MiniGPT-4/train_configs/minigpt_v2_finetune.yaml

model:
  arch: minigpt_v2
  model_type: pretrain
  freeze_vit: True
  freeze_qformer: True
  max_txt_len: 1024
  low_resource: False
  image_size: 448
  end_sym: "</s>"
  llama_model: "/ibex/project/c2133/llama_v2/llama-2-7b-chat-pytorch_update"
  ckpt: "/ibex/project/c2090/minigpt4_ckpt/448_perforamnce_correct_v10_vg/20230925064/checkpoint_32.pth"
  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: 40

  llava_conversation:   # 77k
    batch_size: 2
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 10


  unnatural_instruction:
    batch_size: 1
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 15


  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: #23K
    batch_size: 4
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 20

  llava_reason:   # 77k
    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: 10

  refcoco:  # 142k
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 15


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

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



  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:    # 82K
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 15

  ok_vqa:   # 9k
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 8

  aok_vqa:  # 17k
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 12

  gqa:    # 82K
    batch_size: 6
    vis_processor:
      train:
        name: "blip2_image_train"
        image_size: 448
    text_processor:
      train:
        name: "blip_caption"
    sample_ratio: 40

  ocrvqa:  # 800K
    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: 50
  num_workers: 6
  warmup_steps: 1000
  iters_per_epoch: 1000

  seed: 42
  output_dir: "/ibex/project/c2090/minigpt4_ckpt/448_finetune_test_online"

  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