lavis/projects/blip2/train/retrieval_coco_ft.yaml

 # 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
  model_type: coco
  load_pretrained: True
  load_finetuned: False 
  freeze_vit: False
  
  # # finetune blip2 with clip-vit-large
  # use_grad_checkpoint: False
  # pretrained: "https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/blip2_pretrained_vitL.pth"
  # vit_model: clip_L  

datasets:
  coco_retrieval: # name of the dataset builder
    vis_processor:
        train:
          name: "blip2_image_train"
          image_size: 364
        eval:
          name: "blip_image_eval"
          image_size: 364
    text_processor:
        train:
          name: "blip_caption"
        eval:
          name: "blip_caption"
    build_info:
        images:
            storage: '/export/share/datasets/vision/coco/images/'

run:
  task: retrieval
  # optimizer
  lr_sched: "linear_warmup_cosine_lr"
  init_lr: 1e-5
  min_lr: 0
  warmup_lr: 1e-8
  warmup_steps: 1000
  weight_decay: 0.05
  max_epoch: 5
  batch_size_train: 14
  batch_size_eval: 16
  lr_layer_decay: 0.95 # layer-wise learning rate decay for the ViT 
  num_workers: 4
  accum_grad_iters: 1

  seed: 42
  output_dir: "output/BLIP2/Retrieval_coco"

  amp: True
  resume_ckpt_path: null

  evaluate: False
  train_splits: ["train"]
  valid_splits: ["test"]
  # test_splits: ["test"]
  k_test: 128

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