-
Notifications
You must be signed in to change notification settings - Fork 3
/
dataset_TM_train.py
161 lines (127 loc) · 6.04 KB
/
dataset_TM_train.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
import torch
from torch.utils import data
import numpy as np
from os.path import join as pjoin
import random
import codecs as cs
from tqdm import tqdm
import utils.paramUtil as paramUtil
from torch.utils.data._utils.collate import default_collate
def collate_fn(batch):
batch.sort(key=lambda x: x[3], reverse=True)
return default_collate(batch)
'''For use of training text-2-motion generative model'''
class Text2MotionDataset(data.Dataset):
def __init__(self, dataset_name, feat_bias = 5, unit_length = 4, codebook_size = 1024, tokenizer_name=None):
self.max_length = 64
self.pointer = 0
self.dataset_name = dataset_name
self.unit_length = unit_length
# self.mot_start_idx = codebook_size
self.mot_end_idx = codebook_size
self.mot_pad_idx = codebook_size + 1
if dataset_name == 't2m':
self.data_root = '/data/zhongchongyang/TTM/HumanML3D-main/HumanML3D'
self.motion_dir = pjoin(self.data_root, 'new_joint_vecs')
self.text_dir = pjoin(self.data_root, 'texts')
self.joints_num = 22
radius = 4
fps = 20
self.max_motion_length = 26 if unit_length == 8 else 51
dim_pose = 263
kinematic_chain = paramUtil.t2m_kinematic_chain
elif dataset_name == 'kit':
self.data_root = '/data/zhongchongyang/motiondiffuse/text2motion/data/KIT-ML'
self.motion_dir = pjoin(self.data_root, 'new_joint_vecs')
self.text_dir = pjoin(self.data_root, 'texts')
self.joints_num = 21
radius = 240 * 8
fps = 12.5
dim_pose = 251
self.max_motion_length = 26 if unit_length == 8 else 51
kinematic_chain = paramUtil.kit_kinematic_chain
split_file = pjoin(self.data_root, 'train.txt')
id_list = []
with cs.open(split_file, 'r') as f:
for line in f.readlines():
id_list.append(line.strip())
new_name_list = []
data_dict = {}
for name in tqdm(id_list):
try:
m_token_list = np.load(pjoin(self.data_root, tokenizer_name, '%s.npy'%name))
# Read text
with cs.open(pjoin(self.text_dir, name + '.txt')) as f:
text_data = []
flag = False
lines = f.readlines()
for line in lines:
try:
text_dict = {}
line_split = line.strip().split('#')
caption = line_split[0]
t_tokens = line_split[1].split(' ')
f_tag = float(line_split[2])
to_tag = float(line_split[3])
f_tag = 0.0 if np.isnan(f_tag) else f_tag
to_tag = 0.0 if np.isnan(to_tag) else to_tag
text_dict['caption'] = caption
text_dict['tokens'] = t_tokens
if f_tag == 0.0 and to_tag == 0.0:
flag = True
text_data.append(text_dict)
else:
m_token_list_new = [tokens[int(f_tag*fps/unit_length) : int(to_tag*fps/unit_length)] for tokens in m_token_list if int(f_tag*fps/unit_length) < int(to_tag*fps/unit_length)]
if len(m_token_list_new) == 0:
continue
new_name = '%s_%f_%f'%(name, f_tag, to_tag)
data_dict[new_name] = {'m_token_list': m_token_list_new,
'text':[text_dict]}
new_name_list.append(new_name)
except:
pass
if flag:
data_dict[name] = {'m_token_list': m_token_list,
'text':text_data}
new_name_list.append(name)
except:
pass
self.data_dict = data_dict
self.name_list = new_name_list
def __len__(self):
return len(self.data_dict)
def __getitem__(self, item):
data = self.data_dict[self.name_list[item]]
m_token_list, text_list = data['m_token_list'], data['text']
m_tokens = random.choice(m_token_list)
text_data = random.choice(text_list)
caption= text_data['caption']
coin = np.random.choice([False, False, True])
# print(len(m_tokens))
if coin:
# drop one token at the head or tail
coin2 = np.random.choice([True, False])
if coin2:
m_tokens = m_tokens[:-1]
else:
m_tokens = m_tokens[1:]
m_tokens_len = m_tokens.shape[0]
if m_tokens_len+1 < self.max_motion_length:
m_tokens = np.concatenate([m_tokens, np.ones((1), dtype=int) * self.mot_end_idx, np.ones((self.max_motion_length-1-m_tokens_len), dtype=int) * self.mot_pad_idx], axis=0)
else:
m_tokens = np.concatenate([m_tokens, np.ones((1), dtype=int) * self.mot_end_idx], axis=0)
return caption, m_tokens.reshape(-1), m_tokens_len
def DATALoader(dataset_name,
batch_size, codebook_size, tokenizer_name, unit_length=4,
num_workers = 8) :
train_loader = torch.utils.data.DataLoader(Text2MotionDataset(dataset_name, codebook_size = codebook_size, tokenizer_name = tokenizer_name, unit_length=unit_length),
batch_size,
shuffle=True,
num_workers=num_workers,
#collate_fn=collate_fn,
drop_last = True)
return train_loader
def cycle(iterable):
while True:
for x in iterable:
yield x