config.yaml

dataset: point_maze
data_path: /data/user_data/mchen5/dino_wm_dataset/point_maze
pretrained_model_path: /data/hf_cache/hub/models--maxin-cn--Latte/snapshots/fc35c8c44c38e0b446158ace47e7de642066f0d9/vae
n_rollout: null
split_ratio: 0.9
validation_size: 32
use_wandb: true
results_dir: ./results
pretrained: null
model: Latte-S/2
use_dynamic_tokenizer: true
num_slots: 256
slot_dim: 16
dec_hidden_size: 384
dec_depth: 12
dec_num_heads: 6
num_frames: 4
n_context_frames: 1
scheduling_mode: full_sequence
image_size: 256
num_sampling_steps: 250
frame_interval: 1
fixed_spatial: false
attention_bias: true
learn_sigma: false
diffusion_mode: diffusion_forcing
extras: 1
use_action: true
action_dim: 2
normalize_action: true
save_ceph: false
learning_rate: 0.0001
clip_max_norm: 0.1
start_clip_iter: 20000
local_batch_size: 16
max_train_steps: 1000000
global_seed: 3407
num_workers: 4
log_every: 10
val_every_n_steps: 1000
lr_warmup_steps: 0
resume_from_checkpoint: null
gradient_accumulation_steps: 1
num_classes: null
use_compile: false
mixed_precision: 32
gradient_checkpointing: true
sequential_loss_grad: true
callbacks:
  model_checkpoint_across_timesteps:
    every_n_train_steps: 10000
    save_top_k: -1
    save_on_train_epoch_end: true
    save_weights_only: false
    filename: '{epoch}-{step}'
  model_checkpoint_latest:
    every_n_train_steps: 1000
    save_top_k: 1
    save_on_train_epoch_end: false
    save_weights_only: false
    filename: latest-{epoch}-{step}
  metrics_logger:
    log_every_n_train_steps: 10000
    save_videos: true
    evaluate: true
    max_batchsize: 4
    buffer_size: 32
    i3d_model_path: checkpoints/i3d/i3d_torchscript.pt
    inference_n_slots:
    - 64
    - 128
    - 256
train: true
validate: false
exp_name: tok2_comp=False_mp=32_gc=True
run_id: 6002143-0