/
train_foldingNet.py
235 lines (200 loc) · 11.9 KB
/
train_foldingNet.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
from torch.utils.data import DataLoader
import torch
import torch.optim as optim
import torch.backends.cudnn as cudnn
from torch.autograd import Variable
import argparse
import numpy as np
# import models.PointNetFCAE as PointNetFCAE
from models.Folding import FlodingNet
import sys
sys.path.append('./models')
sys.path.append('./reference')
from kaolin.metrics.point import chamfer_distance
import utliz
from reference.dataset_shapenet_completion import PartDataset_Completion
parser = argparse.ArgumentParser('Completion Trainer')
parser.add_argument('--batch_size', type=int, help='batch size', default=64)
parser.add_argument('--num_epochs', type=int, help='number of epochs', default=9999)
parser.add_argument('--num_points', type=int, help='input batch size', default=2048)
parser.add_argument('--log_dir', type=str, help='directory to save checkpoints and summaries', default='')
parser.add_argument('--checkpoint_path', type=str, help='path to a specific checkpoint to load', default='')
parser.add_argument('--code_nfts', default=1024, type=int, help='Encoder output feature size')
args = parser.parse_args()
###################### ShapeNet Completion Dataset ########################
train_dataset = PartDataset_Completion(train=True, npoints=1024)
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=0)
test_dataset = PartDataset_Completion(train=False, npoints=1024)
test_dataloader = torch.utils.data.DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=0)
train_examples = len(train_dataset)
train_batches = len(train_dataloader)
test_examples = len(test_dataset)
test_batches = len(test_dataloader)
global plotter
plotter = utliz.VisdomLinePlotter(env_name='QH Completion_ShapeNet')
struc_name_list = ['ShapeNet']
strucNum = 0
###########################################################################
############################ Model & Optimizer ##############################
# Choose model #
completion_model = FlodingNet()
# completion_model = ConstrucNet()
# completion_model = ReshapeNet()
# completion_model = PCN()
# completion_model = Vanilla_GAN()
# completion_model = point_cloud_generator()
# discriminator_model = mlp_discriminator()
# discriminator_model.cuda()
# optimizer_D = optim.Adam(discriminator_model.parameters(), lr=1e-3)
completion_model.cuda()
optimizer = optim.Adam(completion_model.parameters(), lr=1e-4)
# optimizer = optim.Adagrad(completion_model.parameters(), lr=1e-2)
#############################################################################
grads = {}
epoch_base = 0
if args.log_dir != '':
checkpoint = torch.load(args.log_dir)
completion_model.load_state_dict(checkpoint['model_state_dict'])
epoch_base = checkpoint['epoch']
print('training from epoch {}'.format(epoch_base))
print("Train examples: {}".format(train_examples))
print("Evaluation examples: {}".format(test_examples))
print("Start training...")
cudnn.benchmark = True
total_test_loss_min = 999
best_epoch = 0
for epoch in range(2002):
print("--------Epoch {}--------".format(epoch+epoch_base))
# train one epoch
completion_model.train()
total_train_loss_test = 0
total_train_loss_repulsion = 0
train_pc_loss = [utliz.AverageMeter(), utliz.AverageMeter()] #partial/pred CD loss
for batch_idx, data in enumerate(train_dataloader, 0):
point_partial_global, point_gt_global = data
point_partial_global = point_partial_global.permute(0, 2, 1) # Bs*3*1024
point_partial_global, point_gt_global = Variable(point_partial_global.float(), requires_grad=True), Variable(point_gt_global)
point_partial_global, point_gt_global = point_partial_global.cuda(), point_gt_global.cuda()
optimizer.zero_grad()
point_partial_global_center = point_partial_global.mean(2).unsqueeze(1).repeat(1,1024,1).permute(0,2,1)
point_partial_local = point_partial_global - point_partial_global_center
################################ FoldingNet ##################################
_, pred_local = completion_model(point_partial_local)
# move pred to partial or align pred with partial
pred_global = pred_local + point_partial_global_center.permute(0, 2, 1) - pred_local.mean(1).unsqueeze(1)
# loss of FoldingNet
for i in range(point_partial_global.size()[0]):
if i==0:
loss = chamfer_distance(pred_global[i], point_gt_global[i])
else:
loss += chamfer_distance(pred_global[i], point_gt_global[i])
train_pc_loss[1].update(loss.item())
loss.backward()
optimizer.step()
##############################################################################
if False:
print(
"batch: {}/{}, Train loss: {:.4f}".format(batch_idx, len(train_dataloader), loss.item() ))
# torch.save({
# 'epoch': epoch+epoch_base,
# 'model_state_dict': completion_model.state_dict(),
# 'optimizer_state_dict': optimizer.state_dict(),
# 'loss': loss,
# }, './checkpoints/epoch_'+str(epoch+epoch_base)+'.pt')
print("Train pc loss: {:.4f}".format(train_pc_loss[1].avg))
plotter.plot('Epoch_loss', 'train_cd', 'cd loss & dice', epoch, train_pc_loss[1].sum)
if (epoch % 20 == 0):
color_grid = utliz.get_color()
plotter.scatter('pointcloud_train_pred' + str(epoch), 'train', 'pointcloud_train_pred' + str(epoch), pred_global[0, :, :].data,
size=1, color=np.round(color_grid.reshape(-1, 3) * 255))#np.array([0]))
plotter.scatter('pointcloud_train_part' + str(epoch), 'train', 'pointcloud_train_part' + str(epoch), point_partial_global.permute(0,2,1)[0, :, :].data,
size=1, color=np.array([[100, 0, 0]]))
plotter.scatter('pointcloud_train_gt' + str(epoch), 'train', 'pointcloud_train_gt' + str(epoch), point_gt_global[0, :, :].data,
size=1, color=np.array([[0, 0, 100]]))
### EVAL ###
completion_model.eval()
test_pc_loss = [utliz.AverageMeter(), utliz.AverageMeter()] #partial/pred CD loss
for batch_idx, data in enumerate(test_dataloader, 0):
point_partial_global, point_gt_global = data
point_partial_global = point_partial_global.permute(0, 2, 1) # Bs*3*2048
# point_gt = point_gt.permute(0, 2, 1) # Bs*3*2048
point_partial_global, point_gt_global = Variable(point_partial_global.float()), Variable(point_gt_global)
point_partial_global, point_gt_global = point_partial_global.cuda(), point_gt_global.cuda()
point_partial_global_center = point_partial_global.mean(2).unsqueeze(1).repeat(1,1024,1).permute(0,2,1)
point_partial_local = point_partial_global - point_partial_global_center
################################ FoldingNet ##################################
_, pred_local = completion_model(point_partial_local)
# move pred to partial or align pred with partial
pred_global = pred_local + point_partial_global_center.permute(0, 2, 1) - pred_local.mean(1).unsqueeze(1)
# loss of FoldingNet
for i in range(point_partial_global.size()[0]):
if i==0:
loss = chamfer_distance(pred_global[i], point_gt_global[i])
# loss_partial = chamfer_distance(point_partial_global.permute(0,2,1)[i], point_gt_global[i])
else:
loss += chamfer_distance(pred_global[i], point_gt_global[i])
# loss_partial += chamfer_distance(point_partial_global.permute(0,2,1)[i], point_gt_global[i])
test_pc_loss[1].update(loss.item())
# total_test_loss_partial += loss_partial.item()
##############################################################################
# if batch_idx % 1000 == 1:
if False:
print(
"batch: {}/{}, Train loss: {:.4f}".format(
batch_idx, len(train_dataloader), loss.item()))
# torch.save({
# 'epoch': epoch+epoch_base,
# 'model_state_dict': completion_model.state_dict(),
# 'optimizer_state_dict': optimizer.state_dict(),
# 'loss': loss,
# }, './checkpoints/epoch_'+str(epoch+epoch_base)+'.pt')
print("MIN LOSS: ", total_test_loss_min, "BEST EPOCH: ", best_epoch)
print("Test loss: {:.4f}".format(test_pc_loss[1].avg ))
plotter.plot('Epoch_loss', 'test_cd', 'cd loss & dice', epoch, test_pc_loss[1].sum)
if (epoch % 100 == 1):
plotter.scatter('pointcloud_test_pred' + str(epoch), 'test', 'pointcloud_test_pred' + str(epoch),
pred_global[0, :, :].data,
size=1, color=np.round(color_grid.reshape(-1, 3) * 255))#np.array([0]))
plotter.scatter('pointcloud_test_part' + str(epoch), 'train', 'pointcloud_test_part' + str(epoch),
point_partial_global.permute(0, 2, 1)[0, :, :].data,
size=1, color=np.array([[100, 0, 0]]))
plotter.scatter('pointcloud_test_gt' + str(epoch), 'test', 'pointcloud_test_gt' + str(epoch),
point_gt_global[0, :, :].data,
size=1, color=np.array([[0, 0, 100]]))
print('END')
plotter_pause = utliz.VisdomLinePlotter(env_name='Completion_debug'+struc_name_list[strucNum])
plotter_pause.scatters2D = {}
plotter_pause.scatter2D('pc_xy', 'gt', 'gt', x=point_gt[0,:,0:2],color=[0,255,0],size=1)
plotter_pause.scatter2D('pc_xy', 'partial', 'partial', x=point_partial[0,0:2,:].permute(1,0),color=[255,0,0],size=2)
plotter_pause.scatter2D('pc_xy', 'pred', 'pred', x=pred[0,:,0:2].data,color=[0,0,255],size=2)
i_ = 2
plotter_pause.scatters2D = {}
plotter_pause.scatter2D('pc_xy', 'gt', 'gt', x=point_gt[i_,:,0:2],color=[0,255,0],size=1)
plotter_pause.scatter2D('pc_xy', 'partial', 'partial', x=point_partial[i_,0:2,:].permute(1,0),color=[255,0,0],size=2)
plotter_pause.scatters2D = {}
plotter_pause.scatter2D('pc_xy', 'gt', 'gt', x=point_gt[i_,:,0:2],color=[0,255,0],size=1)
plotter_pause.scatter2D('pc_xy', 'pred', 'pred', x=pred[i_,:,0:2].data,color=[0,0,255],size=2)
for i in range(8):
plotter_pause.scatters2D = {}
plotter_pause.scatter2D('pc_xy_gpar_'+str(i), 'gt', 'gpar_'+str(i), x=point_gt_global[i, :, 0:2], color=[0, 255, 0], size=1)
plotter_pause.scatter2D('pc_xy_gpar_'+str(i), 'partial', 'gpar_'+str(i), x=point_partial_global[i, 0:2, :].permute(1, 0), color=[255, 0, 0],
size=2)
plotter_pause.scatter2D('pc_xy_gpre_' + str(i), 'gt', 'gpre_'+str(i), x=point_gt_global[i, :, 0:2], color=[0, 255, 0], size=1)
plotter_pause.scatter2D('pc_xy_gpre_'+str(i), 'pred', 'gpre_'+str(i), x=pred_global[i, :, 0:2].data, color=[0, 0, 255], size=2)
plotter_pause.scatter2D('pc_xy_pp_' + str(i), 'partial', 'pp_'+str(i), x=point_partial_global[i, 0:2, :].permute(1, 0),
color=[255, 0, 0],
size=2)
plotter_pause.scatter2D('pc_xy_pp_' + str(i), 'pred', 'pp_'+str(i), x=pred_global[i, :, 0:2].data, color=[0, 0, 255], size=2)
for i in range(8):
plotter_pause.scatters = {}
plotter_pause.scatter('pc_xy_gpar_' + str(i), 'gpar', 'gpar_' + str(i),
x=torch.cat((point_gt, point_partial.permute(0, 2, 1)), 1)[i, :, :].data,
color=np.array([[255, 0, 0], [0, 255, 0]]).repeat(1024, 0), size=2)
plotter_pause.scatter('pc_xy_gpre_' + str(i), 'gpre', 'gpre_' + str(i),
x=torch.cat((point_gt, pred), 1)[i, :, :].data,
color=np.array([[255, 0, 0], [0, 0, 255]]).repeat(1024, 0), size=2)
plotter_pause.scatter('pc_xy_pp_' + str(i), 'pp', 'pp_' + str(i),
x=torch.cat((point_partial.permute(0,2,1), pred), 1)[i, :, :].data,
color=np.array([[0, 255, 0], [0, 0, 255]]).repeat(1024, 0), size=2)
point_gt_partial_pred = torch.cat((point_gt, point_partial.permute(0, 2, 1), pred), 2)
np.save('./results/pc_gt_partial_pred/pc_gt_partial_pred', point_gt_partial_pred.cpu().detach().numpy())