mirror of
https://github.com/open-compass/opencompass.git
synced 2025-05-30 16:03:24 +08:00
3f601f420b
* [Feat] Add public dataset support of VisualGLM. * [Feat] Refactor LLaVA. * [Feat] Add public dataset support of LlaVA. * [Fix] Add arg.
53 lines
1.5 KiB
Python
53 lines
1.5 KiB
Python
from opencompass.multimodal.models.llava import LLaVABasePromptConstructor, LLaVABasePostProcessor
|
|
|
|
# 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=['image_id']),
|
|
]
|
|
|
|
|
|
dataset = dict(type='mmpretrain.Flickr30kCaption',
|
|
data_root='data/flickr30k',
|
|
ann_file='annotations/dataset_flickr30k.json',
|
|
data_prefix='images',
|
|
split='val',
|
|
pipeline=val_pipeline)
|
|
|
|
llava_flickr30k_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_flickr30k_model = dict(
|
|
type='llava',
|
|
model_path='/path/to/llava',
|
|
is_caption_task=True,
|
|
prompt_constructor=dict(type=LLaVABasePromptConstructor),
|
|
post_processor=dict(type=LLaVABasePostProcessor)
|
|
) # noqa
|
|
|
|
# evaluation settings
|
|
llava_flickr30k_evaluator = [
|
|
dict(
|
|
type='mmpretrain.COCOCaption',
|
|
ann_file='data/flickr30k/annotations/flickr30k_val_gt.json',
|
|
) # noqa
|
|
]
|
|
|
|
|