TANGO / configs /gradio_speaker1.yaml
H-Liu1997's picture
init
31f2f28
wandb_project: 'TANGO'
exp_name: 'debug'
wandb_entity: ''
wandb_key: ""
wandb_log_dir: '/content/outputs/wandb'
output_dir: ./outputs/
log_period: 1
seed: 42
data:
name_pyfile: "datasets.beat2_v5"
class_name: "BEAT2Dataset"
train_bs: 2
meta_paths:
- "./datasets/data_json/show-oliver-s40_w128.json"
# test_meta_paths: "./datasets/data_json/show_oliver_test/Abortion_Laws_-_Last_Week_Tonight_with_John_Oliver_HBO-DRauXXz6t0Y.webm.json"
"test_meta_paths": "./datasets/data_json/youtube_test/speaker1.json"
pose_norm: False
pose_fps: 30
rot6d: True
pose_dims: 825
pose_length: 128
stride: 20
test_length: 128
audio_sr: 16000
audio_fps: 16000
model:
name_pyfile: "models.jointembedding_high_env0"
class_name: "JointEmbedding"
motion_f: 256
audio_rep: wave16k
audio_sr: 16000
audio_fps: 16000
audio_norm: False
audio_f: 256
word_rep: textgrid
word_index_num: 11195
word_dims: 300
facial_rep: smplxflame_30
facial_dims: 100
facial_norm: False
facial_f: 0
f_pre_encoder: null
f_encoder: null
f_fix_pre: False
id_rep: onehot
speaker_f: 0
hidden_size: 512
n_layer: 1
motion_dim: 825
validation:
val_loss_steps: 1
validation_steps: 1000
# guidance_scale: 3.5
# denoising_steps: 20
solver:
gradient_accumulation_steps: 1
# mixed_precision: 'fp16'
# enable_xformers_memory_efficient_attention: True
gradient_checkpointing: False
max_train_steps: 5000000
max_grad_norm: 1.0
# lr
learning_rate: 2e-5
scale_lr: False
lr_warmup_steps: 50
lr_scheduler: 'constant'
# optimizer
use_8bit_adam: False
adam_beta1: 0.9
adam_beta2: 0.999
adam_weight_decay: 1.0e-2
adam_epsilon: 1.0e-8