BLIVA/train_configs/finetune_bliva_flant5.yaml at main · mlpc-ucsd/BLIVA · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
model:
  arch: bliva_flant5
  model_type: flant5xxl
  load_finetuned: False
  load_pretrained: True

  pretrained: "please specify the path to the pretrained model"
  finetuned: ""

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

  # Q-Former
  num_query_token: 32

  # T5
  t5_model: "google/flan-t5-xxl"

  # generation configs
  prompt: ""

datasets:
  ocrvqa:
    vis_processor:
        train:
          name: "blip2_image_train"
          image_size: 224
    text_processor:
        train:
          name: "blip_caption"
  coco_vqa:
    vis_processor:
        train:
          name: "blip_image_train"
          image_size: 224
    text_processor:
        train:
          name: "blip_caption"
  ok_vqa:
    vis_processor:
        train:
          name: "blip_image_train"
          image_size: 224
    text_processor:
        train:
          name: "blip_caption"
  aok_vqa:
    vis_processor:
        train:
          name: "blip_image_train"
          image_size: 224
    text_processor:
        train:
          name: "blip_caption"
  coco_caption:
    vis_processor:
        train:
          name: "blip_image_train"
          image_size: 224
    text_processor:
        train:
          name: "blip_caption"
  llavavqa:
    vis_processor:
        train:
          name: "blip_image_train"
          image_size: 224
    text_processor:
        train:
          name: "blip_caption"
  textcaps:
    vis_processor:
        train:
          name: "blip_image_train"
          image_size: 224
    text_processor:
        train:
          name: "blip_caption"
  vqg_coco_vqa:
    vis_processor:
        train:
          name: "blip_image_train"
          image_size: 224
    text_processor:
        train:
          name: "blip_caption"
  vqg_ok_vqa:
    vis_processor:
        train:
          name: "blip_image_train"
          image_size: 224
    text_processor:
        train:
          name: "blip_caption"
  vqg_aok_vqa:
    vis_processor:
        train:
          name: "blip_image_train"
          image_size: 224
    text_processor:
        train:
          name: "blip_caption"

run:
  runner: runner_iter

  task: image_text_pretrain
  # optimizer
  lr_sched: "linear_warmup_cosine_lr"
  init_lr: 1e-5
  min_lr: 0
  warmup_lr: 1e-8
  accum_grad_iters: 1

  weight_decay: 0.05
  max_epoch: 2

  max_iters: 200000
  iters_per_inner_epoch: 100000

  batch_size_train: 3
  batch_size_eval: 1
  num_workers: 8
  warmup_steps: 1000

  seed: 42
  output_dir: "output/finetuned_bliva_flant5/"

  amp: True
  resume_ckpt_path: null

  evaluate: False
  train_splits: ["train"]

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

  train_dataset_ratios: {ocrvqa:  0.20482112476204395, coco_vqa: 0.15271327995696837,
    ok_vqa: 0.02175930084085304, aok_vqa: 0.02993954782443368, coco_caption: 0.17258428920102808, llavavqa: 0.1378184390146113,
    textcaps: 0.07595188977780638,
    vqg_coco_vqa: 0.15271327995696837, vqg_ok_vqa: 0.02175930084085304 , vqg_aok_vqa:  0.02993954782443368 }