lavis/projects/albef/train/vqa_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: albef_vqa
  model_type: vqav2
  load_finetuned: False

  image_size: 384

datasets:
  coco_vqa: # name of the dataset builder
    vis_processor:
        train:
          name: "blip_image_train"
          image_size: 384
        eval:
          name: "blip_image_eval"
          image_size: 384
    text_processor:
        train:
          name: "blip_question"
        eval:
          name: "blip_question"
  vg_vqa: # name of the dataset builder
    vis_processor:
        train:
          name: "blip_image_train"
          image_size: 384
    text_processor:
        train:
          name: "blip_question"

run:
  task: vqa
  # optimization-specific
  lr_sched: "linear_warmup_cosine_lr"
  init_lr: 2e-5
  min_lr: 1e-6
  weight_decay: 0.02
  max_epoch: 8
  batch_size_train: 32
  batch_size_eval: 64
  num_workers: 4

  # inference-specific
  max_len: 10
  min_len: 1
  num_beams: 3
  num_ans_candidates: 128
  inference_method: "rank"

  seed: 42
  output_dir: "output/ALBEF/VQA"

  amp: False
  resume_ckpt_path: null

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

  # distribution-specific
  device: "cuda"
  world_size: 1
  dist_url: "env://"
  distributed: True