/
GEOMetrics.py
273 lines (221 loc) · 11.7 KB
/
GEOMetrics.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
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
from utils import *
from layers import *
from models import *
# Training settings
parser = argparse.ArgumentParser()
parser.add_argument('--render', action='store_true', default=False,
help='render images from validation set')
parser.add_argument('--seed', type=int, default=41, help='Random seed.')
parser.add_argument('--lr', type=float, default=0.0001,
help='Initial learning rate.')
# parser.add_argument('--weight_decay', type=float, default=5e-6,
# help='Weight decay (L2 loss on parameters).')
parser.add_argument('--object', type=str, default='chair',
help='The object... ')
args = parser.parse_args()
batch_size = 5
loss_term = point_to_point
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
# establish directories
images = 'data/images/' + args.object +'/'
meshes = 'data/mesh_info/' + args.object +'/'
samples = 'data/surfaces/' + args.object +'/'
checkpoint_dir = "checkpoint/" + args.object +'/'
save_dir = "plots/" + args.object +'/'
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
# gather data
adj_info, initial_positions= load_initial()
initial_positions = Variable(initial_positions.cuda())
data_train = Mesh_loader(images + 'train/' ,meshes,samples)
data_valid = Mesh_loader(images + 'valid/' ,meshes, samples)
# initialize models
encoder1 = VGG()
encoder2 = VGG()
encoder3 = VGG()
modelA = MeshDeformationBlock(963)
modelB = MeshDeformationBlock(1155)
modelC = MeshDeformationBlock(1155)
encoder_mesh = MeshEncoder(50)
modelA.cuda(), encoder1.cuda(), encoder2.cuda(),encoder3.cuda(), modelB.cuda(), modelC.cuda(), encoder_mesh.cuda()
params = list(modelA.parameters()) + list(encoder1.parameters()) + list(encoder2.parameters()) + list(encoder3.parameters()) +list(modelB.parameters()) +list(modelC.parameters())
optimizer = optim.Adam(params,lr=args.lr)
try:
encoder_mesh.load_state_dict(torch.load('checkpoint/'+ args.object + '/encoder'))
except:
print 'You forgot to train the auto-encoder'
exit()
if args.render:
modelA.load_state_dict(torch.load(checkpoint_dir + 'model_1'))
modelB.load_state_dict(torch.load(checkpoint_dir + 'model_2'))
modelC.load_state_dict(torch.load(checkpoint_dir + 'model_3'))
encoder1.load_state_dict(torch.load(checkpoint_dir + 'encoder_1'))
encoder2.load_state_dict(torch.load(checkpoint_dir + 'encoder_2'))
encoder3.load_state_dict(torch.load(checkpoint_dir + 'encoder_3'))
def train(loader, epoch, num_interations, iteration, best):
# initializatoins
t = time()
edge_loss, lap_loss = 0,0
modelA.train(), modelB.train(),modelC.train(), encoder1.train(), encoder2.train(), encoder3.train()
optimizer.zero_grad()
# load training batch
batch = loader.load_batch(batch_size)
# extract image features
A1, B1, C1, D1 = encoder1(batch['imgs'])
A2, B2, C2, D2 = encoder2(batch['imgs'])
A3, B3, C3, D3 = encoder3(batch['imgs'])
# initializations for graph convolutions
full_info = zip( batch['verts'],batch['adjs'],batch['samples'], batch['latent'],batch['mats'], range(batch_size))
loss = torch.FloatTensor(np.array(0.)).cuda()
track_loss, track_lap, track_edge, track_lat, track_surface = 0, 0, 0, 0, 0
real_distance = 0
latent_loss = 0
lap_loss = 0
size = 0
# loop for every image
for gt,gt_adj,gt_samp,late, mat, i in full_info :
# 3 mesh deformation layers
# stage 1
vertex_features = pooling([A1[i], B1[i], C1[i], D1[i]] , initial_positions, mat)
vertex_features, vertex_positions_1 =modelA(initial_positions, vertex_features,adj_info)
vertex_positions_1 = vertex_positions_1 + initial_positions
splitting_faces = calc_curve(vertex_positions_1, adj_info)
adj_info_2,vertex_positions_split_1,vertex_features= split_info(adj_info, vertex_positions_1, vertex_features, splitting_faces, number =splitting_faces.shape[0])
# stage 2
vertex_features = torch.cat((vertex_features, pooling([A2[i], B2[i], C2[i], D2[i]], vertex_positions_split_1, mat)), dim= 1 )
vertex_features, vertex_positions_2 = modelB( vertex_positions_split_1, vertex_features, adj_info_2)
vertex_positions_2 = vertex_positions_2 + vertex_positions_split_1
splitting_faces = calc_curve(vertex_positions_2, adj_info_2)
adj_info_3,vertex_positions_split_2,vertex_features= split_info(adj_info_2, vertex_positions_2, vertex_features, splitting_faces, number = splitting_faces.shape[0])
# stage 3
vertex_features = torch.cat((vertex_features, pooling([A3[i], B3[i], C3[i], D3[i]], vertex_positions_split_2, mat)), dim= 1 )
_,vertex_positions_3 = modelC(vertex_positions_split_2,vertex_features, adj_info_3)
vertex_positions_3 = vertex_positions_3 + vertex_positions_split_2
size += vertex_positions_3.shape[0]
## latent loss, possible encoding of graph does not exist
if late:
copy_adj = adj_info_3.copy()
latent_pred = encoder_mesh(vertex_positions_3, copy_adj )
latent_gt = encoder_mesh(gt, gt_adj )
latent_loss = torch.mean(torch.abs(latent_pred - latent_gt)) * .001
# surface loss
surface_loss_1,_ = loss_term(vertex_positions_1, adj_info, gt_samp[:3000] ,num = 3000 )
surface_loss_2,_ = loss_term(vertex_positions_2, adj_info_2, gt_samp[:3000],num = 3000 )
surface_loss_3,_ = loss_term(vertex_positions_3, adj_info_3, gt_samp[:3000] ,num = 3000 )
surface_loss = surface_loss_1 + surface_loss_2 + surface_loss_3
real_distance += surface_loss_3.data.cpu().numpy()
# edge length loss
edge_loss = calc_edge(vertex_positions_1, adj_info) * 300
edge_loss += calc_edge(vertex_positions_2, adj_info_2) * 300
edge_loss += calc_edge(vertex_positions_3, adj_info_3) * 300
edge_loss *= .6
# laplacian loss
lap_loss_1 = torch.mean(torch.sum((get_lap_info( initial_positions, adj_info) - get_lap_info( vertex_positions_1, adj_info))**2, 1 ))* 1500
lap_loss_2 = torch.mean(torch.sum((get_lap_info( vertex_positions_split_1, adj_info_2) - get_lap_info( vertex_positions_2, adj_info_2))**2, 1)) * 1500
lap_loss_2 += torch.mean(torch.sum((vertex_positions_split_1 - vertex_positions_2)**2, 1)) * 100
lap_loss_3 = torch.mean(torch.sum((get_lap_info( vertex_positions_split_2, adj_info_3) - get_lap_info( vertex_positions_3, adj_info_3))**2, 1)) * 1500
lap_loss_3 += torch.mean(torch.sum((vertex_positions_split_2- vertex_positions_3)**2, 1)) * 100
lap_loss = lap_loss_1*.3 + lap_loss_2 + lap_loss_3
# combination of losses
loss += edge_loss + latent_loss + surface_loss + lap_loss
# saving losses for printing
track_edge += edge_loss / batch_size
track_lap += lap_loss / batch_size
track_lat += latent_loss / batch_size
track_surface += surface_loss / batch_size
loss /= batch_size
loss.backward()
optimizer.step()
print 'Epoch: {:d}'.format(epoch),
print ' Iter: {:d}/{:d}'.format(iteration, num_interations),
print ' loss: {:.4f}'.format(track_loss),
print ' surf_loss: {:.4f}'.format(track_surface),
print ' latent: {:.4f}'.format(latent_loss),
print ' e_loss: {:.4f}'.format(track_edge),
print ' l_loss: {:.4f}'.format(track_lap),
print ' size {:.4f}'.format(size/batch_size)
print ' time: {:.2f}s'.format(time() - t),
print ' best: {:.4f}'.format(best)
return real_distance / batch_size
# calculate validation loss
def val( A1, B1, C1, D1 ,A2, B2, C2, D2 , A3, B3, C3, D3 , mat,samps, img , i, init ):
with torch.no_grad():
vertex_positions = init
vertex_features = pooling([A1, B1, C1, D1 ], init, mat)
vertex_features, vertex_positions =modelA(init, vertex_features,adj_info)
vertex_positions = vertex_positions + init
splitting_faces = calc_curve(vertex_positions, adj_info)
adj_info_2,vertex_positions, vertex_features= split_info(adj_info, vertex_positions, vertex_features, splitting_faces, number = splitting_faces.shape[0])
vertex_features = torch.cat((vertex_features, pooling([A2, B2, C2, D2], vertex_positions, mat)), dim= 1 )
vertex_features, vertex_positions_ = modelB(vertex_positions, vertex_features, adj_info_2)
vertex_positions = vertex_positions + vertex_positions_
splitting_faces = calc_curve(vertex_positions, adj_info_2)
adj_info_3,vertex_positions, vertex_features= split_info(adj_info_2, vertex_positions, vertex_features, splitting_faces, number = splitting_faces.shape[0])
vertex_features = torch.cat((vertex_features, pooling([A3, B3, C3, D3], vertex_positions, mat)), dim= 1 )
_,vertex_positions_ = modelC(vertex_positions, vertex_features, adj_info_3)
vertex_positions = vertex_positions + vertex_positions_
loss, pred_points = point_to_point(vertex_positions, adj_info_3, samps[:2000] ,num = 2000 )
pred_points.detach()
loss.detach()
optimizer.zero_grad()
# to see output
if args.render:
active_faces_indecies = adj_info_3['face_list'][:,0,2].astype(int)
all_faces = adj_info_3['faces']
pred_face = all_faces[active_faces_indecies]
pred_faces = torch.LongTensor(pred_face).cuda()
render_mesh( vertex_positions.data.cpu().numpy(), pred_faces+1, img.permute(1,2,0).data.cpu().numpy())
raw_input()
return float(loss.data.cpu().numpy())
def validate(batch):
with torch.no_grad():
optimizer.zero_grad()
A1, B1, C1, D1 = encoder1(batch['imgs'])
A2, B2, C2, D2 = encoder2(batch['imgs'])
A3, B3, C3, D3 = encoder3(batch['imgs'])
A1.detach() , B1.detach(), C1.detach(), D1.detach()
A2.detach() , B2.detach(), C2.detach(), D2.detach()
A3.detach() , B3.detach(), C3.detach(), D3.detach()
full_info = zip( batch['mats'],batch['samples'], range(batch['mats'].shape[0]))
loss = 0
for mat,samps, i in full_info :
cur_loss = val(A1[i].clone(), B1[i].clone(), C1[i].clone(), D1[i].clone(),
A2[i].clone(), B2[i].clone(), C2[i].clone(), D2[i].clone(),
A3[i].clone(), B3[i].clone(), C3[i].clone(), D3[i].clone()
, mat,samps, batch['imgs'][i], i, initial_positions)
torch.cuda.empty_cache()
gc.collect()
loss = loss + cur_loss
valid_loss = loss/ (i+1)
return valid_loss
full_valid_loss, full_train_loss, length, min_val, real_val = [],[], 50, 1000, 0
valid_batch = data_valid.load_batch(2, ordered = True)
if args.render:
validate(valid_batch)
exit()
for epoch in range(3000):
if epoch == 1000:
for param_group in optimizer.param_groups:
param_group['lr'] = 0.00003
if epoch == 2000:
for param_group in optimizer.param_groups:
param_group['lr'] = 0.00001
loss_term = point_to_surface
for i in range(length):
full_train_loss.append(train(data_train, epoch, length, i, min_val))
v_loss = validate(valid_batch)
full_valid_loss.append(v_loss)
if full_valid_loss[-1] < min_val:
min_val = full_valid_loss[-1]
torch.save(modelA.state_dict(), checkpoint_dir + 'model_1')
torch.save(modelB.state_dict(), checkpoint_dir + 'model_2')
torch.save(modelC.state_dict(), checkpoint_dir + 'model_3')
torch.save(encoder1.state_dict(), checkpoint_dir + 'encoder_1')
torch.save(encoder2.state_dict(), checkpoint_dir + 'encoder_2')
torch.save(encoder3.state_dict(), checkpoint_dir + 'encoder_3')
if epoch>5:
graph(save_dir, full_train_loss[3:], full_valid_loss[3:])