-
Notifications
You must be signed in to change notification settings - Fork 13
/
data_loader.py
145 lines (118 loc) · 5.83 KB
/
data_loader.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
import os
import random
import torch
import numpy as np
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
class MotionDataset(Dataset):
def __init__(self, phase, data_dir, transform=None):
super(MotionDataset, self).__init__()
if phase == 'train':
data_npz_path = os.path.join(data_dir, 'train_dataset.npz')
else:
data_npz_path = os.path.join(data_dir, 'test_dataset.npz')
mdataset = np.load(data_npz_path, allow_pickle=True)
self.motions = mdataset["motion"]
# self.roots = mdataset['root']
# self.foot_contacts = mdataset["foot_contact"]
data_norm_dir = os.path.join(data_dir, 'norm')
motion_mean_path = os.path.join(data_norm_dir, "motion_mean.npy")
motion_std_path = os.path.join(data_norm_dir, "motion_std.npy")
# root_mean_path = os.path.join(data_norm_dir, "root_mean.npy")
# root_std_path = os.path.join(data_norm_dir, "root_std.npy")
if os.path.exists(motion_mean_path) and os.path.exists(motion_std_path):
self.motion_mean = np.load(motion_mean_path, allow_pickle=True).astype(np.float32)
self.motion_std = np.load(motion_std_path, allow_pickle=True).astype(np.float32)
# self.root_mean = np.load(root_mean_path, allow_pickle=True).astype(np.float32)
# self.root_std = np.load(root_std_path, allow_pickle=True).astype(np.float32)
else:
assert self.motion_mean and self.motion_std, 'no motion_mean or no motion_std'
self.transform = transform
def __len__(self):
return len(self.motions)
def __getitem__(self, index):
if torch.is_tensor(index):
index = index.tolist()
motion_raw = self.motions[index].astype(np.float32)
motion = np.transpose(motion_raw, (2, 1, 0)) # (seq, joint, dim) -> (dim, joint, seq)
motion = torch.from_numpy(motion)
trans_p = float(np.random.rand(1))
if self.transform and trans_p < 0.2:
motion = self.transform(motion)
motion = (motion - self.motion_mean[:, np.newaxis, np.newaxis]) \
/ self.motion_std[:, np.newaxis, np.newaxis] # normalization
# root_raw = self.roots[index].astype(np.float32)
# root = np.transpose(root_raw, (2, 1, 0)) # (seq, joint, dim) -> (dim, joint, seq)
# root = torch.from_numpy(root)
# root = (root - self.root_mean[:, np.newaxis, np.newaxis]) \
# / self.root_std[:, np.newaxis, np.newaxis] # normalization
# foot_contact = self.foot_contacts[index].astype(np.float32)
data = {
"motion_raw": motion_raw,
"motion": motion,
# "root_raw": root_raw,
# "root": root,
# "foot_contact": foot_contact
}
return data
class RandomResizedCrop(object):
"""Crop and resize randomly the motion in a sample."""
def __call__(self, sample):
global crop
c, j, s = sample.shape # (dim, joint, seq)
idx = random.randint(30, 90)
size = random.randint(60, 120)
if idx > (size//2)+(size%2) and idx+(size//2) < 120:
crop = sample[..., idx-(size//2)-(size%2):idx+(size//2)]
elif idx <= (size//2)+(size%2):
crop = sample[..., :idx+(size//2)]
elif idx+(size//2) >= 120:
crop = sample[..., idx-(size//2)-(size%2):]
if size < 90:
scale = random.uniform(1, 2)
else:
scale = random.uniform(0.5, 1)
crop = crop.unsqueeze(0)
# crop = torch.from_numpy(crop).unsqueeze(0)
scale_crop = F.interpolate(crop, scale_factor=(1, scale), mode='bilinear',
align_corners=True, recompute_scale_factor=True)
scale_crop = scale_crop.squeeze(0)
if scale_crop.shape[-1] > 120:
scale_crop = scale_crop[..., scale_crop.shape[-1]//2-60:scale_crop.shape[-1]//2+60]
return scale_crop
else:
# padding
left = scale_crop[..., :1].repeat_interleave(
(120-scale_crop.shape[-1])//2 + (120-scale_crop.shape[-1]) % 2, dim=-1)
left[-3:] = 0.0
right = scale_crop[..., -1:].repeat_interleave((120-scale_crop.shape[-1])//2, dim=-1)
right[-3:] = 0.0
padding_scale_crop = torch.cat([left, scale_crop, right], dim=-1)
return padding_scale_crop
def get_dataloader(subset_name, config, seed=None, shuffle=None, transform=None):
dataset = MotionDataset(subset_name, config['data_dir'], transform)
batch_size = config['batch_size'] if subset_name == 'train' else 1 # since dataloader
return DataLoader(dataset, batch_size=batch_size,
shuffle=(subset_name == 'train') if shuffle is None else shuffle,
num_workers=config['num_workers'] if subset_name == 'train' else 0,
# worker_init_fn=np.random.seed(seed) if seed else None,
pin_memory=True,
drop_last=False)
if __name__ == '__main__':
import sys
from etc.utils import print_composite
sys.path.append('./motion')
sys.path.append('./etc')
from viz_motion import animation_plot # for checking dataloader
data_dir = './datasets/cmu/'
batch_size = 2
dataset = MotionDataset('train', data_dir)
data_loader = DataLoader(dataset, batch_size=batch_size, shuffle=True)
for batch in data_loader:
print_composite(batch)
motion_raw = batch['motion_raw'].cpu().numpy()
root_raw = batch['root_raw'].cpu().numpy()
foot_contact = batch['foot_contact'].cpu().numpy()
anim1 = [motion_raw[0], root_raw[0], foot_contact[0]]
anim2 = [motion_raw[1], root_raw[1], foot_contact[1]]
animation_plot([anim1, anim2])