/
train_fn.py
176 lines (137 loc) · 6.59 KB
/
train_fn.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
import numpy as np
import torch
import torch.nn.functional as F
import os, time
import matplotlib.pyplot as plt
from torch import nn, optim
from models import pointnet
from utils import random_rotate_batch, random_rotate_y_batch
from visualize import output_meshes
def unfold_rotenc(data, rotenc, iters=5):
R_cum = torch.eye(3).unsqueeze(0).repeat(data.size(0), 1, 1).to(data.device)
for _ in range(iters):
R = rotenc(data.transpose(1, 2).contiguous())
data = torch.matmul(data, R).detach()
R_cum = torch.matmul(R_cum, R)
return R_cum
def train_autoencoder(models, losses, optimizers, data, epoch, opt):
optimizers['opt'].zero_grad()
if opt.azimuthal:
random_rotate = random_rotate_y_batch
else:
random_rotate = random_rotate_batch
if opt.art:
data_rot_1, rotmat_1 = random_rotate(data)
data_rot_2, rotmat_2 = random_rotate(data)
data_rot_3, rotmat_3 = random_rotate(data)
R_1 = unfold_rotenc(data_rot_1, models['rot_enc'], opt.iters)
R_2 = unfold_rotenc(data_rot_2, models['rot_enc'], opt.iters)
R_3 = unfold_rotenc(data_rot_3, models['rot_enc'], opt.iters)
R = unfold_rotenc(data, models['rot_enc'], opt.iters)
rotprod_1 = torch.matmul(R, R_1.transpose(1, 2))
rotprod_2 = torch.matmul(R, R_2.transpose(1, 2))
rotprod_3 = torch.matmul(R, R_3.transpose(1, 2))
rot_loss_mse = (F.mse_loss(rotmat_1, rotprod_1) + \
F.mse_loss(rotmat_2, rotprod_2) + \
F.mse_loss(rotmat_3, rotprod_3)) / 3
z = models['enc'](data, R)
y = models['dec'](z, R)
with torch.cuda.device(data.device):
rot_loss_chamfer = (losses['chamfer'](torch.matmul(data, rotprod_1), data_rot_1) + \
losses['chamfer'](torch.matmul(data, rotprod_2), data_rot_2) + \
losses['chamfer'](torch.matmul(data, rotprod_3), data_rot_3)) / 3
elif opt.itn:
R = unfold_rotenc(data, models['rot_enc'], opt.iters)
z = models['enc'](data, R)
y = models['dec'](z, R)
elif opt.tnet:
R = models['rot_enc'](data.transpose(1, 2).contiguous())
z = models['enc'](data, R)
y = models['dec'](z, torch.inverse(R))
else:
data, _ = random_rotate(data)
z = models['enc'](data)
y = models['dec'](z)
with torch.cuda.device(data.device):
chamfer_dist = losses['chamfer'](data, y)
if opt.art:
loss = chamfer_dist + rot_loss_mse * 0.02 + rot_loss_chamfer * opt.lambda2
else:
loss = chamfer_dist
rot_loss_mse = torch.tensor(0)
rot_loss_chamfer = torch.tensor(0)
loss.backward()
optimizers['opt'].step()
return chamfer_dist, rot_loss_mse, rot_loss_chamfer
def train_model(models, losses, optimizers, train_loader, vald_loader, device, opt, save_path=None):
num_epochs = 500
start_epoch = 1
vis_step = 500
log_step = 1
best_loss = 1000
ckpt_files = sorted(os.listdir(save_path))
if opt.resume and len(ckpt_files) > 0:
ckpt_file = ckpt_files[-1]
ckpt = torch.load(os.path.join(save_path, ckpt_file), map_location=device)
for k in models:
if models[k]:
models[k].load_state_dict(ckpt['m_'+k])
for k in optimizers:
optimizers[k].load_state_dict(ckpt['o_'+k])
start_epoch = int(ckpt_file.split('.')[0]) + 1
print('Training started')
print('azimuthal?', opt.azimuthal)
for epoch in range(start_epoch, 1+num_epochs):
t1 = time.time()
models['enc'].train()
models['dec'].train()
if opt.art:
models['rot_enc'].train()
train_loader.dataset.resample()
train_loss_dict = {'chamfer_dist': 0, 'rot_loss_mse': 0, 'rot_loss_chamfer': 0}
vald_loss_dict = {'chamfer_dist': 0}
for i, data in enumerate(train_loader):
data = data.to(device)
recon_loss, rot_loss_mse, rot_loss_chamfer = train_autoencoder(models, losses, optimizers, data, epoch, opt)
train_loss_dict['chamfer_dist'] += recon_loss.item() * data.size(0)
train_loss_dict['rot_loss_mse'] += rot_loss_mse.item() * data.size(0)
train_loss_dict['rot_loss_chamfer'] += rot_loss_chamfer.item() * data.size(0)
t2 = time.time()
print(t2-t1)
if epoch > 0:
models['enc'].eval()
models['dec'].eval()
if opt.art:
models['rot_enc'].eval()
with torch.no_grad():
for batch_idx, x in enumerate(vald_loader):
x = x.to(device)
if opt.art:
R = unfold_rotenc(x, models['rot_enc'], opt.iters)
z = models['enc'](x, R)
y = models['dec'](z, R)
else:
z = models['enc'](x)
y = models['dec'](z)
with torch.cuda.device(device):
recon_loss = losses['chamfer'](x, y)
vald_loss_dict['chamfer_dist'] += recon_loss.item() * x.size(0)
if epoch % vis_step == 0 and batch_idx == 0:
x = x.cpu().numpy().reshape(x.shape[0], 1, -1, 3)
y = y.cpu().numpy().reshape(y.shape[0], 1, -1, 3)
meshes = np.concatenate([x, y], axis=1)
output_meshes(meshes, epoch)
if epoch % log_step == 0:
print('====> Epoch {}/{}: Training'.format(epoch, num_epochs), flush=True)
for term in train_loss_dict:
print('\t{} {:.5f}'.format(term, train_loss_dict[term] / len(train_loader.dataset)), flush=True)
vald_loss = vald_loss_dict['chamfer_dist'] / len(vald_loader.dataset)
print('====> Epoch {}/{}: Validation'.format(epoch, num_epochs), flush=True)
for term in vald_loss_dict:
print('\t{} {:.5f}'.format(term, vald_loss), flush=True)
if vald_loss < best_loss:
best_loss = vald_loss
checkpoint = dict([('m_'+t, models[t].state_dict() if models[t] else None) for t in models])
checkpoint.update(dict([('o_'+t, optimizers[t].state_dict()) for t in optimizers]))
checkpoint.update({'torch_rnd': torch.get_rng_state(), 'numpy_rnd': np.random.get_state()})
torch.save(checkpoint, os.path.join(save_path, '{}.pth'.format(epoch)))