Compute

We aim to support local, ssh, cloud and HPC environments.

SSH

Use tmux for long-running sessions.

Google Colab

Not very performant, but enough to run benchmarks. For an example, see this eval_coco notebook that runs a full COCO keypoint benchmark.

Google Cloud

This gives you a V100 GPU to benchmark OpenPifPaf. You can ssh into the box directly and use tmux. Not all regions/zones support recent GPUs.

gcloud compute instances create instance-2 \
  --zone=europe-west4-a \
  --image-family=pytorch-latest-gpu \
  --image-project=deeplearning-platform-release \
  --machine-type=n1-standard-4 \
  --maintenance-policy=TERMINATE \
  --accelerator="type=nvidia-tesla-v100,count=1" \
  --metadata="install-nvidia-driver=True" \
  --preemptible

HPC

See Training document.