Spaces:
Running
on
T4
Running
on
T4
File size: 10,688 Bytes
b03a8f2 |
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 |
"""
motion_in and motion_out are all (bs, t, c), not (bs, t, j, c//j)
input:
audio: (bs, audio_t)
speaker_id: (bs, 1)
seed_frames: int
seed_motion: (bs, t, j*6) # rot6d
output:
motion: (bs, t, j*6) # rot6d
motion_axis_angle: (bs, t, j*3) # axis-angle
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import PreTrainedModel
from .configuration_disco_audio import DiscoAudioConfig
# ------------------ utils ---------------------- #
MASK_DICT = {
"local_upper": [
False, False, False, True, False, False, True, False, False, True,
False, False, True, True, True, True, True, True, True, True,
True, True, False, False, False, True, True, True, True, True,
True, True, True, True, True, True, True, True, True, True,
True, True, True, True, True, True, True, True, True, True,
True, True, True, True, True
],
"local_full": [False] + [True]*54
}
def _copysign(a, b):
signs_differ = (a < 0) != (b < 0)
return torch.where(signs_differ, -a, a)
def _sqrt_positive_part(x):
ret = torch.zeros_like(x)
positive_mask = x > 0
ret[positive_mask] = torch.sqrt(x[positive_mask])
return ret
def matrix_to_quaternion(matrix):
if matrix.size(-1) != 3 or matrix.size(-2) != 3:
raise ValueError(f"Invalid rotation matrix shape {matrix.shape}.")
m00 = matrix[..., 0, 0]
m11 = matrix[..., 1, 1]
m22 = matrix[..., 2, 2]
o0 = 0.5 * _sqrt_positive_part(1 + m00 + m11 + m22)
x = 0.5 * _sqrt_positive_part(1 + m00 - m11 - m22)
y = 0.5 * _sqrt_positive_part(1 - m00 + m11 - m22)
z = 0.5 * _sqrt_positive_part(1 - m00 - m11 + m22)
o1 = _copysign(x, matrix[..., 2, 1] - matrix[..., 1, 2])
o2 = _copysign(y, matrix[..., 0, 2] - matrix[..., 2, 0])
o3 = _copysign(z, matrix[..., 1, 0] - matrix[..., 0, 1])
return torch.stack((o0, o1, o2, o3), -1)
def quaternion_to_axis_angle(quaternions):
norms = torch.norm(quaternions[..., 1:], p=2, dim=-1, keepdim=True)
half_angles = torch.atan2(norms, quaternions[..., :1])
angles = 2 * half_angles
eps = 1e-6
small_angles = angles.abs() < eps
sin_half_angles_over_angles = torch.empty_like(angles)
sin_half_angles_over_angles[~small_angles] = (
torch.sin(half_angles[~small_angles]) / angles[~small_angles]
)
sin_half_angles_over_angles[small_angles] = (
0.5 - (angles[small_angles] * angles[small_angles]) / 48
)
return quaternions[..., 1:] / sin_half_angles_over_angles
def matrix_to_axis_angle(matrix):
return quaternion_to_axis_angle(matrix_to_quaternion(matrix))
def rotation_6d_to_matrix(d6: torch.Tensor) -> torch.Tensor:
a1, a2 = d6[..., :3], d6[..., 3:]
b1 = F.normalize(a1, dim=-1)
b2 = a2 - (b1 * a2).sum(-1, keepdim=True) * b1
b2 = F.normalize(b2, dim=-1)
b3 = torch.cross(b1, b2, dim=-1)
return torch.stack((b1, b2, b3), dim=-2)
def rotation_6d_to_axis_angle(rot6d):
return matrix_to_axis_angle(rotation_6d_to_matrix(rot6d))
def recover_from_mask_ts(selected_motion: torch.Tensor, mask: list[bool]) -> torch.Tensor:
device = selected_motion.device
dtype = selected_motion.dtype
mask_arr = torch.tensor(mask, dtype=torch.bool, device=device)
j = len(mask_arr)
sum_mask = mask_arr.sum().item()
c_channels = selected_motion.shape[-1] // sum_mask
new_shape = selected_motion.shape[:-1] + (sum_mask, c_channels)
selected_motion = selected_motion.reshape(new_shape)
out_shape = list(selected_motion.shape[:-2]) + [j, c_channels]
recovered = torch.zeros(out_shape, dtype=dtype, device=device)
recovered[..., mask_arr, :] = selected_motion
final_shape = list(recovered.shape[:-2]) + [j * c_channels]
recovered = recovered.reshape(final_shape)
return recovered
# ------------------ network ---------------------- #
class BasicBlock(nn.Module):
"""Basic 1D residual block."""
def __init__(self, inplanes, planes, ker_size, stride=1, first_dilation=None, norm_layer=nn.BatchNorm1d, act_layer=nn.LeakyReLU):
super().__init__()
self.conv1 = nn.Conv1d(inplanes, planes, kernel_size=ker_size, stride=stride,
padding=first_dilation, dilation=1, bias=True)
self.bn1 = norm_layer(planes)
self.act1 = act_layer(inplace=True)
self.conv2 = nn.Conv1d(planes, planes, kernel_size=ker_size, padding=ker_size//2, bias=True)
self.bn2 = norm_layer(planes)
self.act2 = act_layer(inplace=True)
self.downsample = None
if stride != 1 or inplanes != planes:
self.downsample = nn.Sequential(
nn.Conv1d(inplanes, planes, stride=stride, kernel_size=ker_size, padding=first_dilation, bias=True),
norm_layer(planes)
)
def forward(self, x):
shortcut = x
x = self.conv1(x)
x = self.bn1(x)
x = self.act1(x)
x = self.conv2(x)
x = self.bn2(x)
if self.downsample is not None:
shortcut = self.downsample(shortcut)
x += shortcut
x = self.act2(x)
return x
class WavEncoder(nn.Module):
"""Waveform encoder that uses stacked residual blocks."""
def __init__(self, out_dim):
super().__init__()
self.feat_extractor = nn.Sequential(
BasicBlock(1, 32, 15, 5, first_dilation=1600),
BasicBlock(32,32,15,6,first_dilation=0),
BasicBlock(32,32,15,1,first_dilation=7),
BasicBlock(32,64,15,6,first_dilation=0),
BasicBlock(64,64,15,1,first_dilation=7),
BasicBlock(64,128,15,6,first_dilation=0),
)
def forward(self, wav_data):
wav_data = wav_data.unsqueeze(1)
out = self.feat_extractor(wav_data)
return out.transpose(1, 2)
class MLP(nn.Module):
"""A simple MLP for projection."""
def __init__(self, in_dim, middle_dim, out_dim):
super().__init__()
self.fc1 = nn.Linear(in_dim, middle_dim)
self.fc2 = nn.Linear(middle_dim, out_dim)
self.act = nn.LeakyReLU(0.1, True)
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.fc2(x)
return x
class Empty(nn.Module):
"""Empty module that returns input as is."""
def forward(self, x):
return x
class DiscoAudioPreTrainedModel(PreTrainedModel):
config_class = DiscoAudioConfig
base_model_prefix = "camn_audio"
def _init_weights(self, module):
pass
class DiscoAudioModel(DiscoAudioPreTrainedModel):
"""DiscoAudio model for audio-driven motion generation.
This model assumes that the config (DiscoAudioConfig) can be initialized from a dict-like object
or OmegaConf directly by passing them as kwargs. For example:
from omegaconf import OmegaConf
cfg = OmegaConf.load("configs/camn_audio.yaml")
config = DiscoAudioConfig(config_obj=cfg.model)
This way all attributes from cfg.model become config attributes without having to manually map each one.
"""
def __init__(self, config: DiscoAudioConfig):
super().__init__(config)
self.pose_rep = config.pose_rep
self.cfg = config
self.audio_encoder = WavEncoder(self.cfg.audio_f)
self.speaker_embedding = nn.Embedding(self.cfg.speaker_dims, self.cfg.speaker_f) if self.cfg.speaker_f > 0 else None
self.motion_encoder = Empty()
self.joint_mask = MASK_DICT[config.joint_mask]
self.audio_encoder_c1 = MLP(self.cfg.audio_f, self.cfg.hidden_size, self.cfg.audio_f)
self.audio_encoder_c2 = MLP(self.cfg.audio_f, self.cfg.hidden_size, self.cfg.audio_f)
self.audio_encoder_r = MLP(self.cfg.audio_f, self.cfg.hidden_size, self.cfg.audio_f)
self.selector = MLP(self.cfg.audio_f, self.cfg.hidden_size, 2)
self.softmax = nn.Softmax(dim=2)
input_dim_body = self.cfg.pose_dims+1+self.cfg.speaker_f+self.cfg.audio_f*2
self.body_motion_decoder = nn.LSTM(
input_dim_body, hidden_size=self.cfg.hidden_size,
num_layers=self.cfg.n_layer, batch_first=True,
bidirectional=True, dropout=self.cfg.dropout_prob
)
self.body_out = MLP(self.cfg.hidden_size, self.cfg.hidden_size, self.cfg.pose_dims)
def forward(self, audio, speaker_id, seed_frames=4, seed_motion=None, return_axis_angle=True):
audio_feat = self.audio_encoder(audio)
bs, t, _ = audio_feat.shape
if self.speaker_embedding is not None:
speaker_feat = self.speaker_embedding(speaker_id)
speaker_feat = speaker_feat.repeat(1, t, 1)
else:
speaker_feat = torch.zeros(bs, t, 0, device=audio.device)
if seed_motion is None:
seed_motion = torch.zeros(bs, t, self.cfg.pose_dims+1, device=audio.device)
seed_motion[:, :seed_frames, -1] = 1
else:
_, t_m, _ = seed_motion.shape
seed_motion_pad = torch.zeros(bs, t_m, self.cfg.pose_dims+1, device=audio.device)
seed_motion_pad[:, :seed_frames, :-1] = seed_motion[:, :seed_frames]
seed_motion_pad[:, :seed_frames, -1] = 1
seed_motion = seed_motion_pad
if t_m != t:
diff_length = t_m - t
if diff_length > 0:
seed_motion = seed_motion[:, :t, :]
else:
seed_motion = torch.cat((seed_motion, seed_motion[:, -diff_length:, :]), 1)
audio_feat_c1 = self.audio_encoder_c1(audio_feat)
audio_feat_c2 = self.audio_encoder_c2(audio_feat)
audio_feat_r = self.audio_encoder_r(audio_feat)
weight_c = self.softmax(self.selector(audio_feat))
audio_feat_c = weight_c[:, :, 0:1] * audio_feat_c1 + weight_c[:, :, 1:2] * audio_feat_c2
audio_feat = torch.cat((audio_feat_c, audio_feat_r), dim=2)
in_fea = torch.cat((audio_feat, speaker_feat, seed_motion), dim=2)
body_out, _ = self.body_motion_decoder(in_fea)
body_out = body_out[:, :, :self.cfg.hidden_size] + body_out[:, :, self.cfg.hidden_size:]
recombine = self.body_out(body_out)
motion_axis_angle = None
if return_axis_angle:
motion_axis_angle = rotation_6d_to_axis_angle(recombine.reshape(-1, self.cfg.pose_dims//6, 6)).reshape(bs, t, -1)
motion_axis_angle = recover_from_mask_ts(motion_axis_angle, self.joint_mask)
return {
"motion": recombine,
"motion_axis_angle": motion_axis_angle,
"audio_fea_c": audio_feat_c,
"audio_fea_r": audio_feat_r,
} |