3 Cloud Deep Learning Notebooks in 2021
In a warm winter sun, you sit at a cafe by the lake, open a browser on your laptop, log in the “notebook” in the cloud, type in a deep learning model you designed, press Run All, and continue to enjoy the breath-taking natural beauty in front of you. After a while, finish your coffee, the training results are already stored in your cloud account. The cloud-based deep learning notebooks allow you to have powerful computing power anytime, anywhere.
In 2021, there are 3 proven platforms that can bring you the above experience. They are all cloud solutions based on Jupyter notebook. If you have used Jupyter notebook, there is no difficulty in getting started with them.
1. Google Colaboratory
Colaboratory, or “Colab”, is a product from Google. It is a hosted Jupyter notebook service that requires no setup to use while providing free access to computing resources including GPUs and TPUs.
Colab is free to use, but there are restrictions on its resources. Resources available in Colab vary over time to accommodate fluctuations in demand. This means that idle timeout periods, maximum VM lifetime, memory, GPU types available, and other factors vary over time.
The GPUs available in Colab often include Nvidia K80, T4, P4 and P100. But there is no way to choose what type of GPU in the free edition.
Although the official claim that the VM connecting maximum lifetime can be as much as 12 hours, Notebooks will also disconnect from VMs when left idle for too long. Therefore, when using Colab, it is important to estimate the training time and to prevent excessive idle time. So it is more suitable for training “smaller” models or just for experiments.
If you are interested in faster GPUs, longer runtimes, and more memory, you may be interested in Colab Pro. For now, Colab Pro is only available in the US and Canada.
Colab notebooks are stored in Google Drive, or can be loaded from GitHub. You can share your Colab notebooks just as you would with Google Docs.
Colab focuses on supporting Python and its ecosystem of third-party tools
2. Kaggle Notebooks
Every data science and machine learning developer should be familiar with Kaggle. Google acquired Kaggle in 2017, so it is technically backed by Google.
Kaggle Notebooks is a cloud computational environment for data science and machine learning. Its predecessor was Kaggle Kernels.
To use the Kaggle notebook, simply sign up for Kaggle.
Unlike Google Colab, there are two different types of Notebooks on Kaggle: Scripts and Jupyter notebooks. They all may be written in either R or Python.
Kaggle Notebooks provide the following resources: 9 hours execution time, 20 G disk space, 4 Cores CPU with 16G RAM, GPU with 2 Cores CPU and 13G RAM, TPU with 4 Cores CPU and 16G RAM, and 1 hour of idle time. This resource configuration is better than the Colab free.
Besides this, you can add multiple data sources to your Notebook’s environment. Kaggle Datasets provides a rich mix of interesting datasets for any kind of data science project. And Kaggle Community also provides lots of great notebooks by other “Kaggle Master.” There are many free courses on data science and machine learning in Kaggle.
3. Amazon SageMaker notebooks
Amzon SageMaker is a cloud machine-learning platform at the AWS. You can use Amazon SageMake Stuido(like JupyterLab) to build, train, debug, deploy, and monitor your deep learning models.
You can write and run your deep learning code with the SageMaker notebooks or the SageMaker Python SDK.
There is a fee for using Amazon SageMaker notebooks. For example, Works on a Amzon SageMaker Stuido Notebook in a
TensorFlow kernel on an
ml.c5.xlarge(4 vCPu, 8GiB Memory) instance for 1 hour, the total cost is less than $0.5.
Besides the usual deep learning frameworks, TensorFlow and PyTorch, Amazon SageMaker also supports MXNet. If you use the MXNet to build your deep learning models, the Amazon SageMaker is your preferred platform.
I personally recommend Kaggle Notebooks, for the free cloud deep learning platform, because of its high performance, rich datasets, and excellent notebooks, as well as its healthy community.
There is a post “Training in Kaggle vs Colab vs SageMaker (ml.p2.xlarge)” on Kaggle. The author was running the same PyTorch training script on 3 platforms. The training speed is:
Kaggle > Google Colab(Free) > SageMaker(ml.p2.xlarge)
If you need higher performance hardware and more professional service, you can customize the Amazon SageMaker or Google Colab Pro on-demand.
Future — Codespaces
Originally, this list also included Microsoft Azure Notebooks, but it will be retired on January 15th, 2021. Its most likely successor is Codespaces.
Codespaces is an online development environment, hosted by GitHub and powered by Visual Studio Code, that allows you to develop entirely in the cloud.
Combining the strongest scripting IDE-VS Code and the largest open-source code repository and community-Github, with the support of Microsoft, the former dominant IDE. Codespaces is the future of IDEs.
For now, Codespaces is still in beta version, you need to request early access.
Code anywhere, anytime, training in the air.
Thanks for reading and Happy 2021.