src/configs/config_dit_base.yaml

image_size: 128
output_dir: ./outputs/dit_${model_config.n_layers}layers_${pc_size}p_${train_dataset.category}_${train_dataset.n_training_frames}frames_pointembed_latent${model_config.latent_dim}_deform${lambda_deform}_8gpus_base
seed: 0
train_batch_size: 2
eval_batch_size: 4
num_train_epochs: 50
max_train_steps: 60000
gradient_accumulation_steps: 25
gradient_checkpointing: true
learning_rate: 1e-4
scale_lr: false
lr_scheduler: "costant_with_warmup"
lr_warmup_steps: 100
use_8bit_adam: false
allow_tf32: true
dataloader_num_workers: 48
adam_beta1: 0.9
adam_beta2: 0.999
adam_weight_decay: 1.e-2
adam_epsilon: 1.e-08
max_grad_norm: 1.0
prediction_type: null
vis_dir: vis
logging_dir: logs
mixed_precision: 'bf16'
report_to: 'tensorboard'
local_rank: -1
checkpointing_steps: 2500
checkpoints_total_limit: 10
resume_from_checkpoint: latest
enable_xformers_memory_efficient_attention: true
validation_steps: 500
validation_train_steps: 2000
validation_sanity_check: true
tracker_project_name: 'diffusion'
push_to_hub: false
set_grads_to_none: true
lambda_vel: 1.0
lambda_mask: 0.0
lambda_momentum: 0.0
lambda_deform: 0.001
pc_size: 2048
condition_drop_rate: 0.0
model_type: 'dit_st'
model_config:
  n_layers: 8
  latent_dim: 256
  frame_cond: true
  point_embed: true
  mask_cond: false
  pred_offset: true
  num_neighbors: -1
  floor_cond: false
  max_num_forces: 1
  force_as_token: false
  force_as_latent: false
  coeff_cond: false
  class_token: false
  transformer_block: SpatialTemporalTransformerBlock
train_dataset:
  category: hf-objaverse-v1
  dataset_path: DATASET_FOLDER
  dataset_list: DATASET_ITEM_LIST
  has_gravity: true # no gravity
  max_num_forces: ${model_config.max_num_forces}
  norm_fac: 5
  stage: 'deform'
  mode: 'diff'
  pc_size: ${pc_size}
  repeat: 1
  seed: 0
  n_sample_pro_model: 300
  n_frames_interval: 2
  n_training_frames: 24
  batch_size: 20
  overfit: false