configs/emage.yaml

is_train: True
ddp: False
stat: ts
root_path: ./
out_path: ./outputs/audio2pose/
project: s2g
data_path: ./BEAT2/beat_english_v2.0.0/
e_path:  weights/AESKConv_240_100.bin
eval_model: motion_representation
e_name: VAESKConv
test_ckpt: ./EMAGE/emage_240.bin
data_path_1: ./EMAGE/
vae_test_len: 32
vae_test_dim: 330
vae_test_stride: 20
vae_length: 240
vae_codebook_size: 256
vae_layer: 4
vae_grow: [1,1,2,1]
variational: False

# data config
training_speakers: [2] #[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]
additional_data: False
cache_path: datasets/beat_cache/beat_smplx_en_emage/
dataset: beat_sep_lower
new_cache: False

# motion config
ori_joints: beat_smplx_joints
tar_joints: beat_smplx_full
pose_rep: smplxflame_30
pose_norm: False
pose_fps: 30
rot6d: True
pre_frames: 4
pose_dims: 330
pose_length: 64
stride: 20
test_length: 64
motion_f: 256
m_pre_encoder: null
m_encoder: null
m_fix_pre: False

# audio config
audio_rep: onset+amplitude
audio_sr: 16000
audio_fps: 16000
audio_norm: False
audio_f: 256
# a_pre_encoder: tcn_camn
# a_encoder: none
# a_fix_pre: False

# text config
word_rep: textgrid
word_index_num: 11195
word_dims: 300
freeze_wordembed: False
word_f: 256
t_pre_encoder: fasttext
t_encoder: null
t_fix_pre: False

# facial config
facial_rep: smplxflame_30
facial_dims: 100
facial_norm: False
facial_f: 0
f_pre_encoder: null
f_encoder: null
f_fix_pre: False

# speaker config
id_rep: onehot
speaker_f: 0

# model config
batch_size: 64
# warmup_epochs: 1
# warmup_lr: 1e-6
lr_base: 5e-4
model: emage
g_name: MAGE_Transformer
trainer: emage
hidden_size: 768
n_layer: 1
 
rec_weight: 1
grad_norm: 0.99
epochs: 400
test_period: 20
ll: 3
lf: 3
lu: 3
lh: 3
cl: 1
cf: 0
cu: 1
ch: 1