0.0, np.sqrt(2 / weight.shape[0]))
torch.nn.init.constant_(bias, 0.0)
weight, bias = net[-1].linear.weight, net[-1].linear.bias
torch.nn.init.constant_(bias, -0.6)
torch.nn.init.normal_(weight, mean=np.sqrt(np.pi / weight.shape[1]), std=1e-5)
# optmizer
lr = flags_solver.learning_rate
optim = torch.optim.Adam(lr=lr, params=model.parameters())
if flags_solver.optim_ckpt:
print('loading checkpoint %s' % flags_solver.optim_ckpt)
optim.load_state_dict(torch.load(flags_solver.optim_ckpt))
# summaries
logdir = flags_solver.logdir
ckpt_dir = os.path.join(logdir, 'checkpoints')
writer = SummaryWriter(logdir)
if not os.path.exists(ckpt_dir):
os.makedirs(ckpt_dir)
# latent code regularization
def shape_code_reg(idx):
shape_code = model.pos_enc.get_shape_code(idx)
code_loss = shape_code.square().mean() # or sum()
return code_loss
# train
def train_step(model_train, global_step):
model_train.train()
avg_loss = []
for i, data in enumerate(dataloader):
coords = data[0].cuda().requires_grad_()
sdf_gt, normal_gt, idx = data[1].cuda(), data[2].cuda(), data[3].cuda()
sdf = model_train(coords, idx)
losses = sdf_loss(sdf, coords, sdf_gt, normal_gt,
normal_weight=FLAGS.LOSS.normal_weight,
grad_weight=FLAGS.LOSS.grad_weight)
total_train_loss = losses['total_train_loss']
# latent code regularization
code_loss = shape_code_reg(idx)
total_loss = total_train_loss + code_loss * 1e-4
optim.zero_grad()
total_loss.backward()
optim.step()
# tqdm.write("step %d" % (global_step + i))
for k, v in losses.items():
writer.add_scalar(k, v.detach().cpu().item(), global_step + i)
writer.add_scalar('latent', code_loss.detach().cpu().item(), global_step+1)
avg_loss.append(total_loss.detach().cpu().item())
return np.mean(avg_loss)
# test
def test_step(epoch=0, idx=None, save_sdf=True):
model.eval()
if idx is None:
idx = np.random.randint(len(dfaust_dataset))
output_path = os.path.join(logdir, 'mesh')
if not os.path.exists(output_path): os.makedirs(output_path)
filename = '%s_%04d_%04d.ply' % (flags_solver.alias, epoch, idx)
filename = os.path.join(output_path, filename)
model_test = partial(model, idx=idx)
create_mesh(model_test, filename, N=flags_solver.resolution,
save_sdf=save_sdf, level=flags_solver.level_set)
# run train
def train():
model_train = model
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
model_train = torch.nn.DataParallel(model) # use multiple gpus
num = len(dataloader)
rng = range(flags_solver.start_epoch, flags_solver.num_epochs)
for epoch in tqdm(rng, ncols=80):
global_step = epoch * num
if epoch % flags_solver.test_every_epoch == 0:
write_sdf_summary(model, writer, global_step)
save_state(filename='model_%05d' % epoch)
test_step(epoch, save_sdf=False)
train_loss = train_step(model_train, global_step)
tqdm.write("Epoch %d, Total loss %0.6f" % (epoch, train_loss))
save_state(filename='model_final')
upsample_code()
# run test
def test():
num = FLAGS.MODEL.shape_num
for i in tqdm(range(num), ncols=80):
test_step(idx=i, save_sdf=True)
# upsample the hidden code
def upsample_code():
size = flags_solver.upsample_size
if size < 0: return
# upsample
model_dict = model.state_dict()
with torch.no_grad():
code = model.pos_enc.upsample(size)
model_dict['pos_enc.shape_code'] = code
# save checkpoints
ckpt_name = os.path.join(ckpt_dir, 'model_final_upsample_%03d.pth' % size)
torch.save(model_dict, ckpt_name)
# save model and solver state
def save_state(filename):
model_dict = model.state_dict()
ckpt_name = os.path.join(ckpt_dir, filename + '.pth')
torch.save(model_dict, ckpt_name)
ckpt_name = os.path.join(ckpt_dir, filename + '.mean.pth')
model_dict['pos_enc.shape_code'] = model.pos_enc.get_mean_code()
torch.save(model_dict, ckpt_name)
ckpt_name = os.path.join(ckpt_dir, filename + '.solver.pth')
torch.save(optim.state_dict(), ckpt_name)
if __name__ == '__main__':
eval('{}()'.format(flags_solver.run))