search

Henry1iu / TNT-Trajectory-Prediction

A Unofficial Pytorch Implementation of TNT: Target-driveN Trajectory Prediction
504 stars 95 forks source link

questions about subgraph implementation #44

Open ImCabbage opened 1 year ago

ImCabbage commented 1 year ago

Hi, Great implementation of VectorNet. I have one doubt about the subgraph part. The subgraph is stacked in a row to encoding the features. Then why there is a hidden unit coded in this part. From the perspective of the paper, should't the dimensions keeps the same? What is the benefits?

Henry1iu commented 1 year ago

Hi,

Thanks for your appreciation. Could you please explain what you mean by "a hidden unit"?

Best Regards, Henry

ImCabbage commented 1 year ago

In the implementaion of subgraph, I see the "in_channels" variable is set as two times of hidden_unit, and at end step, you have one more linear layer added which is not mentioned in the paper. I just wanna know what is the idea behind such design.

Henry1iu commented 1 year ago

The idea came from the bottleneck of the Resnet block.