configurations/huggingface.yaml

n_tokens: 3
pose_cond_dim: 5
use_plucker: true
focal_length: 0.35
customized_validation: true
condition_similar_length: 8
log_video: true
relative_embedding: true
cond_only_on_qk: true
add_pose_embed: false
use_domain_adapter: false
use_reference_attention: true
add_frame_timestep_embedder: true
is_interactive: true
diffusion:
  sampling_timesteps: 20
  beta_schedule: sigmoid
  objective: pred_v
  use_fused_snr: True
  cum_snr_decay: 0.96
  clip_noise: 20.
  ddim_sampling_eta: 0.0
  stabilization_level: 15
  schedule_fn_kwargs: {}
  use_snr: False
  use_cum_snr: False
  snr_clip: 5.0
  timesteps: 1000
  # architecture
  architecture:
    network_size: 64
    attn_heads: 4
    attn_dim_head: 64
    dim_mults: [1, 2, 4, 8]
    resolution: ${dataset.resolution}
    attn_resolutions: [16, 32, 64, 128]
    use_init_temporal_attn: True
    use_linear_attn: True
    time_emb_type: rotary

weight_decay: 2e-3
warmup_steps: 10000
optimizer_beta: [0.9, 0.99]
action_cond_dim: 25
n_frames: 8
frame_skip: 1
frame_stack: 1
uncertainty_scale: 1
guidance_scale: 0.0
chunk_size: 1 # -1 for full trajectory diffusion, number to specify diffusion chunk size
scheduling_matrix: autoregressive
noise_level: random_all
causal: True
x_shape: [3, 360, 640]
context_frames: 1
diffusion_path: yslan/worldmem_checkpoints/diffusion_only.ckpt
vae_path: yslan/worldmem_checkpoints/vae_only.ckpt
pose_predictor_path: yslan/worldmem_checkpoints/pose_prediction_model_only.ckpt