MotionLCM / configs /mld_t2m.yaml
wxDai's picture
[Init]
c64dfa4
FOLDER: './experiments_t2m'
TEST_FOLDER: './experiments_t2m_test'
NAME: 'mld_humanml'
SEED_VALUE: 1234
TRAIN:
BATCH_SIZE: 64
SPLIT: 'train'
NUM_WORKERS: 8
PERSISTENT_WORKERS: true
PRETRAINED: 'experiments_recons/vae_humanml/vae_humanml.ckpt'
validation_steps: -1
validation_epochs: 50
checkpointing_steps: -1
checkpointing_epochs: 50
max_train_steps: -1
max_train_epochs: 3000
learning_rate: 1e-4
lr_scheduler: "cosine"
lr_warmup_steps: 1000
adam_beta1: 0.9
adam_beta2: 0.999
adam_weight_decay: 0.0
adam_epsilon: 1e-08
max_grad_norm: 1.0
model_ema: false
model_ema_steps: 32
model_ema_decay: 0.999
VAL:
BATCH_SIZE: 32
SPLIT: 'test'
NUM_WORKERS: 12
PERSISTENT_WORKERS: true
TEST:
BATCH_SIZE: 32
SPLIT: 'test'
NUM_WORKERS: 12
PERSISTENT_WORKERS: true
CHECKPOINTS: 'experiments_t2m/mld_humanml/mld_humanml.ckpt'
# Testing Args
REPLICATION_TIMES: 20
MM_NUM_SAMPLES: 100
MM_NUM_REPEATS: 30
MM_NUM_TIMES: 10
DIVERSITY_TIMES: 300
DO_MM_TEST: true
DATASET:
NAME: 'humanml3d'
SMPL_PATH: './deps/smpl'
WORD_VERTILIZER_PATH: './deps/glove/'
HUMANML3D:
FRAME_RATE: 20.0
UNIT_LEN: 4
ROOT: './datasets/humanml3d'
CONTROL_ARGS:
CONTROL: false
TEMPORAL: false
TRAIN_JOINTS: [0]
TEST_JOINTS: [0]
TRAIN_DENSITY: 'random'
TEST_DENSITY: 100
MEAN_STD_PATH: './datasets/humanml_spatial_norm'
SAMPLER:
MAX_LEN: 200
MIN_LEN: 40
MAX_TEXT_LEN: 20
PADDING_TO_MAX: false
WINDOW_SIZE: null
METRIC:
DIST_SYNC_ON_STEP: true
TYPE: ['TM2TMetrics']
model:
target: ['motion_vae', 'text_encoder', 'denoiser', 'scheduler_ddim', 'noise_optimizer']
latent_dim: [16, 32]
guidance_scale: 7.5
guidance_uncondp: 0.1
t2m_textencoder:
dim_word: 300
dim_pos_ohot: 15
dim_text_hidden: 512
dim_coemb_hidden: 512
t2m_motionencoder:
dim_move_hidden: 512
dim_move_latent: 512
dim_motion_hidden: 1024
dim_motion_latent: 512
bert_path: './deps/distilbert-base-uncased'
clip_path: './deps/clip-vit-large-patch14'
t5_path: './deps/sentence-t5-large'
t2m_path: './deps/t2m/'