training object class

[btjhjeon]

In your code, the linear layer of object classifier (net.fc_obj) is untrainable because you set net.fc_obj untrainable and net.fc_obj layer is not included to the optimizer. Beside, the pretrained model (VGG_imagenet.npy) does not contains the parameters of net.fc_obj. How can this model classify the object properly? thank you in advance.

[GriffinLiang]

For relationship detection, our proposed pipeline contains two stages. The first stage is object detection which provides the object prediction. Therefore, in the second stage, net.fc_obj is not included for training.

[btjhjeon]

you mean that the obj_score (the output of net.fc_obj) does not affect the recall performance, right?

[GriffinLiang]

Yes

[btjhjeon]

thanks a lot for prompt response!