search

wudongming97 / TopoMLP

[ICLR2024] TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning
Apache License 2.0
131 stars 11 forks source link

About the lane to lane topology loss #1

Closed lizimo061 closed 8 months ago

lizimo061 commented 8 months ago

In topo_ll_head part L157, are the targets and rel_preds equations wrong when batch size greater than 1, since the xs_new and ys_new are defined for the last batch?

wudongming97 commented 8 months ago

Our code does not support batch size > 1.

lizimo061 commented 8 months ago

It is said in the literature that experiments are using batch size of 8. Any plan for updating the code?

wudongming97 commented 8 months ago

It means we use a batch size of 1 on each card. Using 8 cards refers to a batch size of 8.

lizimo061 commented 8 months ago

Got it, thanks.

Wolfybox commented 2 months ago

It means we use a batch size of 1 on each card. Using 8 cards refers to a batch size of 8.

@wudongming97 is it possible to extend it to support multi-batches by padding ?

wudongming97 commented 2 months ago

You need to change the topology head to support multi-batches. From my view, lane detection branch has supported multi-batches.