There is a checkpoint named ''vinvl_vg_x152c4.pth'', and I have tried loading the checkpoint. I found the num_classes in roi heads is 1595, which is the classes number of VG. I have some questions about it:
Is the od model is trained only on VG, or trained on FourSets( which is mentioned in the paper) and fine-tuned on VG?
Is this model also used to extract features of Pre-exacted Image Features which is posted below Pre-trained Models?
And how should this model be used if I want to extract features by myself?
Thank you for sharing your working on VinVL. I have downloaded the od_models using:
There is a checkpoint named ''vinvl_vg_x152c4.pth'', and I have tried loading the checkpoint. I found the num_classes in roi heads is 1595, which is the classes number of VG. I have some questions about it:
Thanks!