MiniGPT_3D_conv_UI_demo.yaml

model:
  arch: minigpt_3d
  model_type: pretrain
  max_txt_len: 160
  end_sym: "###"
  prompt_template: 'Instruct: {} /n Output: '
  llama_model: "./params_weight/Phi_2"
  ckpt: './params_weight/MiniGPT_3D_stage_3/MiniGPT_3D_stage_3.pth'
  second_ckpt: "./params_weight/MiniGPT_3D_stage_4/MiniGPT_3D_stage_4.pth"
  lora_r: 64
  lora_alpha: 16
  pc_linear_layer: 2
  use_MQE: True  #  False: not use  query expert
  query_expert_router_type: 'Sparse Router'  #  three types： Constant Router, Sparse Router,  Soft Router
  query_expert_num: 8
  query_expert_top_k: 2  #  the top_k numbers of  query expert in 'Sparse Router'
  QFormer_lora_r: 8   # 0 : not use QFormer lora
  QFormer_lora_alpha: 16
  QFormer_lora_module: [ "query", "key",  "value" ]


datasets:
  Objaverse_brief:
    batch_size: 9   # 9
    text_processor:
      train:
        name: "blip_caption"

run:
  task: image_text_pretrain