Open jjl212 opened 2 months ago
The VG annotation file is the same as the initial annotation from vg_motif_anno. The codebase will select different splits for training and testing. You can create those files through soft links.
The test set of our work in PSG is the official validation set of PSG. The validation set is randomly selected from the official training set by us. You can rename the official split and create your own validation set.
Another file is missing during the test: zeroshot_triplet_lavis.pytorch
The VG annotation file is the same as the initial annotation from vg_motif_anno.
Sorry I did not get which file you are referring. what should VG-SGG-train.h5
be create from?
The VG annotation file is the same as the initial annotation from vg_motif_anno.
Sorry I did not get which file you are referring. what should
VG-SGG-train.h5
be create from?
I assume that all these files (VG-SGG-train.h5
, VG-SGG-test.h5
, VG-SGG-val.h5
) are created as soft links to the VG-SGG-with-attri.h5
file. The codebase automatically selects the appropriate data splits for further processing.
Another file is missing during the test: zeroshot_triplet_lavis.pytorch
This file stores the triplet combination that is unseen in the training set and shares the same format with the previous zero-shot SGG codebase. I will update this file later.
The VG annotation file is the same as the initial annotation from vg_motif_anno.
Sorry I did not get which file you are referring. what should
VG-SGG-train.h5
be create from?
The VG-SGG-train.h5
is created from the initial annotation of the VG dataset: vg_motif_anno.h5
.
I couldn't find the file VG-SGG-train_zs_pred.h5/VG-SGG-train.h5 in the download link .Is this file the same as 'VG-SGG-with-attri.h5'? Or is VG-SGG-train_zs_pred.h5/VG-SGG-train.h5 a dataset file that has been repartitioned based on VG-SGG-with-attri.h5? I have the same confusion regarding the val and test dataset files named vg/vg_motif_anno/VG-SGG-val.h5 and vg/vg_motif_anno/VG-SGG-test.h5 in lavis/configs/datasets/vg/defaults_rel_det_train_zs_pred.yaml. Looking forward to your response.