Add a Tensorflow with GPU template

[MarkEdmondson1234]

Like this one https://tensorflow.rstudio.com/blog/rstudio-gpu-paperspace.html

[MarkEdmondson1234]

Rocker-org are supporting a GPU template now so will use that:
https://github.com/rocker-org/ml/tree/master/gpu

Close to this working:

library(googleComputeEngineR)
gce_global_project("mark-edmondson-gde")
gce_global_zone("europe-west4-b")

# https://cloud.google.com/deep-learning-vm/docs/quickstart-cli
IMAGE_FAMILY="tf-latest-cu92"
IMAGE_PROJECT="deeplearning-platform-release"
INSTANCE_NAME="gpu-vm"
INSTANCE_TYPE="n1-standard-8"


# how to keep GPU available when you start and stop instance?
gce_vm(INSTANCE_NAME,
       predefined_type = INSTANCE_TYPE,
       image_project = IMAGE_PROJECT,
       image_family = IMAGE_FAMILY,
       acceleratorCount = 1,
       acceleratorType = "nvidia-tesla-p4",
       disk_size_gb=120,
       scheduling = list(
         onHostMaintenance = "TERMINATE",
         automaticRestart = FALSE
       ),
       metadata = list("install-nvidia-driver" = "True"))

# wait for it to launch

dpvm <- gce_get_instance(INSTANCE_NAME)

docker_run(dpvm,
           image = "rocker/ml",
           docker_opts = "-e USER=gpu -e PASSWORD=gpu -d -p 80:8787 -v ~/:/home/rstudio",
           name = "gpu-r",
           detach = TRUE,
           nvidia = TRUE)

[MarkEdmondson1234]

Template code:

library(googleComputeEngineR) # assume auto-auth, project settings etc

# set up VM with nvidia-tesla-p4 GPU, RStudio with Tensorflow, keras etc based on rocker/ml
vm <- gce_vm(template = "rstudio-gpu",
             name = "deeplearning-ml",
             username = "mark", password = "mark1234")

[MarkEdmondson1234]

Well seems nvidia-docker is not on the container optimised OS's (meh) so need to hack through this
https://github.com/GoogleCloudPlatform/cos-gpu-installer or wait for it to be supported, or just use normal VMs and install docker myself to run the startup script.