dpo_iter_parm_gudie/config.yaml

wandb:
  entity: null
  resume: auto
  run_id: c27m34aq
experiment:
  project: training
  name: show-o-dpo
  output_dir: train_set_hq_prune_full_t2i_dpo_10k_1e-5_v2_ad_orm_RM
  save_every: 10000
  eval_every: 2500
  generate_every: 1000
  log_every: 1
  log_grad_norm_every: 20
  logging_dir: train_set_hq_prune_full_t2i_dpo_10k_1e-5_v2_ad_orm_RM/logs
model:
  rw_model: llava-onevision-google_siglip-so400m-patch14-384-Qwen_Qwen2-7B-Instruct-ov_stage_am9--train_ad_orm
  vq_model:
    type: magvitv2
    vq_model_name: showlab/magvitv2
  showo:
    load_from_showo: false
    pretrained_model_path: train_set_hq_prune_full_t2i_dpo_10k_1e-5
    w_clip_vit: false
    vocab_size: 58498
    llm_vocab_size: 50295
    llm_model_path: microsoft/phi-1_5
    codebook_size: 8192
    num_vq_tokens: 256
    num_new_special_tokens: 10
  gradient_checkpointing: true
dataset:
  gen_type: t2i
  params:
    data_path: t2i_dpo_draft.yaml
    validation_prompts_file: validation_prompts/showoprompts.txt
    shuffle_buffer_size: 1000
    num_workers: 32
    resolution: 256
    pin_memory: true
    persistent_workers: true
  preprocessing:
    max_seq_length: 128
    resolution: 256
    center_crop: false
    random_flip: false
optimizer:
  name: adamw
  params:
    learning_rate: 1.0e-05
    scale_lr: false
    beta1: 0.9
    beta2: 0.999
    weight_decay: 0.01
    epsilon: 1.0e-08
lr_scheduler:
  scheduler: cosine
  params:
    learning_rate: ${optimizer.params.learning_rate}
    warmup_ratio: 0.1
training:
  gradient_accumulation_steps: 1
  batch_size_t2i: 10
  mixed_precision: bf16
  enable_tf32: true
  seed: 10086
  overfit_one_batch: false
  cond_dropout_prob: 0.1
  min_masking_rate: 0.0
  label_smoothing: 0.0
  max_grad_norm: null
  guidance_scale: 0.0
  generation_timesteps: 12
  beta: 0.1
  reward_coef: 0.1
  dpo_coef: 1
  sft_coef: 0
  num_epoch: 1
inference:
  generation_timesteps: 18
  guidance_scale: 1.75