question about batch size and learnrate
gallenszl opened this issue · comments
Hello, I want to know that in the paper you set batch size to 3 for training on 3 1080TI with learning rate 0.001 in seceneflow,but in the README you said that you use 8 GPUS to train(batch size 16) with learning rate 0.001,so which one did you get the best result in kitti?
Thanks
Hi, @gallenszl,
The setting in our paper is quite different from the setting in this new architecture. But it's fine as long as you train our model with lr=0.001 and GPUs=3 or 4, as more GPUs will lead to fewer iterations.
To be clear, I detail the training schedule in this framework here:
GPUs <= 4, we find GPUs=4 works well in this framework.
batch_size=1
For SceneFlow, training with constant lr=0.001 for 20 epochs;
For KITTI, training with start lr=0.001 for 600 epochs, and decay 1/3 at epochs=100, 300.
I hope this information helps you, and I'll release the checkpoint this week for your reimplementation.
Hi, @gallenszl
I'm sorry to tell you that the KITTI 2015 benchmark crashed, framework updating and checkpoint release are both delayed until it works normally.
Thank you for your notice.
updated, enjoy it.