train_configs/minigptv2_finetune.yaml

model:
  arch: minigpt_v2
  model_type: pretrain
  max_txt_len: 1024
  image_size: 448
  end_sym: "</s>"
  llama_model: "/path/to/llama_checkpoint"
  ckpt: "/path/to/pretrained_checkpoint"
  use_grad_checkpoint: True
  chat_template: True
  lora_r: 64
  lora_alpha: 16

datasets:
  multitask_conversation:
    batch_size: 2
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 50

  llava_conversation: 
    batch_size: 2
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 30

  unnatural_instruction:
    batch_size: 1
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 10


  refvg:
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 40

  llava_detail:
    batch_size: 4
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 20

  llava_reason: 
    batch_size: 4
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 80
    

  flickr_grounded_caption:
    batch_size: 2
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 80

  flickr_CaptionToPhrase:
    batch_size: 2
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 80

  flickr_ObjectToPhrase:
    batch_size: 2
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 80

  coco_caption:
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 10  

    
  textcaps_caption:  #
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 30

  refcoco: 
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 25


  refcocop:
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 25

  refcocog:
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 25



  invrefcoco:
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 10

  invrefcocop:
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 10

  invrefcocog:
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 10


  coco_vqa:    
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 15

  ok_vqa:   
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 8

  aok_vqa: 
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 12

  gqa:  
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 50

  ocrvqa: 
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 30


run:
  task: image_text_pretrain
  # optimizer
  lr_sched: "linear_warmup_cosine_lr"
  init_lr: 1e-5
  min_lr: 1e-6
  warmup_lr: 1e-6

  weight_decay: 0.05
  max_epoch: 50
  num_workers: 6
  warmup_steps: 1000
  iters_per_epoch: 1000

  seed: 42
  output_dir: "/path/to/save_checkpoint"

  amp: True
  resume_ckpt_path: null

  evaluate: False 
  train_splits: ["train"]

  device: "cuda"
  world_size: 1
  dist_url: "env://"
  distributed: True

  wandb_log: True
  job_name: minigptv2_finetune