-
Notifications
You must be signed in to change notification settings - Fork 30
/
Step4_UNet_MoFA_EM.py
439 lines (343 loc) · 16.4 KB
/
Step4_UNet_MoFA_EM.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
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Tue Feb 23 13:12:25 2021
UNet + MoFA
@author: li0005
"""
import os
import torch
import math
import torch.optim as optim
import util.util as util
import csv
import util.load_dataset as load_dataset
import util.load_object as lob
import renderer.rendering as ren
import encoder.encoder as enc
import time
import UNet.UNet as unet
import argparse
from datetime import date
import util.advanced_losses as adlosses
from facenet_pytorch import InceptionResnetV1
from models import networks
import pickle
#hyper-parameters
par = argparse.ArgumentParser(description='MoFA')
par.add_argument('--learning_rate',default=0.1,type=float,help='The learning rate')
par.add_argument('--epochs',default=130,type=int,help='Total epochs')
par.add_argument('--batch_size',default=12,type=int,help='Batch sizes')
par.add_argument('--gpu',default=0,type=int,help='The GPU ID')
par.add_argument('--pretrained_model',default=00,type=int,help='Pretrained model')
par.add_argument('--img_path',type=str,help='Root of the training samples')
args = par.parse_args()
GPU_no = args.gpu
dist_weight={'neighbour':15,'dist': 3,'area':0.5,'preserve':0.25,'binary': 10}
ct = args.pretrained_model #load trained mofa model
output_name = 'MoFA_UNet'
device = torch.device("cuda:{}".format(util.device_ids[GPU_no]) if torch.cuda.is_available() else "cpu")
begin_learning_rate = args.learning_rate
decay_step_size=5000
decay_rate_gamma =0.99
decay_rate_unet =0.95
learning_rate_begin=begin_learning_rate*(decay_rate_gamma ** ((300000)//decay_step_size)) *0.8
mofa_lr_begin = learning_rate_begin * (decay_rate_gamma ** (ct//decay_step_size))
unet_lr_begin = learning_rate_begin *0.06* (decay_rate_unet**(ct//decay_step_size))
ct_begin=ct
today = date.today()
current_path = os.getcwd()
model_path = current_path+'/basel_3DMM/model2017-1_bfm_nomouth.h5'
image_path = (args.img_path + '/' ).replace('//','/')
output_path = current_path+'/MoFA_UNet_Save/robustness/'+output_name + '/'
if not os.path.exists(output_path):
os.makedirs(output_path)
loss_log_path_train = output_path+today.strftime("%b-%d-%Y")+"loss_train.csv"
loss_log_path_test = output_path+today.strftime("%b-%d-%Y")+"loss_test.csv"
weight_log_path_train = output_path+today.strftime("%b-%d-%Y")+"weight_train.pkl"
with open(weight_log_path_train, 'wb') as f:
pickle.dump(dist_weight, f)
'''------------------
Prepare Log Files
------------------'''
if ct != 0:
try:
fid_train = open(loss_log_path_train, 'a')
fid_test = open(loss_log_path_test, 'a')
except:
fid_train = open(loss_log_path_train, 'w')
fid_test = open(loss_log_path_test, 'w')
else:
fid_train = open(loss_log_path_train, 'w')
fid_test = open(loss_log_path_test, 'w')
writer_train = csv.writer(fid_train, lineterminator="\r\n")
writer_test = csv.writer(fid_test, lineterminator="\r\n")
'''------------------
Load Data & Models
------------------'''
#parameters
batch = args.batch_size
width = 224
height = 224
epoch = args.epochs
test_batch_num = 5
cos = torch.nn.CosineSimilarity(dim=1, eps=1e-6)
'''-------------
Load Dataset
-------------'''
testset = load_dataset.CelebDataset(device,image_path, False, height,width,1)
testloader = torch.utils.data.DataLoader(testset, batch_size=batch,shuffle=False, num_workers=0)
# 3dmm data
obj = lob.Object3DMM(model_path,device,is_crop = True)
A = torch.Tensor([[9.06*224/2, 0, (width-1)/2.0, 0, 9.06*224/2, (height-1)/2.0, 0, 0, 1]]).view(-1, 3, 3).to(device) #intrinsic camera mat
T_ini = torch.Tensor([0, 0, 1000]).to(device) #camera translation(direction of conversion will be set by flg later)
sh_ini = torch.zeros(3, 9,device=device) #offset of spherical harmonics coefficient
sh_ini[:, 0] = 0.7 * 2 * math.pi
sh_ini = sh_ini.reshape(-1)
'''----------------
Prepare Network and Optimizer
----------------'''
#renderer and encoder and UNet
render_net = ren.Renderer(32) #block_size^2 pixels are simultaneously processed in renderer, lager block_size consumes lager memory
enc_net = enc.FaceEncoder(obj).to(device)
cos = torch.nn.CosineSimilarity(dim=1, eps=1e-6)
net_recog = networks.define_net_recog(net_recog='r50', pretrained_path='models/ms1mv3_arcface_r50_fp16/backbone.pth')
net_recog = net_recog.to(device)
'''----------------------------------
Fixed Testing Images for Observation
----------------------------------'''
test_input_images = []
test_landmarks = []
test_landmark_masks = []
for i_test, data_test in enumerate(testloader, 0):
if i_test >= test_batch_num:
break
images, landmarks = data_test
test_input_images +=[images]
test_landmarks +=[landmarks]
util.write_tiled_image(torch.cat(test_input_images,dim=0), output_path + 'test_gt.png',10)
def occlusionPhotometricLossWithoutBackground(gt,rendered,fgmask,standardDeviation=0.043,backgroundStDevsFromMean=3.0):
normalizer = (-3 / 2 * math.log(2 * math.pi) - 3 * math.log(standardDeviation))
fullForegroundLogLikelihood = (torch.sum(torch.pow(gt - rendered,2), axis=1)) * -0.5 / standardDeviation / standardDeviation + normalizer
uniformBackgroundLogLikelihood = math.pow(backgroundStDevsFromMean * standardDeviation, 2) * -0.5 / standardDeviation / standardDeviation + normalizer
occlusionForegroundMask = fgmask * (fullForegroundLogLikelihood > uniformBackgroundLogLikelihood).type(torch.FloatTensor).cuda(util.device_ids[GPU_no])
foregroundLogLikelihood = occlusionForegroundMask*fullForegroundLogLikelihood
lh = torch.mean(foregroundLogLikelihood)
return -lh, occlusionForegroundMask
'''-------------
Network Forward
-------------'''
#################################################################
def proc_mofaunet(images, landmarks, render_mode,train_net=False,occlusion_mode=False, valid_mask=None,image_org=None,is_cutmix_mode=False):
#valid_mask: 1 indicating unoccluded part of faces, vice versa
'''
images: network_input
landmarks: landmark ground truth
render_mode: renderer mode
occlusion mode: use occlusion robust loss or not
landmark_vmask: landmark valid mask
valid_mask: use the valid region
image_org: if is supervised mode, image_org
'''
shape_param, exp_param, color_param, camera_param, sh_param = enc_net(images)
color_param *= 3 #adjust learning rate
camera_param[:,:3] *= 0.3
camera_param[:,5] *= 0.005
shape_param[:,80:] *= 0 #ignore high dimensional component of BFM
exp_param[:,64:] *= 0
color_param[:,80:] *= 0
#convert parameters to mesh, camera, and lighting
#If camera2world=False, generated R and T are converstion from world to camera, which might be different from MoFA paper formulation)
vertex, color, R, T, sh_coef = enc.convert_params(shape_param, exp_param, color_param, camera_param, sh_param,obj,T_ini,sh_ini,False)
projected_vertex, sampled_color, shaded_color, occlusion, raster_image, raster_mask = render_net(obj.face, vertex,
color, sh_coef, A,
R, T, images,ren.RASTERIZE_DIFFERENTIABLE_IMAGE,False, 5, True)
'''-------------------------------------
U-Net input: Raster [RGB] + ORG [RGB]
----------------------------------------'''
lm68 = projected_vertex[:,0:2,obj.landmark]
image_concatenated = torch.cat(( raster_image,images),axis = 1)
unet_est_mask = unet_for_mask(image_concatenated)
valid_loss_mask = raster_mask.unsqueeze(1)*unet_est_mask
masked_rec_loss = torch.mean(torch.sum(torch.norm(valid_loss_mask*(images - raster_image), 2, 1)) / torch.clamp(torch.sum(raster_mask.unsqueeze(1)*unet_est_mask),min=1))
bg_unet_loss = torch.mean(torch.sum(raster_mask.unsqueeze(1)*(1-unet_est_mask),axis=[2,3])/torch.clamp(torch.sum(raster_mask.unsqueeze(1),axis=[2,3]),min=1))#area loss
mask_binary_loss= torch.zeros([1])
perceptual_loss= torch.zeros([1])
land_loss= torch.zeros([1])
stat_reg= torch.zeros([1])
if train_net=='unet':
mask_binary_loss= (0.5- torch.mean(torch.norm(valid_loss_mask-0.5,2,1)))
loss_unet = mask_binary_loss*dist_weight['binary'] + bg_unet_loss*dist_weight['area']
if train_net == 'mofa':
pred_feat = net_recog(image=raster_image,pred_lm=lm68.transpose(1,2))
gt_feat = net_recog(images,landmarks.transpose(1,2))
cosine_d = torch.sum(pred_feat * gt_feat, dim=-1)
perceptual_loss = torch.sum(1 - cosine_d) / cosine_d.shape[0]
land_loss = torch.mean((obj.weight_lm*(landmarks-lm68))**2)
stat_reg = (torch.sum(shape_param ** 2) + torch.sum(exp_param ** 2) + torch.sum(color_param ** 2))/float(batch)/224.0
loss_mofa = masked_rec_loss*0.5 + perceptual_loss*0.25+ 1e-1 * stat_reg + 5e-4 * land_loss +6e-2* bg_unet_loss
if train_net == False:
mask_binary_loss= (0.5- torch.mean(torch.norm(valid_loss_mask-0.5,2,1)))
pred_feat = net_recog(image=raster_image,pred_lm=lm68.transpose(1,2))
gt_feat = net_recog(images,landmarks.transpose(1,2))
cosine_d = torch.sum(pred_feat * gt_feat, dim=-1)
perceptual_loss = torch.sum(1 - cosine_d) / cosine_d.shape[0]
land_loss = torch.mean((obj.weight_lm*(landmarks-projected_vertex[:,0:2,obj.landmark]))**2)
stat_reg = (torch.sum(shape_param ** 2) + torch.sum(exp_param ** 2) + torch.sum(color_param ** 2))/float(batch)/224.0
loss_test = mask_binary_loss*dist_weight['binary'] + masked_rec_loss*0.5 + bg_unet_loss*dist_weight['area'] + perceptual_loss*0.25+ 1e-1 * stat_reg + 5e-4 * land_loss
I_target_masked=images* valid_loss_mask
id_target_masked = net_recog(I_target_masked, landmarks.transpose(1,2),is_shallow =True)
id_target = net_recog(images, landmarks.transpose(1,2),is_shallow =True)
id_reconstruct_masked = net_recog(raster_image* valid_loss_mask , pred_lm=lm68.transpose(1,2),is_shallow =True)
I_IM_Per_loss = torch.mean(1-cos(id_target, id_target_masked))
IRM_IM_Per_loss = torch.mean(1-cos(id_reconstruct_masked, id_target_masked))
if train_net=='unet':
loss_unet += I_IM_Per_loss*dist_weight['preserve'] + IRM_IM_Per_loss*dist_weight['dist']
if train_net == False:
loss_test += I_IM_Per_loss*dist_weight['preserve'] + IRM_IM_Per_loss*dist_weight['dist']
#force it to be binary mask
loss_mask_neighbor = torch.zeros([1])
if train_net=='unet':
loss_mask_neighbor = adlosses.neighbor_unet_loss(images, valid_loss_mask, raster_image)
loss_unet +=loss_mask_neighbor*dist_weight['neighbour']
loss=loss_unet
if train_net == 'mofa':
loss=loss_mofa
if train_net == False:
loss=loss_test
losses_return = torch.FloatTensor([loss.item(),land_loss.item(),masked_rec_loss.item(),stat_reg.item(),bg_unet_loss.item(),\
perceptual_loss.item(),I_IM_Per_loss.item(),IRM_IM_Per_loss.item(),loss_mask_neighbor.item(),mask_binary_loss.item()])
if train_net=='unet':
return loss_unet, losses_return, raster_image,raster_mask,unet_est_mask, valid_loss_mask
if train_net == 'mofa':
return loss_mofa, losses_return, raster_image,raster_mask,unet_est_mask, valid_loss_mask
if train_net == False:
return loss_test, losses_return, raster_image,raster_mask,unet_est_mask, valid_loss_mask
#################################################################
'''-----------------------------------------
load pretrained model and continue training
-----------------------------------------'''
if ct!=0:
trained_model_path = output_path+'enc_net_{:06d}.model'.format(ct)
enc_net = torch.load(trained_model_path, map_location='cuda:{}'.format(util.device_ids[GPU_no]))
unet_model_path = output_path+'unet_{:06d}.model'.format(ct)
unet_for_mask = torch.load(unet_model_path, map_location='cuda:{}'.format(util.device_ids[GPU_no]))
else:
trained_model_path = current_path+'/MoFA_UNet_Save/pretrain_mofa'+'/enc_net_300000.model'
enc_net = torch.load(trained_model_path, map_location='cuda:{}'.format(util.device_ids[GPU_no]))
unet_model_path = current_path+'/MoFA_UNet_Save/Pretrain_UNet'+'/unet_mask_030000.model'
unet_for_mask = torch.load(unet_model_path, map_location='cuda:{}'.format(util.device_ids[GPU_no]))
print('Loading pre-trained unet: \n'+trained_model_path +'\n' +unet_model_path)
'''----------
Set Optimizer
----------'''
optimizer_mofa = optim.Adadelta(enc_net.parameters(), lr=mofa_lr_begin)
optimizer_unet = optim.Adadelta(unet_for_mask.parameters(), lr=unet_lr_begin)
scheduler_mofa = torch.optim.lr_scheduler.StepLR(optimizer_mofa,step_size=decay_step_size,gamma=0.99)
scheduler_unet = torch.optim.lr_scheduler.StepLR(optimizer_unet,step_size=decay_step_size,gamma=decay_rate_unet)
print('Training ...')
start = time.time()
mean_losses_mofa = torch.zeros([10])
mean_losses_unet = torch.zeros([10])
trainset = load_dataset.CelebDataset(device,image_path, True, height,width,1)
def count_parameters(model):
return sum(p.numel() for p in model.parameters() if p.requires_grad)
for ep in range(0,epoch):
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch,shuffle=True, num_workers=0)
for i, data in enumerate(trainloader, 0):
'''-------------------------
Save images for observation
--------------------------'''
a=count_parameters(enc_net)
b=count_parameters(unet_for_mask)
if (ct-ct_begin) % 1000 == 0:
enc_net.eval()
unet_for_mask.eval()
test_raster_images = []
valid_loss_mask_temp = []
for images, landmarks in zip(test_input_images,test_landmarks ):
with torch.no_grad():
_, _, raster_image,raster_mask,fg_mask, valid_loss_mask = proc_mofaunet(images,landmarks,True,False)
test_raster_images += [images*(1-raster_mask.unsqueeze(1))+raster_image*raster_mask.unsqueeze(1)]
valid_loss_mask_temp += [valid_loss_mask]
util.write_tiled_image(torch.cat(test_raster_images,dim=0),output_path+'test_image_{}.png'.format(ct),10)
util.write_tiled_image(torch.cat(valid_loss_mask_temp , dim=0), output_path + 'valid_loss_mask_{}.png'.format(ct),10)
'''-------------------------
Save Model every 5000 iters
--------------------------'''
if (ct-ct_begin) % 10000 ==0 and ct>ct_begin:
torch.save(enc_net, output_path + 'enc_net_{:06d}.model'.format(ct))
torch.save(unet_for_mask, output_path + 'unet_{:06d}.model'.format(ct))
#validating
'''-------------------------
Validate Model every 1000 iters
--------------------------'''
if (ct-ct_begin) % 5000 == 0 and ct>ct_begin:
print('Training mode:'+output_name)
c_test=0
mean_test_losses = torch.zeros([10])
for i_test, data_test in enumerate(testloader,0):
image, landmark = data_test
c_test+=1
with torch.no_grad():
loss_, losses_return_,_, _,_,_ = proc_mofaunet(image,landmark,True,False)
mean_test_losses += losses_return_
mean_test_losses = mean_test_losses/c_test
str = 'test loss:{}'.format(ct)
for loss_temp in losses_return_:
str+=' {:05f}'.format(loss_temp)
print(str)
writer_test.writerow(str)
fid_train.close()
fid_train = open(loss_log_path_train , 'a')
writer_train = csv.writer(fid_train, lineterminator="\r\n")
fid_test.close()
fid_test = open(loss_log_path_test, 'a')
writer_test = csv.writer(fid_test, lineterminator="\r\n")
'''-------------------------
Model Training
--------------------------'''
images, landmarks = data
if images.shape[0]!=batch:
continue
if ct %30000 >5000:
enc_net.train()
unet_for_mask.eval()
loss_mofa, losses_return_mofa, _,_,_,_= proc_mofaunet(images,landmarks,True,'mofa')
loss_mofa.backward()
optimizer_mofa.step()
mean_losses_mofa+= losses_return_mofa
#optimizer_mofa.zero_grad()
else:
unet_for_mask.train()
enc_net.eval()
loss_unet, losses_return_unet, _,_,_,_= proc_mofaunet(images,landmarks,True,'unet')
loss_unet.backward()
optimizer_unet.step()
#optimizer_unet.zero_grad()
mean_losses_unet+= losses_return_unet
ct += 1
scheduler_unet.step()
scheduler_mofa.step()
optimizer_unet.zero_grad()
optimizer_mofa.zero_grad()
'''-------------------------
Show Training Loss
--------------------------'''
if (ct-ct_begin) % 100 == 0 and ct>ct_begin:
end = time.time()
mean_losses_unet = mean_losses_unet/100
mean_losses_mofa = mean_losses_mofa/100
str = 'mofa loss:{}'.format(ct)
for loss_temp in mean_losses_mofa:
str+=' {:05f}'.format(loss_temp)
str += '\nunet loss:{}'.format(ct)
for loss_temp in mean_losses_unet:
str+=' {:05f}'.format(loss_temp)
str += ' time: {}'.format(end-start)
print(str)
writer_train.writerow(str)
start = time.time()
mean_losses_unet = torch.zeros([10])
mean_losses_mofa = torch.zeros([10])
torch.save(enc_net, output_path + 'enc_net_{:06d}.model'.format(ct))