How to train with other depth annotations
ruili3 opened this issue · comments
Hi,
I'm following your work and wish to train the model with the dense annotated depth. Can I simply set lidar_depth=true and dso_depth=false in monorec_depth.json file?
Moreover, in the .json file, the batchsize=8 and n_gpu=8, so is there totally 8*8=64 batches during the distributed learning? Since I do not have so many cards, so I cant train the model with 64 total batches a time. With your experience, should I set the learning rate with a higher value or try to train with more epochs than 70? Thank you a lot!
Hi,
in theory, it should basically work out of the box with dense depths, except for the correct loading of the files. (lidar depth often has a weird format).
No, the batch size is total combined over all cards. That means, in this configuration it would be 1 sample per GPU. But the script will take up to 8 GPUs, if they are available. I mostly trained with one or two GPUs.
Best,
Felix