cciiplab / dpt Goto Github PK

The BaiduYun links in README has been cancelled. The new link is: https://pan.baidu.com/s/1HhPSum0aCqvEtgCe1Yxl0Q?pwd=8888

Hi,

The link to download object features is unavailable now (microsoft/Oscar#198).
Could you provide the object feature file of GQA and COCO?
Thanks a lot!

Hi,
Thanks for the open-source code provided in this repository for the paper, and I'm very interested in the zero/few-shot ability as shown in your great work, but as I investigate to evaluate the model under zero-shot MLM+ITM setting in run_gqa_prompt_zero_few.py, I found that the ITM head (i.e. model.cls_ans.seq_relationship) is newly initialized (and of different shape from original pretrained ITM head) and actually not used in the evaluation. Am I wrong on the usage of this code or is ITM head not participating in the evaluation?
Thanks a lot for your time and attention.

Thank you very much for provide this repo! When I ran the training by bash run_train.sh, I got the following error.

How to get the sacrebleu.py ? Look forward to your favourable reply.

你好，在finetuning_T5.py文件中提示import datasets和from datasets import load_dataset, load_metric错误，请问可以在哪获取datasets这个包呢？谢谢

Thank you for providing your project. But I need about 6 hours to train one epoch of MLM tasks even if per_gpu_train_batch_size is halved to 16. I ran on the single 3090. Is this normal? Look forward to your reply.

Does the model supports declaration generation on custom datasets such as CLEVRER, or just can be applied on GQA and VQA v2.0?