Open jwyang opened 6 years ago
The training is different for the three settings, the network is either provided ground truth detections as input (in PR for example) or needs to produce detections from scratch (for SG). This changes the network architecture slightly.
Hi, thanks for sharing the code.
I have a question about the training settings. To train the model, we need to specify the sg_task to one of three options [PR|CL|SG]. I am not familiar with tensorflow. Maybe I am wrong, but I am wondering why it is needed to specify the sg_task during training. As far as I know about the scene graph generation task, the three options are merely specified during evaluation, which means that the training should be the same to different settings, right?
thanks in advance for your reply.