Cloud Utilities for Deep Learning ⛅️

A super lightweight cloud management tool designed with deep learning applications in mind.

Built with the belief that managing cloud resources should be as easy as:

import cloud

cloud.connect()
train_my_network()
cloud.down()

We welcome all contributions, suggestions, and use-cases. Reach out to us over GitHub or at [email protected] with ideas!

Contents

• Quickstart
  • Install
  • Config
  • Usage
• Documentation
  • Amazon EC2
  • Azure
  • Google Cloud

Quickstart

Install:

Sort of stable:

sudo pip install dl-cloud

Bleeding edge:

git clone git@github.com:for-ai/cloud.git
sudo pip install -e cloud

Config:

See configs/cloud.toml-* for instructions on how to authenticate for each provider (Google Cloud, AWS EC2, and Azure).

Place your completed configuration file (named cloud.toml) in either root / or $HOME. Otherwise, provide a full path to the file in $CLOUD_CFG.

If you use GCP as a provider for your cloud.toml it will use GCP Instance metadata APIs to fetch APIs. If you want to configure for Google Cloud Build, please use;

is_gcb = true
zone = '{{DESIRED_ZONE}}' 

Usage:

GPU

import cloud
cloud.connect()

# gpu instances have a dedicated GPU so we don't need to worry
# about preemption or acquiring/releasing accelerators online.

while True:
  # train your model or w/e

cloud.down()  # stop the instance (does not delete instance)

TPU (Only on GCP)

import cloud
cloud.connect()

tpu = cloud.instance.tpu.get(preemptible=True)  # acquire an accelerator
while True:
  if not tpu.usable:
    tpu.delete(background=True)  # release the accelerator in the background
    tpu = cloud.instance.tpu.get(preemptible=True)  # acquire a new accelerator
  else:
    # train your model or w/e
    
cloud.down()  # release all resources, then stop the instance (does not delete instance)

---

Documentation

cloud.connect()

Takes/Creates a cloud.Instance object and sets cloud.instance to it.

returns	desc.
cloud_env	a cloud.Instance.

cloud.down()

Calls cloud.instance.down().

cloud.delete(confirm=True)

Calls cloud.instance.delete(confirm).

cloud.Resource

Takes/Creates a cloud.Instance object and sets cloud.instance to it.

properties	desc.
name	str, name of the instance
usable	bool, whether this resource is usable
methods	desc.
up(background=False)	start an existing stopped resource
down(background=False)	stop the resource. Note: this should not necessarily delete this resource
delete(background=False)	delete this resource

cloud.Instance(Resource)

An object representing a cloud instance with a set of Resources that can be allocated/deallocated.

properties	desc.
resource_managers	list of ResourceManagers
methods	desc.
down(background=False, delete_resources=True)	stop this instance and optionally delete all managed resources
delete(background=False, confirm=True)	delete this instance with optional user confirmation

cloud.ResourceManager

Class for managing the creation and maintanence of cloud.Resources.

properties	desc.
instance	cloud.Instance instance owning this resource manager
resource_cls	cloud.Resource type, the class of the resource to be managed
resources	list of cloud.Resources, managed resources
methods	desc.
__init__(instance, resource_cls)	instance: the cloud.Instance object operating this ResourceManager
	resource_cls : the cloud.Resource class this object manages
add(*args, **kwargs)	add an existing resource to this manager
remove(*args, **kwargs)	remove an existing resource from this manager

Amazon EC2

cloud.AWSInstance(Instance)

A cloud.Instance object for AWS EC2 instances.

Azure

cloud.AzureInstance(Instance)

A cloud.Instance object for Microsoft Azure instances.

Google Cloud

Our GCPInstance requires that your instances have gcloud installed and properly authenticated so that gcloud alpha compute tpus create test_name runs without issue.

cloud.GCPInstance(Instance)

A cloud.Instance object for Google Cloud instances.

properties	desc.
tpu	cloud.TPUManager, a resource manager for this instance's TPUs
resource_managers	list of owned cloud.ResourceManagers
methods	desc.
__init__(collect_existing_tpus=True, **kwargs)	collect_existing_tpus : bool, whether to add existing TPUs to this manager
	**kwargs : passed to cloud.Instance's initializer

cloud.TPU(Resource)

Resource class for TPU accelerators.

properties	desc.
ip	str, IP address of the TPU
preemptible	bool, whether this TPU is preemptible or not
details	dict {str: str}, properties of this TPU
methods	desc.
up(background=False)	start this TPU
down(background=False)	stop this TPU
delete(background=False)	delete this TPU

cloud.TPUManager(ResourceManager)

ResourceManager class for TPU accelerators.

properties	desc.
names	list of str, names of the managed TPUs
ips	list of str, ips of the managed TPUs
methods	desc.
__init__(instance, collect_existing=True)	instance: the cloud.GCPInstance object operating this TPUManager
	collect_existing: bool, whether to add existing TPUs to this manager
clean(background=True)	delete all managed TPUs with unhealthy states
get(preemptible=True)	get an available TPU, or create one using up() if none exist
up(preemptible=True, background=False)	allocate and manage a new instance of resource_cls