# Copyright (c) 2022, salesforce.com, inc.
 # All rights reserved.
 # SPDX-License-Identifier: BSD-3-Clause
 # For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause

model:
  arch: blip2_t5_instruct_pro_moe
  model_type: flant5xxl
  load_finetuned: False
  load_pretrained: True
  vit_model: eva_clip_g
  pretrained: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/blip2/blip2-flant5-xxl/blip2_pretrained_flant5xxl.pth"
  finetuned: ""

  # vit encoder
  image_size: 224
  drop_path_rate: 0
  use_grad_checkpoint: False
  vit_precision: "fp16"

  # Q-Former
  num_query_token: 32
  qformer_text_input: True

  # T5
  t5_model: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/google-flan-t5-xxl"
  prompt: ""
  max_txt_len: 256
  max_output_txt_len: 256

  # freeze
  freeze_vit: True
  freeze_llm: True
  freeze_qformer: False
  freeze_t5_proj: False

  # moe
  moe_position: "pre" # post (position to insert PromptMoE Part)
  embed_extract: "blip2_pretrain" # t5, random (way to extract embeddings of task instruction if moe_position is pre)
  repeat_to_init_qt_candidates: True
  num_qt_candidates: 20
  moe_topk: 2
  eval_gate_save: False
  train_gate_save: False
  gate_save_path: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/flant5xxl/prompt_moe/llava_st_257k_raw_train_qf_train_qt_linear_gate_textt5_20ex_3loss_textinqf_epo3_1012/"

preprocess:
    vis_processor:
        train:
          name: "blip_image_train"
          image_size: 224
        eval:
          name: "blip_image_eval"
          image_size: 224
    text_processor:
        train:
          name: "blip_caption"
        eval:
          name: "blip_caption"