scwangdyd / promting_hoi Goto Github PK

hi, how to generate unseen interactions in your swig_hoi datasets?

When inferencing, does the network combine all possible interaction and object categories into text encoder to get the output feature and then calculate the similarity with the predicted feature of visual encoder? does it put all the combinations into it? Is it too computationally intensive? I can't find the relevant explanation in the paper, please ask for the answer, thank you

There are eight 3080 video cards in my server, all of which have 10GB of memory. It's strange that I can train with a single card (batch size is set to 1), but I can't train with multiple cards in a single computer using the author's code. Does anyone know why?

Command：python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py
--batch_size 1
--output_dir runs
--epochs 100
--lr 1e-4 --min-lr 1e-7
--enable_dec
--dataset_file hico

Has anyone ever encountered this problem during training？

"RuntimeError: expected scalar type Float but found Half"
I am getting the above error shown in bold when I run the inference on HICO-DET using the pretrained weights given in your repo. Could you please help?

Attaching the traceback for your reference ->

Here is the inference command I used (referring to your repo's README) -
python main.py --eval --batch_size 1 --output_dir hico_results --hoi_token_length 50 --enable_dec --pretrained downloaded_checkpoint/thid_hico_token50_epoch100.pth --eval_size 256 --test_score_thresh 1e-4 --dataset_file hico