-
Notifications
You must be signed in to change notification settings - Fork 68
/
train_landmarks_generator.py
337 lines (308 loc) · 15.6 KB
/
train_landmarks_generator.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
import os
from tqdm import tqdm
import torch
import numpy as np
from glob import glob
from os.path import join, isfile
import random
from tensorboardX import SummaryWriter
from models import Landmark_generator as Landmark_transformer
import argparse
parser=argparse.ArgumentParser()
parser.add_argument('--pre_audio_root',default='...../Dataset/lrs2_preprocessed_audio',
help='root path for preprocessed audio')
parser.add_argument('--landmarks_root',default='...../Dataset/lrs2_landmarks',
help='root path for preprocessed landmarks')
args=parser.parse_args()
#network parameters
d_model=512
dim_feedforward=1024
nlayers=4
nhead=4
dropout=0.1 # 0.5
Nl=15
T = 5
Project_name = 'landmarkT5_d512_fe1024_lay4_head4'
print('Project_name:', Project_name)
finetune_path =None
num_workers = 8
batch_size =128 # 512
batch_size_val =128 #512
evaluate_interval = 5000 #
checkpoint_interval = evaluate_interval
mel_step_size = 16
fps = 25
lr = 1e-4
global_step, global_epoch = 0, 0
landmark_root=args.landmarks_root
filelist_name = 'lrs2'
checkpoint_root = './checkpoints/landmark_generation/'
checkpoint_dir = os.path.join(checkpoint_root, 'Pro_' + Project_name)
reset_optimizer = False
save_optimizer_state = True
writer = SummaryWriter('tensorboard_runs/Project_{}'.format(Project_name))
#we arrange the landmarks in some order
ori_sequence_idx=[162,127,234,93,132,58,172,136,150,149,176,148,152,377,400,378,379,365,397,288,361,323,454,356,389, #
70,63,105,66,107,55,65,52,53,46,#
336,296,334,293,300,276,283,282,295,285,#
168,6,197,195,5,#
48,115,220,45,4,275,440,344,278,#
33,246,161,160,159,158,157,173,133,155,154,153,145,144,163,7,#
362,398,384,385,386,387,388,466,263,249,390,373,374,380,381,382,#
61,185,40,39,37,0,267,269,270,409,291,375,321,405,314,17,84,181,91,146,#
78,191,80,81,82,13,312,311,310,415,308,324,318,402,317,14,87,178,88,95]
full_face_sequence=[*list(range(0, 4)), *list(range(21, 25)), *list(range(25, 91)), *list(range(4, 21)), *list(range(91, 131))]
class LandmarkDict(dict):
def __init__(self, idx, x, y):
self['idx'] = idx
self['x'] = x
self['y'] = y
def __getattr__(self, name):
try:
return self[name]
except:
raise AttributeError(name)
def __setattr__(self, name, value):
self[name] = value
class Dataset(object):
def get_vidname_list(self, split):
vid_name_list = []
with open('filelists/{}/{}.txt'.format(filelist_name, split)) as f:
for line in f:
line = line.strip()
if ' ' in line: line = line.split()[0]
vid_name_list.append(line)
return vid_name_list
def __init__(self, split):
min_len = 25 #filter videos that is too short
vid_name_lists = self.get_vidname_list(split)
self.all_video_names = []
print("init dataset,filtering very short videos.....")
for vid_name in tqdm(vid_name_lists, total=len(vid_name_lists)):
pkl_paths = list(glob(join(landmark_root,vid_name, '*.npy')))
vid_len=len(pkl_paths)
if vid_len >= min_len:
self.all_video_names.append((vid_name, vid_len))
print("complete,with available vids: ", len(self.all_video_names), '\n')
def __len__(self):
return len(self.all_video_names)
def __getitem__(self, idx):
while 1:
vid_idx = random.randint(0, len(self.all_video_names) - 1)
vid_name = self.all_video_names[vid_idx][0]
vid_len=self.all_video_names[vid_idx][1]
# 00.randomly select a window of T video frames
random_start_idx = random.randint(2, vid_len - T - 2)
T_idxs = list(range(random_start_idx, random_start_idx + T))
# 01. get reference landmarks
all_list=[i for i in range(vid_len) if i not in T_idxs]
Nl_idxs = random.sample(all_list, Nl)
Nl_landmarks_paths = [os.path.join(landmark_root, vid_name, str(idx) + '.npy') for idx in Nl_idxs]
Nl_pose_landmarks,Nl_content_landmarks= [],[]
for frame_landmark_path in Nl_landmarks_paths:
if not os.path.exists(frame_landmark_path):
break
landmarks=np.load(frame_landmark_path,allow_pickle=True).item()
Nl_pose_landmarks.append(landmarks['pose_landmarks'])
Nl_content_landmarks.append(landmarks['content_landmarks'])
if len(Nl_pose_landmarks) != Nl:
continue
Nl_pose = torch.zeros((Nl, 2, 74)) # 74 landmark
Nl_content = torch.zeros((Nl, 2, 57)) # 57 landmark
for idx in range(Nl):
Nl_pose_landmarks[idx] = sorted(Nl_pose_landmarks[idx],
key=lambda land_tuple: ori_sequence_idx.index(land_tuple[0]))
Nl_content_landmarks[idx] = sorted(Nl_content_landmarks[idx],
key=lambda land_tuple: ori_sequence_idx.index(land_tuple[0]))
Nl_pose[idx, 0, :] = torch.FloatTensor(
[Nl_pose_landmarks[idx][i][1] for i in range(len(Nl_pose_landmarks[idx]))]) # x
Nl_pose[idx, 1, :] = torch.FloatTensor(
[Nl_pose_landmarks[idx][i][2] for i in range(len(Nl_pose_landmarks[idx]))]) # y
Nl_content[idx, 0, :] = torch.FloatTensor(
[Nl_content_landmarks[idx][i][1] for i in range(len(Nl_content_landmarks[idx]))]) # x
Nl_content[idx, 1, :] = torch.FloatTensor(
[Nl_content_landmarks[idx][i][2] for i in range(len(Nl_content_landmarks[idx]))]) # y
# 02. get T pose landmark and content landmark
T_ladnmark_paths = [os.path.join(landmark_root, vid_name, str(idx) + '.npy') for idx in T_idxs]
T_pose_landmarks,T_content_landmarks=[],[]
for frame_landmark_path in T_ladnmark_paths:
if not os.path.exists(frame_landmark_path):
break
landmarks=np.load(frame_landmark_path,allow_pickle=True).item()
T_pose_landmarks.append(landmarks['pose_landmarks'])
T_content_landmarks.append(landmarks['content_landmarks'])
if len(T_pose_landmarks)!=T:
continue
T_pose=torch.zeros((T,2,74)) #74 landmark
T_content=torch.zeros((T,2,57)) #57 landmark
for idx in range(T):
T_pose_landmarks[idx]=sorted(T_pose_landmarks[idx],key=lambda land_tuple:ori_sequence_idx.index(land_tuple[0]))
T_content_landmarks[idx] = sorted(T_content_landmarks[idx],key=lambda land_tuple: ori_sequence_idx.index(land_tuple[0]))
T_pose[idx,0,:]=torch.FloatTensor([T_pose_landmarks[idx][i][1] for i in range(len(T_pose_landmarks[idx]))] ) #x
T_pose[idx,1,:]=torch.FloatTensor([T_pose_landmarks[idx][i][2] for i in range(len(T_pose_landmarks[idx]))]) #y
T_content[idx, 0, :] = torch.FloatTensor([T_content_landmarks[idx][i][1] for i in range(len(T_content_landmarks[idx]))]) # x
T_content[idx, 1, :] = torch.FloatTensor([T_content_landmarks[idx][i][2] for i in range(len(T_content_landmarks[idx]))]) # y
# 03. get T audio
try:
audio_mel = np.load(join(args.pre_audio_root,vid_name, "audio.npy"))
except Exception as e:
continue
T_mels = []
for frame_idx in T_idxs:
mel_start_frame_idx = frame_idx - 2 ###around the frame
if mel_start_frame_idx < 0:
break
start_idx = int(80. * (mel_start_frame_idx / float(fps)))
m = audio_mel[start_idx: start_idx + mel_step_size, :] # get five frames around
if m.shape[0] != mel_step_size: # in the end of vid
break
T_mels.append(m.T) # transpose
if len(T_mels) != T:
continue
T_mels = np.asarray(T_mels) # (T,hv,wv)
T_mels = torch.FloatTensor(T_mels).unsqueeze(1) # (T,1,hv,wv)
# return value
return T_mels, T_pose, T_content,Nl_pose,Nl_content
# (T,1,hv,wv) (T,2,74) (T,2,57)
def load_checkpoint(path, model, optimizer, reset_optimizer=False, overwrite_global_states=True):
global global_step
global global_epoch
print("Load checkpoint from: {}".format(path))
checkpoint = torch.load(path)
s = checkpoint["state_dict"]
new_s = {}
for k, v in s.items():
new_s[k.replace('module.', '')] = v
# for k, v in s.items():
# new_s['module.'+k] = v
model.load_state_dict(new_s)
if not reset_optimizer:
optimizer_state = checkpoint["optimizer"]
if optimizer_state is not None:
print("Load optimizer state from {}".format(path))
optimizer.load_state_dict(checkpoint["optimizer"])
if overwrite_global_states:
global_step = checkpoint["global_step"]
global_epoch = checkpoint["global_epoch"]
return model
def save_checkpoint(model, optimizer, step, checkpoint_dir, epoch, prefix=''):
checkpoint_path = join(
checkpoint_dir, "{}_epoch_{}_checkpoint_step{:09d}.pth".format(prefix, epoch, global_step))
if isfile(checkpoint_path):
os.remove(checkpoint_path)
optimizer_state = optimizer.state_dict() if save_optimizer_state else None
torch.save({
"state_dict": model.state_dict(),
"optimizer": optimizer_state,
"global_step": step,
"global_epoch": epoch,
}, checkpoint_path)
print("Saved checkpoint:", checkpoint_path)
criterion_L1 = torch.nn.L1Loss()
def get_velocity_loss (pred, gt): #(B*T,2,57) (B*T,2,57)
pred=torch.stack(torch.split(pred,T,dim=0),dim=0) #(B,T,2,57)
gt = torch.stack(torch.split(gt, T, dim=0), dim=0) # (B,T,2,57)
pred=torch.cat([pred[:,:,:,i] for i in range(pred.size(3))],dim=2) #(B,T,57*2)
gt = torch.cat([gt[:, :, :, i] for i in range(gt.size(3))], dim=2) # (B,T,57*2)
b, t, c = pred.shape
pred_spiky = pred[:, 1:, :] - pred[:, :-1, :] #
gt_spiky = gt[:, 1:, :] - gt[:, :-1, :]
pred_spiky = pred_spiky.view(b * (t - 1), c)
gt_spiky = gt_spiky.view(b * (t - 1), c)
pairwise_distance = torch.nn.functional.pairwise_distance(pred_spiky, gt_spiky)
return torch.mean(pairwise_distance)
def evaluate(model, val_data_loader):
global global_epoch, global_step
eval_epochs = 25
print('Evaluating model for {} epochs'.format(eval_epochs))
eval_L1_loss = 0.
eval_velocity_loss=0.
count = 0
folder = join(checkpoint_dir, "samples_step{:09d}".format(global_step))
if not os.path.exists(folder):
os.mkdir(folder)
for epoch in tqdm(range(eval_epochs),total=eval_epochs):
prog_bar = enumerate(val_data_loader)
for step, (T_mels,T_pose,T_content,Nl_pose,Nl_content) in prog_bar:
model.eval()
T_mels, T_pose, T_content,Nl_pose,Nl_content = T_mels.cuda(non_blocking=True), T_pose.cuda(non_blocking=True), T_content.cuda(non_blocking=True), \
Nl_pose.cuda(non_blocking=True), Nl_content.cuda(non_blocking=True)
# (B,T,1,hv,wv) (B,T,2,74) (B,T,2,57)
predict_content = model(T_mels, T_pose,Nl_pose,Nl_content) # (B*T,2,57)
T_content = torch.cat([T_content[i] for i in range(T_content.size(0))], dim=0) # (B*T,2,57)
eval_L1_loss += criterion_L1(predict_content, T_content).item()
eval_velocity_loss +=get_velocity_loss(predict_content, T_content).item()
count += 1
writer.add_scalar('eval_L1_loss', eval_L1_loss / count, global_step)
print('eval_L1_loss', eval_L1_loss / count, 'global_step:', global_step)
writer.add_scalar('eval_velocity_loss', eval_velocity_loss / count, global_step)
print('eval_velocity_loss', eval_velocity_loss / count, 'global_step:', global_step)
if __name__ == '__main__':
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir, exist_ok=True)
device = torch.device("cuda")
# create a model and optimizer
model = Landmark_transformer(T,d_model,nlayers,nhead,dim_feedforward,dropout)
if finetune_path is not None: ###fine tune
model_dict = model.state_dict()
print('load module....from :', finetune_path)
checkpoint = torch.load(finetune_path)
s = checkpoint["state_dict"]
new_s = {}
for k, v in s.items():
new_s[k.replace('module.', '')] = v
state_dict_needed = {k: v for k, v in new_s.items() if k in model_dict.keys()} # we need in model
model_dict.update(state_dict_needed)
model.load_state_dict(model_dict)
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
if torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
model = model.cuda()
train_dataset = Dataset('train')
val_dataset = Dataset('test')
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=batch_size,
shuffle=True,
drop_last=True,
num_workers=num_workers,
pin_memory=True
)
val_data_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=batch_size_val,
shuffle=True,
drop_last=True,
num_workers=num_workers,
pin_memory=True
)
while global_epoch < 9999999999:
prog_bar = tqdm(enumerate(train_data_loader), total=len(train_data_loader))
running_L1_loss,running_velocity_loss=0.,0.
for step, (T_mels, T_pose, T_content, Nl_pose, Nl_content) in prog_bar:
if global_step % checkpoint_interval == 0:
save_checkpoint(model, optimizer, global_step, checkpoint_dir, global_epoch, prefix=Project_name)
if global_step % evaluate_interval == 0 or global_step == 100:
with torch.no_grad():
evaluate(model, val_data_loader)
T_mels,T_pose,T_content,Nl_pose,Nl_content= T_mels.cuda(non_blocking=True), T_pose.cuda(non_blocking=True), T_content.cuda(non_blocking=True), \
Nl_pose.cuda(non_blocking=True),Nl_content.cuda(non_blocking=True)
#(B,T,1,hv,wv) (B,T,2,74) (B,T,2,57)
model.train()
optimizer.zero_grad()
predict_content=model(T_mels, T_pose, Nl_pose, Nl_content) #(B*T,2,57)
T_content=torch.cat([T_content[i] for i in range(T_content.size(0))],dim=0) #(B*T,2,57):ground truth lip and jaw landmarks
L1_loss=criterion_L1(predict_content,T_content)
Velocity_loss=get_velocity_loss(predict_content, T_content)
loss= L1_loss + Velocity_loss
loss.backward()
optimizer.step()
running_L1_loss+=L1_loss.item()
running_velocity_loss+=Velocity_loss.item()
prog_bar.set_description('epoch: %d step: %d running_L1_loss: %.4f running_velocity_loss: %.4f '
% (global_epoch, global_step, running_L1_loss / (step + 1), running_velocity_loss / (step + 1)))
writer.add_scalar('running_L1_loss', running_L1_loss / (step + 1), global_step)
writer.add_scalar('running_velocity_loss', running_velocity_loss / (step + 1), global_step)
global_step += 1
global_epoch += 1
print("end")