search

Junction4Nako / mvp_pytorch

pytorch implementation of mvp: a multi-stage vision-language pre-training framework
MIT License
33 stars 8 forks source link

VQA finetune: file "test2015_qla_mrcnn.json" are missing #1

Open maxinge8698 opened 2 years ago

maxinge8698 commented 2 years ago

Hi, thank you for releasing the great work! I am working on the VQA task. May I ask where can I find the annotation file "test2015_qla_mrcnn.json" to make inference on the test set? It seems these files do not exist in the download path https://biglmdiag.blob.core.windows.net/vinvl/datasets/vqa/. Thank you very much.

Junction4Nako commented 2 years ago

Thanks for your interest in our work! Yes, the "test2015_qla_mrcnn.json" file is not provided. We constructed this file ourselves.

  1. download the test split of VQA v2.0 from "https://s3.amazonaws.com/cvmlp/vqa/mscoco/vqa/v2_Questions_Test_mscoco.zip",
  2. for each question in the test split, you can define a dict object, here is an example of the first question in test2015 split {'q': 'Is the ball flying towards the batter?', 'o': '', 'an': [], 's': [], 'img_id': 262144, 'q_id': 262144000};
  3. combine all the dict objects into a list, then you can save it as the file you want!

NOTE: for the test split, there are no answers and scores provided, and we will get the object tags from our COCO data, so the values of 'an', 's', 'o' are empty for each question.