View on GitHub


Depreciation Alert

This tutorial is no longer maintained - please use it accordingly.

Google Cloud for Deep Learning

This small tutorial is a little help for students, who are participating in “DEEP LEARNING A GYAKORLATBAN PYTHON ÉS LUA ALAPON” at Budapest University of Technology and Economics.

If you have any questions, please feel free to contact me! All suggestions are welcome!


If a ~2GB GPU is enough for you I would like to recommend and skip this document. It’s easier to setup and use than the workflow listed below.

GCP Free Tier

Table of content:

  1. How to register
  2. Setup Google Cloud SDK
  3. Google Cloud Compute - Virtual Machines
    1. How to start an instance
    2. Useful commands
  4. Datalab
    1. Create Datalab

WARNING: The registration, and the usage of the services is free for a limited amount of duration, and resource, BUT only for this duration.

How to register

Register with your Google account here.

Setup Google Cloud SDK

It’s much easier to use GCloud with a CLI (command line interface) than with a browser. I highly recommend to use the SDK.

On linux (Ubuntu 16.04LTS): Quick Start Guide by Google

Installing of the SDK:

# Create an environment variable for the correct distribution
export CLOUD_SDK_REPO="cloud-sdk-$(lsb_release -c -s)"

# Add the Cloud SDK distribution URI as a package source
echo "deb $CLOUD_SDK_REPO main" | sudo tee -a /etc/apt/sources.list.d/google-cloud-sdk.list

# Import the Google Cloud Platform public key
curl | sudo apt-key add -

# Update the package list and install the Cloud SDK
sudo apt-get update && sudo apt-get install google-cloud-sdk

Initialization of the SDK on our system:

Type gcloud init and login to your google account (browser needed). After successful login choose between projects (or creat a new one with the CLI).

Google Cloud Compute - Virtual Machines

So google cloud instance is just a virtual machine (just like in AWS), but you have much more freedom to scale it. For Deep Learning it’s a necessary to use GPU (it’s event more important with bigger models, and with more data), therefore we will see how can we setup an instance with GPU. WARNING: I want to repeat myself, that this is not free (however you have free credits).

GPU Quota

If we want a GPU instance, we need to request it first, because it’s a limited resource, and the initial quota for GPU’s is zero.

How to start an instance

Because GUI is much more interactive, I will show this process with web-GUI on Google Cloud Management Console:

Now we need to define the properties of our instance:

Instance Creation Image


If the instance is created, and running, you can connect to it via SSH, or with CLI. To connect with CLI just type the following command: gcloud compute ssh [INSTANCE NAME]

Useful commands:

gcloud compute instances list : List all available instances in a given group.

gcloud compute ssh [INSTANCE NAME] : Connect to an instance via SSH.


Datalab is an environment where you can run and test Python Notebooks easily. Original documentation can be found here.

Create Datalab

Let’s use GCloud CLI for this. Open a terminal and list all the installed components with </br>

gcloud components list

If the Cloud Datalab Command Line Tool isn’t installed, install it with</br>

gcloud components install datalab
# or with
sudo apt-get install google-cloud-sdk-datalab

If it’s installed, create an actual datalab environment with the following command:

datalab create [VM-INSTANCE-NAME]
# For example
datalab create

This will create a new VM instance with the given name. You can check the deployment state on Google Cloud Console on web, or with CLI. If the deployement is done, you can reach your environment on localhost:8081. If the terminal command window used for running the datalab command is closed or interrupted, the connection to your Cloud Datalab VM will terminate, and you will need to run datalab connect [vm-instance-name] to reestablish the connection to your VM.

You can delete the VM with the follwoing command:

datalab delete [vm-instance-name]

So we have an amazing webserver, where we can test our code, but what if we need additional packages? For exaample Keras?

Just like in a simple environment, we can use our notebook to install python packages:

!pip3 install keras

Keras Install Image