-
Notifications
You must be signed in to change notification settings - Fork 6
/
train_FAUST.py
262 lines (180 loc) · 8.81 KB
/
train_FAUST.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
#!/usr/bin/env python
# coding: utf-8
import os
#os.environ["CUDA_VISIBLE_DEVICES"]="0"
import scipy.sparse as sp
import torch
import dill
import pickle
import torch.nn as nn
from random import shuffle
import torch_geometric
from torch_geometric.data.dataloader import DataLoader, DataListLoader
from misc_utils import *
from utils_distance import *
from datasets.faust_2500 import Faust2500Dataset
from models.model import PointNetVAE
device = 'cuda'
# import shutil
# shutil.rmtree('data/faust2500/processed')
dataset_train = Faust2500Dataset('data/faust2500')
#model options
opt = lambda x:x
opt.NUM_POINTS = dataset_train[0].pos.shape[0]
opt.BATCH_SIZE = 16
opt.DESC_SIZE = 512 #pointwise descriptro size after convlolutional layers
opt.LATENT_SPACE = 256 #dimension of the full (pose+style) latent space
opt.POSE_SIZE = 64 #number of dimension dedicated to pose encoding
opt.LOCAL_TH = 0.1
opt.LEARNING_RATE = 2e-5
## get a randomly selected samples inside the same randomly selected class
class IntraSampler(torch.utils.data.sampler.Sampler):
def __init__(self, dataset):
self.class_ids = dataset.class_ids
def __iter__(self):
class_idx = self.class_ids[np.random.randint(len(self.class_ids))]
class_idxs = [i for i,id in enumerate(self.class_ids) if id==class_idx]
shuffle(class_idxs)
return iter(class_idxs)
def __len__(self):
print ('\tcalling Sampler:__len__')
return self.num_samples
intra_loader = DataListLoader(dataset_train, batch_size=2, sampler = IntraSampler(dataset_train))
interp_loader = DataListLoader(dataset_train, batch_size=2, shuffle=True)
rec_loader = DataListLoader(dataset_train, batch_size=opt.BATCH_SIZE, shuffle=True, num_workers=0)
#DEFINE 3 LOSSES
#optimize gloabl euclidean distortions on data points
def optimize_rec(vae, batch_loader):
vae.enable_bn(True)
data = batch_loader.__iter__().next()
x = torch.stack([d.pos for d in data]).to(device)
rec = vae(x)
De_t = pairwise_dists(x)
De_r = pairwise_dists(rec)
loss1 = torch.mean((De_t-De_r)**2) #L1 loss on the absoulute euclidean distortion
# loss2 = torch.mean( ((De_t-De_r)/(De_t+1e-3) )**2 ) #L2 loss on the relative euclidean distortion
loss = loss1#+loss2
kl_loss = 0.5 * torch.mean(torch.exp(vae.z_var) + vae.z_mu**2 - 1.0 - vae.z_var)
loss += 1e-0*kl_loss
return loss
def optimize_intep(vae, batch_loader, opt, geoloss=1e1, eucloss=1e1):
vae.enable_bn(False)
data = batch_loader.__iter__().next()
# data = data.to(device)
T = data[0].faces[None,...].to(device)
x = torch.stack([d.pos for d in data]).to(device)
orix = torch.stack([d.oripos for d in data]).to(device)
Dg = torch.stack([d.Dg for d in data]).to(device)
De = calc_euclidean_dist_matrix(orix)
latent = vae.enc(x)[...,:vae.opt.LATENT_SPACE]
a = torch.rand(1).to(device)
Dg_t = Dg[0]*a+(1-a)*Dg[1]
De_t = De[0]*a+(1-a)*De[1]
latent_t = latent[0:1]*a + (1-a)*latent[1:2]
rec = vae.dec(latent_t)
loss1=torch.zeros(1).to(device)
loss2=torch.zeros(1).to(device)
if geoloss>0:
Dg_r, grad, div, W, S, C = distance_GIH(rec, T)
loss1 = geoloss*torch.mean( ((Dg_t-Dg_r.float()))**2)
if eucloss>0:
localmask = (Dg_t<opt.LOCAL_TH*Dg_t.max()).float()
De_r = calc_euclidean_dist_matrix(rec)
loss2 = eucloss*torch.sum( localmask*((De_t-De_r)/(De_t+1e-3))**2 )/torch.sum(localmask)
loss=loss1 + loss2
return loss
def optimize_disent_ext(vae, batch_loader, opt, geoloss=1e1):
vae.enable_bn(False)
data = batch_loader.__iter__().next()
# data = data.to(device)
T = data[0].faces[None,...].to(device)
x = torch.stack([d.pos for d in data]).to(device)
orix = torch.stack([d.oripos for d in data]).to(device)
Dg = torch.stack([d.Dg for d in data]).to(device)
De = calc_euclidean_dist_matrix(orix)
latent = vae.enc(x)[...,:vae.opt.LATENT_SPACE]
a = torch.rand(1).to(device)
Dg_t = Dg[0]
loss1=torch.zeros(1).to(device)
loss2=torch.zeros(1).to(device)
localmask = (Dg_t<opt.LOCAL_TH*Dg_t.max()).float()
# Interpolationg only the pose latent vector of different subjects should preserve geodesic distances
latent_t = torch.cat([latent[0:1,:opt.POSE_SIZE]*a + (1-a)*latent[1:2,:opt.POSE_SIZE], latent[0:1,opt.POSE_SIZE:]],-1)
rec = vae.dec(latent_t)
Dg_r, grad, div, W, S, C = distance_GIH(rec, T)
loss = geoloss*torch.mean( ((Dg_t-Dg_r.float()))**2)
return loss
def optimize_disent_int(vae, batch_loader, opt, geoloss=1e0, eucloss=1e0):
vae.enable_bn(False)
data = batch_loader.__iter__().next()
# data = data.to(device)
T = data[0].faces[None,...].to(device)
x = torch.stack([d.pos for d in data]).to(device)
orix = torch.stack([d.oripos for d in data]).to(device)
Dg = torch.stack([d.Dg for d in data]).to(device)
De = calc_euclidean_dist_matrix(orix)
latent = vae.enc(x)[...,:vae.opt.LATENT_SPACE]
a = torch.rand(1).to(device)
Dg_t = Dg[0]
loss1=torch.zeros(1).cuda()
loss2=torch.zeros(1).cuda()
localmask = (Dg_t<opt.LOCAL_TH*Dg_t.max()).float()
# Interpolationg only the style latent vector of the same subject should not result in changes on it's embedding
latent_t = torch.cat([latent[0:1,:opt.POSE_SIZE], latent[0:1,opt.POSE_SIZE:]*a + (1-a)*latent[1:2,opt.POSE_SIZE:]],-1)
rec = vae.dec(latent_t)
De_r = calc_euclidean_dist_matrix(rec)
loss = eucloss*torch.mean( ((De[0:1]-De_r)/(De[0:1]+1e-3) )**2 ) #L2 loss on the relative euclidean distortion
return loss
# ######### Optimization procedure
# # 0-7000: we start optimizing only for global euclidean distortion on data points
# # to avoid local minima introduced by local error metrics
# # 7000-10000: we now start optimizing also for linear interpolation and disentanglement of latent codes. Global
# # euclidean distortion is not the ideal loss for itnerpolating shapes, we instead use
# # local euclidean metric. We will also use geodesic metric in the next optimization stage.
# # 10000-15000: now that shapes are clean also in the interpolated latent space we can safely
# # compute geodesic distances (GIH) and use them in the loss terms for interpolation and disentanglement.
# # 15000-: during the last iterations we favor local euclidean loss to remove wrinkles
# # introduced by geodesics preservation (due to the non exact isometries of pose changes)
NUM_ITERATIONS = 20000
virtual_batch_size = 20 #multiplier applied to the interations number (should be adapted to the size and complexity of the dataset)
vae = PointNetVAE(opt).to(device)
i=0
optimizer = torch.optim.Adam(vae.parameters(), lr=opt.LEARNING_RATE)
total_loss=0
losses = []
t=time.time()
for i in range(i,NUM_ITERATIONS+1):
for inner_it in range(virtual_batch_size):
optimizer.zero_grad()
loss1 = loss2 = loss3 = loss4 = torch.zeros((1,)).float().to(device)
loss1 = optimize_rec(vae,rec_loader)
if i>7000 and i<= 10000: #geodesic weight=0, local_euclidean=1e1
loss2 = optimize_intep(vae,interp_loader, opt, 0, 1e1)
loss3 = optimize_disent_int(vae,intra_loader, opt, 0,1e1)
if i>10000 and i<= 15000: #geodesic weight=1e-2, local_euclidean=1e2, dis_euclidean=1e1
loss2 = optimize_intep(vae,interp_loader, opt, 1e-2,1e2)
loss3 = optimize_disent_int(vae,intra_loader, opt, 1e-2,1e1)
loss4 = optimize_disent_ext(vae,interp_loader, opt, 1e1)
if i>15000: #geodesic weight=1e-3, local_euclidean=1e1, dis_euclidean=1e0
loss2 = optimize_intep(vae,interp_loader, opt,1e-3, 1e2)
loss3 = optimize_disent_int(vae,intra_loader, opt, 1e-3,1e0)
loss4 = optimize_disent_ext(vae,interp_loader, opt, 1e0)
loss = loss1 + 1e-1*loss2 + 1e1*(1e0*loss3 + 1e0*loss4)
loss.backward()
optimizer.step()
losses.append([loss.item(),loss1.item(),loss2.item(),loss3.item(),loss4.item()])
if i%10==0:
avg_loss = np.mean(losses[-50:],0)
print('%d] loss: %.2e (%.2e, %.2e, %.2e, %.2e), time: %.2f s' %
(i, avg_loss[0],avg_loss[1],avg_loss[2],avg_loss[3],avg_loss[4], time.time()-t))
t=time.time()
#network weights checkpoints
os.makedirs('trained',exist_ok=True)
prefix = 'FAUST'
if i%100 ==0 :
if i<=3000:
torch.save(vae.state_dict(), 'trained/'+prefix+'_vae.dict')
elif i <= 4000:
torch.save(vae.state_dict(), 'trained/'+prefix+'_vae_euc.dict')
else:
torch.save(vae.state_dict(), 'trained/'+prefix+'_vae_euc_gd.dict')