forked from anliyuan/Ultralight-Digital-Human
-
Notifications
You must be signed in to change notification settings - Fork 0
/
datasetsss.py
143 lines (115 loc) · 4.88 KB
/
datasetsss.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
import os
import cv2
import torch
import random
import numpy as np
import random
from torch.utils.data import Dataset
from torch.utils.data import DataLoader
class MyDataset(Dataset):
def __init__(self, img_dir, mode):
self.img_path_list = []
self.lms_path_list = []
self.mode = mode
for i in range(len(os.listdir(img_dir+"/full_body_img/"))):
img_path = os.path.join(img_dir+"/full_body_img/", str(i)+".jpg")
lms_path = os.path.join(img_dir+"/landmarks/", str(i)+".lms")
self.img_path_list.append(img_path)
self.lms_path_list.append(lms_path)
if self.mode == "wenet":
self.audio_feats = np.load(img_dir+"/aud_wenet.npy")
if self.mode == "hubert":
self.audio_feats = np.load(img_dir+"/aud_hu.npy")
self.audio_feats = self.audio_feats.astype(np.float32)
print(img_dir)
print(self.audio_feats.shape)
print(len(self.img_path_list))
def __len__(self):
# return len(self.img_path_list)-1
# return len(self.img_path_list)
return self.audio_feats.shape[0]-1
def get_audio_features(self, features, index):
left = index - 8
right = index + 8
pad_left = 0
pad_right = 0
if left < 0:
pad_left = -left
left = 0
if right > features.shape[0]:
pad_right = right - features.shape[0]
right = features.shape[0]
auds = torch.from_numpy(features[left:right])
if pad_left > 0:
auds = torch.cat([torch.zeros_like(auds[:pad_left]), auds], dim=0)
if pad_right > 0:
auds = torch.cat([auds, torch.zeros_like(auds[:pad_right])], dim=0) # [8, 16]
return auds
def get_audio_features_1(self, features, index):
left = index - 8
pad_left = 0
if left < 0:
pad_left = -left
left = 0
auds = features[left:index]
auds = torch.from_numpy(auds)
if pad_left > 0:
# pad may be longer than auds, so do not use zeros_like
auds = torch.cat([torch.zeros(pad_left, *auds.shape[1:], device=auds.device, dtype=auds.dtype), auds], dim=0)
return auds
def process_img(self, img, lms_path, img_ex, lms_path_ex):
lms_list = []
with open(lms_path, "r") as f:
lines = f.read().splitlines()
for line in lines:
arr = line.split(" ")
arr = np.array(arr, dtype=np.float32)
lms_list.append(arr)
lms = np.array(lms_list, dtype=np.int32)
xmin = lms[1][0]
ymin = lms[52][1]
xmax = lms[31][0]
width = xmax - xmin
ymax = ymin + width
crop_img = img[ymin:ymax, xmin:xmax]
crop_img = cv2.resize(crop_img, (168, 168), cv2.INTER_AREA)
img_real = crop_img[4:164, 4:164].copy()
img_real_ori = img_real.copy()
img_masked = cv2.rectangle(img_real,(5,5,150,145),(0,0,0),-1)
lms_list = []
with open(lms_path_ex, "r") as f:
lines = f.read().splitlines()
for line in lines:
arr = line.split(" ")
arr = np.array(arr, dtype=np.float32)
lms_list.append(arr)
lms = np.array(lms_list, dtype=np.int32)
xmin = lms[1][0]
ymin = lms[52][1]
xmax = lms[31][0]
width = xmax - xmin
ymax = ymin + width
crop_img = img_ex[ymin:ymax, xmin:xmax]
crop_img = cv2.resize(crop_img, (168, 168), cv2.INTER_AREA)
img_real_ex = crop_img[4:164, 4:164].copy()
img_real_ori = img_real_ori.transpose(2,0,1).astype(np.float32)
img_masked = img_masked.transpose(2,0,1).astype(np.float32)
img_real_ex = img_real_ex.transpose(2,0,1).astype(np.float32)
img_real_ex_T = torch.from_numpy(img_real_ex / 255.0)
img_real_T = torch.from_numpy(img_real_ori / 255.0)
img_masked_T = torch.from_numpy(img_masked / 255.0)
img_concat_T = torch.cat([img_real_ex_T, img_masked_T], axis=0)
return img_concat_T, img_real_T
def __getitem__(self, idx):
img = cv2.imread(self.img_path_list[idx])
lms_path = self.lms_path_list[idx]
ex_int = random.randint(0, self.__len__()-1)
img_ex = cv2.imread(self.img_path_list[ex_int])
lms_path_ex = self.lms_path_list[ex_int]
img_concat_T, img_real_T = self.process_img(img, lms_path, img_ex, lms_path_ex)
audio_feat = self.get_audio_features(self.audio_feats, idx)
if self.mode == "wenet":
audio_feat = audio_feat.reshape(256,16,32)
if self.mode == "hubert":
audio_feat = audio_feat.reshape(32,32,32)
return img_concat_T, img_real_T, audio_feat