Closed itsbman closed 1 week ago
@itsbman As mentioned in paper, since the frequency of the dataset is 10Hz instead of 2Hz in nuScenes, there is lots of very similar samples. Thus, we only train 1/10 epochs - 6 epochs. The loss is normal.
@itsbman Hi, you may modify the lr, epoch_number, and dropout as in the updated config file.
That makes sense. Thanks
Hi, thank you for sharing this work!
I want to know how long does it typically take to train VAD for 60 epochs on the base dataset?
I'm training with 8 A6000 gpus and the eta is around 32 days