MiniGPT-4/minigpt4/projects/minigpt/train/minigptv2_finetune.yaml

model:
  arch: minigpt_v2
  model_type: pretrain
  max_txt_len: 1024
  image_size: 448
  end_sym: "</s>"
  llama_model: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/llama_2_7b_chat"
  ckpt: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/minigptv2/checkpoint_stage2.pth"
  use_grad_checkpoint: True
  chat_template: True
  lora_r: 64
  lora_alpha: 16

datasets:
  # multitask_conversation: # in-house data 12171
  #   batch_size: 2
  #   vis_processor:
  #     train:
  #       name: "blip2_image_train"
  #       image_size: 448
  #   text_processor:
  #     train:
  #       name: "blip_caption"
  #   sample_ratio: 50

  llava_conversation: # 56681
    batch_size: 2
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 30

  unnatural_instruction: # pure text 65852
    batch_size: 1
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 10


  # refvg: # [refer] return the location
  #   batch_size: 6
  #   vis_processor:
  #     train:
  #       name: "blip2_image_train"
  #       image_size: 448
  #   text_processor:
  #     train:
  #       name: "blip_caption"
  #   sample_ratio: 40

  llava_detail: # 23240
    batch_size: 4
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 20

  llava_reason: # 76643
    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: # [grounding] : TODO
  #   batch_size: 2
  #   vis_processor:
  #     train:
  #       name: "blip2_image_train"
  #       image_size: 448
  #   text_processor:
  #     train:
  #       name: "blip_caption"
  #   sample_ratio: 80

  # flickr_CaptionToPhrase: # [detection]
  #   batch_size: 2
  #   vis_processor:
  #     train:
  #       name: "blip2_image_train"
  #       image_size: 448
  #   text_processor:
  #     train:
  #       name: "blip_caption"
  #   sample_ratio: 80

  # flickr_ObjectToPhrase: # [detection]
  #   batch_size: 2
  #   vis_processor:
  #     train:
  #       name: "blip2_image_train"
  #       image_size: 448
  #   text_processor:
  #     train:
  #       name: "blip_caption"
  #   sample_ratio: 80

  coco_caption: # 414113 train
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 10  

    
  textcaps_caption:  # 109765 train
    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:    # 658104
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 15

  ok_vqa: # train, valid (9009, 5046)
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 8

  aok_vqa: # (17056, 1145, 6702)
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 12

  gqa:  # (943000, 12578, 12578)
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 50

  ocrvqa: # 207572
    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: 8e-5
  warmup_lr: 1e-6

  weight_decay: 0.05
  max_epoch: 1
  num_workers: 6
  warmup_steps: 1000
  iters_per_epoch: 1000

  seed: 42
  output_dir: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/minigpt/v2/vqa_pretrain_3B_llama2_7b_chat_stage3_train_linear_lora_test_1030"

  amp: True
  resume_ckpt_path: null

  evaluate: False 
  train_splits: ["train"]
  valid_splits: ["val"]
  test_splits: ["test"]

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

  wandb_log: True
  job_name: minigptv2_finetune