File size: 7,796 Bytes
8b3f571 | 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 | import numpy as np
import random, torch, joblib
from torch.utils.data import Dataset
from mmengine.registry import DATASETS, FUNCTIONS
import time, os
from ..utils.motion import recover_from_local_position_fast
FPS = 30
@DATASETS.register_module()
class AnyMoTextDataset(Dataset):
'''
data format:
dict(
motion_272: np.ndarray (T, 272),
captions: [str]
)
'''
def __init__(self, data_path_file='', unit_length=4, mean_path='', std_path='',
min_motion_length=60, max_motion_length=300, interval=1):
super().__init__()
self.min_motion_length = min_motion_length
self.max_motion_length = max_motion_length
self.unit_length = unit_length
if isinstance(data_path_file, list):
self.data_paths = []
for path in data_path_file:
with open(path, 'r') as f:
self.data_paths.extend([line.strip() for line in f.readlines()])
elif isinstance(data_path_file, str):
with open(data_path_file, 'r') as f:
self.data_paths = [line.strip() for line in f.readlines()]
else:
raise NotImplementedError
self.data_paths = self.data_paths[::interval]
self.motion_mean = torch.from_numpy(np.load(mean_path).astype(np.float32))
self.motion_std = torch.from_numpy(np.load(std_path).astype(np.float32))
def __len__(self):
return len(self.data_paths)
def __getitem__(self, index):
data_infos = joblib.load(self.data_paths[index])
data = []
for data_info in data_infos:
# 1. load motion
motion = torch.as_tensor(data_info['motion_272'], dtype=torch.float32)
caption = random.choice(data_info['captions'])
# 2. clip motion
T = motion.shape[0]
motion_length = (T // self.unit_length) * self.unit_length
if motion_length > self.max_motion_length:
motion_length = self.max_motion_length
idx = random.randint(0, T - motion_length)
motion = motion[idx:idx+motion_length]
motion = (motion - self.motion_mean) / self.motion_std
pad_motion = torch.zeros(self.max_motion_length - motion_length, motion.shape[1]).to(motion)
motion = torch.cat([motion, pad_motion], dim=0)
data.append(dict(
motion=motion,
text_data=caption,
motion_length=motion_length,
mean=self.motion_mean,
std=self.motion_std,
))
return data
@DATASETS.register_module()
class AnyMoTextDataset_v1(AnyMoTextDataset):
'''
加采样
data format:
dict(
motion_272: np.ndarray (T, 272),
captions: [str]
)
'''
def __init__(self, sample_ratio=0.5, **kwargs):
super().__init__(**kwargs)
self.sample_ratio = sample_ratio
def __len__(self):
return int(len(self.data_paths) * sample_ratio)
def __getitem__(self, index):
cap_size = len(self.data_paths)
size = int(cap_size * self.sample_ratio)
begin = random.randint(0, cap_size - size)
data_infos = joblib.load(self.caption_paths[begin + index])
data = []
for data_info in data_infos:
# 1. load motion
motion = torch.as_tensor(data_info['motion_272'], dtype=torch.float32)
caption = random.choice(data_info['captions'])
# 2. clip motion
T = motion.shape[0]
motion_length = (T // self.unit_length) * self.unit_length
if motion_length > self.max_motion_length:
motion_length = self.max_motion_length
idx = random.randint(0, T - motion_length)
motion = motion[idx:idx+motion_length]
motion = (motion - self.motion_mean) / self.motion_std
pad_motion = torch.zeros(self.max_motion_length - motion_length, motion.shape[1]).to(motion)
motion = torch.cat([motion, pad_motion], dim=0)
data.append(dict(
motion=motion,
text_data=caption,
motion_length=motion_length,
mean=self.motion_mean,
std=self.motion_std,
))
return data
@DATASETS.register_module()
class AnyMoTextTrajDataset(AnyMoTextDataset):
def __init__(self, traj_mean_path='', traj_std_path='', **kwargs):
super().__init__(**kwargs)
self.traj_mean = torch.from_numpy(np.load(traj_mean_path).astype(np.float32))
self.traj_std = torch.from_numpy(np.load(traj_std_path).astype(np.float32))
def __getitem__(self, index):
data_infos = joblib.load(self.data_paths[index])
data = []
for data_info in data_infos:
# 1. load motion
motion = torch.as_tensor(data_info['motion_272'], dtype=torch.float32)
caption = random.choice(data_info['captions'])
traj_data = recover_from_local_position_fast(motion[None]).squeeze(0)[:, 0]
# 2. clip motion
T = motion.shape[0]
motion_length = (T // self.unit_length) * self.unit_length
if motion_length > self.max_motion_length:
motion_length = self.max_motion_length
idx = random.randint(0, T - motion_length)
motion = motion[idx:idx+motion_length]
traj_data = traj_data[idx:idx+motion_length]
traj_data = (traj_data - self.traj_mean) / self.traj_std
motion = (motion - self.motion_mean) / self.motion_std
pad_motion = torch.zeros(self.max_motion_length - motion_length, motion.shape[1]).to(motion)
motion = torch.cat([motion, pad_motion], dim=0)
pad_traj = torch.zeros(self.max_motion_length - motion_length, traj_data.shape[1]).to(traj_data)
traj_data = torch.cat([traj_data, pad_traj], dim=0)
data.append(dict(
motion=motion,
traj_data=traj_data,
text_data=caption,
motion_length=motion_length,
mean=self.motion_mean,
std=self.motion_std,
))
return data
@FUNCTIONS.register_module()
def anymo_text_traj_collate_fn(batch):
new_batch = []
for item in batch:
new_batch.extend(item)
motion = [item['motion'] for item in new_batch] # [motion: T, C]
traj_data = [item['traj_data'] for item in new_batch]
traj_data = torch.stack(traj_data)
text_data = [item['text_data'] for item in new_batch]
motion_length = torch.as_tensor([item['motion_length'] for item in new_batch])
motion = torch.stack(motion)
motion = motion[:, :motion_length.max(), :]
traj_data = traj_data[:, :motion_length.max()]
mean = new_batch[0]['mean']
std = new_batch[0]['std']
return dict(
motion=motion,
traj_data=traj_data,
motion_length=motion_length,
text_data=text_data,
mean=mean,
std=std,
)
@FUNCTIONS.register_module()
def anymo_text_collate_fn(batch):
new_batch = []
for item in batch:
new_batch.extend(item)
motion = [item['motion'] for item in new_batch] # [motion: T, C]
text_data = [item['text_data'] for item in new_batch]
motion_length = torch.as_tensor([item['motion_length'] for item in new_batch])
motion = torch.stack(motion)
motion = motion[:, :motion_length.max(), :]
mean = new_batch[0]['mean']
std = new_batch[0]['std']
return dict(
motion=motion,
motion_length=motion_length,
text_data=text_data,
mean=mean,
std=std,
)
|