Giters

sxjscience / HKO-7

Source code of paper "[NIPS2017] Deep Learning for Precipitation Nowcasting: A Benchmark and A New Model"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

GPU memory and the speed of TrajGRU

cunwang-root opened this issue · comments

commented

  1. Could you talk more about the experiment settings especially the GPU memory required?

  2. I implement TrajGRU in tensorflow and I notice that it runs much slower than ConvGRU. Do you have similar situation?

commented

Yes, it will be slower and cost more GPU memory if you implement it using the same approach as in this repo. One solution to accelerate the speed and reduce the memory cost is to write your own kernel which combines "k warp" + "concat" + "conv1x1". (I may add one later but I'm currently busy with other works).

commented

Then could you tell me how much memory is required in your implementation? I can only run a single layer of TrajGRU with 9 links on sequence with total length 16 on a GeForce GTX 1080 Ti GPU (11 GB), it will run out of memory if I try with larger links. I'm not sure if it's the problem of my implementation or TrajGRU requires such much memory.

commented

I use a single GTX1080 (no Ti) to do the MNIST++ experiment and use two GTX1080s to do the HKO7 experiment.