Batch Time is very slow (around 4 sec) with 4 GPUs (1080ti)

[Chandradeep Pokhariya]

BatchTime is very slow. I am using batch size of 12 as of now.

[Chuxin Wang]

We use 4 Tesla V100 GPUs to train our model. With the batch size 32, it takes an average of 1.2 seconds to finish one batch. During the training, we need to calculate EMD loss and also render depth map, so the training process can be time-consuming.