File size: 1,409 Bytes
256a159
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from opencompass.multimodal.models.llava import LLaVAVQAPromptConstructor, LLaVAVSRPostProcessor

# dataloader settings
val_pipeline = [
    dict(type='mmpretrain.LoadImageFromFile'),
    dict(type='mmpretrain.ToPIL', to_rgb=True),
    dict(type='mmpretrain.torchvision/Resize',
         size=(224, 224),
         interpolation=3),
    dict(type='mmpretrain.torchvision/ToTensor'),
    dict(
        type='mmpretrain.torchvision/Normalize',
        mean=(0.48145466, 0.4578275, 0.40821073),
        std=(0.26862954, 0.26130258, 0.27577711),
    ),
    dict(
        type='mmpretrain.PackInputs',
        algorithm_keys=['question', 'gt_answer', 'gt_answer_weight'],
        meta_keys=['question_id', 'image_id'],
    )
]

dataset = dict(type='mmpretrain.VSR',
               data_root='data/vsr/',
               data_prefix='images/',
               ann_file='annotations/test.json',
               pipeline=val_pipeline)

llava_vsr_dataloader = dict(
    batch_size=1,
    num_workers=4,
    dataset=dataset,
    collate_fn=dict(type='pseudo_collate'),
    sampler=dict(type='DefaultSampler', shuffle=False),
)

# model settings
llava_vsr_model = dict(
    type='llava',
    model_path='/path/to/llava',
    prompt_constructor=dict(type=LLaVAVQAPromptConstructor),
    post_processor=dict(type=LLaVAVSRPostProcessor)
)  # noqa

# evaluation settings
llava_vsr_evaluator = [dict(type='mmpretrain.GQAAcc')]