Skip to content

Latest commit

 

History

History
56 lines (44 loc) · 2.01 KB

data.md

File metadata and controls

56 lines (44 loc) · 2.01 KB
Raw

Data

You can access all the data we utilize by downloading the ZIP file from Google Driver or Baidu Netdisk. This compressed file exclusively contains the annotations in jsonl format. Once you extract this ZIP file, please ensure that you place it under the data folder. After extraction, the directory structure should appear as follows:

|- config
|- mllm
|- data
    |-- blip_laion_cc_sbu_558k.jsonl
    |-- CAP_coco2014_train.jsonl
    |-- CWB_flickr30k_train.jsonl
    ...

Please note that the images can be downloaded separately from their official website. You can update the dataset image_folder configuration in the config/_base_/dataset/DEFAULT_XXXX_XXXXXX.py directory accordingly.

For example, if you are working with the Flickr30k trainset in config/_base_/dataset/DEFAULT_TRAIN_DATASET.py, you can update the image_folder field as follows:

flickr=dict(
    type='FlickrDataset',
    filename=r'{{fileDirname}}/../../../data/CWB_flickr30k_train.jsonl',
    image_folder=r'zz1424:s3://production-public-flickr_image/Flickr_Image/unzip/flickr30k_images/flickr30k_images',
    template_file=r'{{fileDirname}}/template/flickr30k.json',
    ),

to

flickr=dict(
    type='FlickrDataset',
    filename=r'{{fileDirname}}/../../../data/CWB_flickr30k_train.jsonl',
    image_folder=r'path/to/flickr30k_images/on/your/computer',
    template_file=r'{{fileDirname}}/template/flickr30k.json',
    ),

if you are working with the pope test set in config/_base_/dataset/DEFAULT_TEST_POPE_VARIANT.py, you can update the image_folder field as follows:

POPE_TEST_COMMON_CFG = dict(
    type='POPEVQADataset',
    image_folder=r'openmmlab1424:s3://openmmlab/datasets/detection/coco/val2014',
)

to

POPE_TEST_COMMON_CFG = dict(
    type='POPEVQADataset',
    image_folder=r'path/to/coco/val2014/on/your/computer',
)