Spaces:
Running
on
Zero
Running
on
Zero
File size: 12,968 Bytes
1da48bb |
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 |
import torch
import torch.nn as nn
import os
import math
import pickle
import numpy as np
import torch.nn.functional as F
from torch.nn.utils import weight_norm
from .utils.build_vocab import Vocab
from torch.nn import TransformerEncoder, TransformerEncoderLayer
from .utils.layer import BasicBlock, TextEncoderTCN
from .utils.wav2vec import Wav2Vec2Model
from .utils.layer import ResBlock, init_weight
from loguru import logger
from .motion_encoder import *
import copy
class WavEncoder(nn.Module):
def __init__(self, out_dim, audio_in=1):
super().__init__()
self.out_dim = out_dim
self.feat_extractor = nn.Sequential(
BasicBlock(audio_in, out_dim//4, 15, 5, first_dilation=1600, downsample=True),
BasicBlock(out_dim//4, out_dim//4, 15, 6, first_dilation=0, downsample=True),
BasicBlock(out_dim//4, out_dim//4, 15, 1, first_dilation=7, ),
BasicBlock(out_dim//4, out_dim//2, 15, 6, first_dilation=0, downsample=True),
BasicBlock(out_dim//2, out_dim//2, 15, 1, first_dilation=7),
BasicBlock(out_dim//2, out_dim, 15, 3, first_dilation=0,downsample=True),
)
def forward(self, wav_data):
if wav_data.dim() == 2:
wav_data = wav_data.unsqueeze(1)
else:
wav_data = wav_data.transpose(1, 2)
out = self.feat_extractor(wav_data)
return out.transpose(1, 2)
class MLP(nn.Module):
def __init__(self, in_dim, hidden_size, out_dim):
super().__init__()
self.mlp = nn.Sequential(
nn.Linear(in_dim, hidden_size),
nn.LeakyReLU(0.2, True),
nn.Linear(hidden_size, out_dim)
)
def forward(self, inputs):
out = self.mlp(inputs)
return out
class PeriodicPositionalEncoding(nn.Module):
def __init__(self, d_model, dropout=0.1, period=15, max_seq_len=60): # 12s
super(PeriodicPositionalEncoding, self).__init__()
self.dropout = nn.Dropout(p=dropout)
pe = torch.zeros(period, d_model)
position = torch.arange(0, period, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(0) # (1, period, d_model)
repeat_num = (max_seq_len//period) + 1
pe = pe.repeat(1, repeat_num, 1) # (1, repeat_num, period, d_model)
self.register_buffer('pe', pe)
def forward(self, x):
# print(self.pe.shape, x.shape)
x = x + self.pe[:, :x.size(1), :]
return self.dropout(x)
class MAGE_Transformer(nn.Module):
def __init__(self, args):
super(MAGE_Transformer, self).__init__()
self.args = args
with open(f"{args.data_path}weights/vocab.pkl", 'rb') as f:
self.lang_model = pickle.load(f)
pre_trained_embedding = self.lang_model.word_embedding_weights
self.text_pre_encoder_face = nn.Embedding.from_pretrained(torch.FloatTensor(pre_trained_embedding),freeze=args.t_fix_pre)
self.text_encoder_face = nn.Linear(300, args.audio_f)
self.text_encoder_face = nn.Linear(300, args.audio_f)
self.text_pre_encoder_body = nn.Embedding.from_pretrained(torch.FloatTensor(pre_trained_embedding),freeze=args.t_fix_pre)
self.text_encoder_body = nn.Linear(300, args.audio_f)
self.text_encoder_body = nn.Linear(300, args.audio_f)
self.audio_pre_encoder_face = WavEncoder(args.audio_f, audio_in=2)
self.audio_pre_encoder_body = WavEncoder(args.audio_f, audio_in=2)
self.at_attn_face = nn.Linear(args.audio_f*2, args.audio_f*2)
self.at_attn_body = nn.Linear(args.audio_f*2, args.audio_f*2)
args_top = copy.deepcopy(self.args)
args_top.vae_layer = 3
args_top.vae_length = args.motion_f
args_top.vae_test_dim = args.pose_dims+3+4
self.motion_encoder = VQEncoderV6(args_top) # masked motion to latent bs t 333 to bs t 256
# face decoder
self.feature2face = nn.Linear(args.audio_f*2, args.hidden_size)
self.face2latent = nn.Linear(args.hidden_size, args.vae_codebook_size)
self.transformer_de_layer = nn.TransformerDecoderLayer(
d_model=self.args.hidden_size,
nhead=4,
dim_feedforward=self.args.hidden_size*2,
batch_first=True
)
self.face_decoder = nn.TransformerDecoder(self.transformer_de_layer, num_layers=4)
self.position_embeddings = PeriodicPositionalEncoding(self.args.hidden_size, period=self.args.pose_length, max_seq_len=self.args.pose_length)
# motion decoder
self.transformer_en_layer = nn.TransformerEncoderLayer(
d_model=self.args.hidden_size,
nhead=4,
dim_feedforward=self.args.hidden_size*2,
batch_first=True
)
self.motion_self_encoder = nn.TransformerEncoder(self.transformer_en_layer, num_layers=1)
self.audio_feature2motion = nn.Linear(args.audio_f, args.hidden_size)
self.feature2motion = nn.Linear(args.motion_f, args.hidden_size)
self.bodyhints_face = MLP(args.motion_f, args.hidden_size, args.motion_f)
self.bodyhints_body = MLP(args.motion_f, args.hidden_size, args.motion_f)
self.motion2latent_upper = MLP(args.hidden_size, args.hidden_size, self.args.hidden_size)
self.motion2latent_hands = MLP(args.hidden_size, args.hidden_size, self.args.hidden_size)
self.motion2latent_lower = MLP(args.hidden_size, args.hidden_size, self.args.hidden_size)
self.wordhints_decoder = nn.TransformerDecoder(self.transformer_de_layer, num_layers=8)
self.upper_decoder = nn.TransformerDecoder(self.transformer_de_layer, num_layers=1)
self.hands_decoder = nn.TransformerDecoder(self.transformer_de_layer, num_layers=1)
self.lower_decoder = nn.TransformerDecoder(self.transformer_de_layer, num_layers=1)
self.face_classifier = MLP(self.args.vae_codebook_size, args.hidden_size, self.args.vae_codebook_size)
self.upper_classifier = MLP(self.args.vae_codebook_size, args.hidden_size, self.args.vae_codebook_size)
self.hands_classifier = MLP(self.args.vae_codebook_size, args.hidden_size, self.args.vae_codebook_size)
self.lower_classifier = MLP(self.args.vae_codebook_size, args.hidden_size, self.args.vae_codebook_size)
self.mask_embeddings = nn.Parameter(torch.zeros(1, 1, self.args.pose_dims+3+4))
self.motion_down_upper = nn.Linear(args.hidden_size, self.args.vae_codebook_size)
self.motion_down_hands = nn.Linear(args.hidden_size, self.args.vae_codebook_size)
self.motion_down_lower = nn.Linear(args.hidden_size, self.args.vae_codebook_size)
self.motion_down_upper = nn.Linear(args.hidden_size, self.args.vae_codebook_size)
self.motion_down_hands = nn.Linear(args.hidden_size, self.args.vae_codebook_size)
self.motion_down_lower = nn.Linear(args.hidden_size, self.args.vae_codebook_size)
self._reset_parameters()
self.spearker_encoder_body = nn.Embedding(25, args.hidden_size)
self.spearker_encoder_face = nn.Embedding(25, args.hidden_size)
def _reset_parameters(self):
nn.init.normal_(self.mask_embeddings, 0, self.args.hidden_size ** -0.5)
def forward(self, in_audio=None, in_word=None, mask=None, is_test=None, in_motion=None, use_attentions=True, use_word=True, in_id = None):
in_word_face = self.text_pre_encoder_face(in_word)
in_word_face = self.text_encoder_face(in_word_face)
in_word_body = self.text_pre_encoder_body(in_word)
in_word_body = self.text_encoder_body(in_word_body)
bs, t, c = in_word_face.shape
in_audio_face = self.audio_pre_encoder_face(in_audio)
in_audio_body = self.audio_pre_encoder_body(in_audio)
if in_audio_face.shape[1] != in_motion.shape[1]:
diff_length = in_motion.shape[1]- in_audio_face.shape[1]
if diff_length < 0:
in_audio_face = in_audio_face[:, :diff_length, :]
in_audio_body = in_audio_body[:, :diff_length, :]
else:
in_audio_face = torch.cat((in_audio_face, in_audio_face[:,-diff_length:]),1)
in_audio_body = torch.cat((in_audio_body, in_audio_body[:,-diff_length:]),1)
if use_attentions:
alpha_at_face = torch.cat([in_word_face, in_audio_face], dim=-1).reshape(bs, t, c*2)
alpha_at_face = self.at_attn_face(alpha_at_face).reshape(bs, t, c, 2)
alpha_at_face = alpha_at_face.softmax(dim=-1)
fusion_face = in_word_face * alpha_at_face[:,:,:,1] + in_audio_face * alpha_at_face[:,:,:,0]
alpha_at_body = torch.cat([in_word_body, in_audio_body], dim=-1).reshape(bs, t, c*2)
alpha_at_body = self.at_attn_body(alpha_at_body).reshape(bs, t, c, 2)
alpha_at_body = alpha_at_body.softmax(dim=-1)
fusion_body = in_word_body * alpha_at_body[:,:,:,1] + in_audio_body * alpha_at_body[:,:,:,0]
else:
fusion_face = in_word_face + in_audio_face
fusion_body = in_word_body + in_audio_body
masked_embeddings = self.mask_embeddings.expand_as(in_motion)
masked_motion = torch.where(mask == 1, masked_embeddings, in_motion) # bs, t, 256
body_hint = self.motion_encoder(masked_motion) # bs t 256
speaker_embedding_face = self.spearker_encoder_face(in_id).squeeze(2)
speaker_embedding_body = self.spearker_encoder_body(in_id).squeeze(2)
# decode face
use_body_hints = True
if use_body_hints:
body_hint_face = self.bodyhints_face(body_hint)
fusion_face = torch.cat([fusion_face, body_hint_face], dim=2)
a2g_face = self.feature2face(fusion_face)
face_embeddings = speaker_embedding_face
face_embeddings = self.position_embeddings(face_embeddings)
decoded_face = self.face_decoder(tgt=face_embeddings, memory=a2g_face)
face_latent = self.face2latent(decoded_face)
cls_face = self.face_classifier(face_latent)
# motion spatial encoder
body_hint_body = self.bodyhints_body(body_hint)
motion_embeddings = self.feature2motion(body_hint_body)
motion_embeddings = speaker_embedding_body + motion_embeddings
motion_embeddings = self.position_embeddings(motion_embeddings)
# bi-directional self-attention
motion_refined_embeddings = self.motion_self_encoder(motion_embeddings)
# audio to gesture cross-modal attention
if use_word :
a2g_motion = self.audio_feature2motion(fusion_body)
motion_refined_embeddings_in = motion_refined_embeddings + speaker_embedding_body
motion_refined_embeddings_in = self.position_embeddings(motion_refined_embeddings)
word_hints = self.wordhints_decoder(tgt=motion_refined_embeddings_in, memory=a2g_motion)
motion_refined_embeddings = motion_refined_embeddings + word_hints
# feedforward
upper_latent = self.motion2latent_upper(motion_refined_embeddings)
hands_latent = self.motion2latent_hands(motion_refined_embeddings)
lower_latent = self.motion2latent_lower(motion_refined_embeddings)
upper_latent_in = upper_latent + speaker_embedding_body
upper_latent_in = self.position_embeddings(upper_latent_in)
hands_latent_in = hands_latent + speaker_embedding_body
hands_latent_in = self.position_embeddings(hands_latent_in)
lower_latent_in = lower_latent + speaker_embedding_body
lower_latent_in = self.position_embeddings(lower_latent_in)
# transformer decoder
motion_upper = self.upper_decoder(tgt=upper_latent_in, memory=hands_latent+lower_latent)
motion_hands = self.hands_decoder(tgt=hands_latent_in, memory=upper_latent+lower_latent)
motion_lower = self.lower_decoder(tgt=lower_latent_in, memory=upper_latent+hands_latent)
upper_latent = self.motion_down_upper(motion_upper+upper_latent)
hands_latent = self.motion_down_hands(motion_hands+hands_latent)
lower_latent = self.motion_down_lower(motion_lower+lower_latent)
cls_lower = self.lower_classifier(lower_latent)
cls_upper = self.upper_classifier(upper_latent)
cls_hands = self.hands_classifier(hands_latent)
return {
"rec_face":face_latent,
"rec_upper":upper_latent,
"rec_lower":lower_latent,
"rec_hands":hands_latent,
"cls_face":cls_face,
"cls_upper":cls_upper,
"cls_lower":cls_lower,
"cls_hands":cls_hands,
} |